• The standalone operating thesis was emphatically validated: revenue rose 136.5%, gross margin reached 53.2%, operating margin hit 42.3%, and December guidance points even higher. Optical scarcity is producing extraordinary economics, although cash conversion remains much weaker than reported profit.

• The merger changes the strategic question: Lumentum had already proved it could win major CPO commitments independently. What Marvell adds is earlier influence over system architecture processors, SerDes, switching, photonics and packaging potentially moving Lumentum from supplying a chosen component to helping define the design itself.

• Shareholders are no longer underwriting pure optical scarcity: the all-stock structure turns Lumentum into a contingent claim on Marvell, with regulatory risk, Marvell share-price risk and uncertainty over whether integration creates enough incremental value to justify giving up independence.

In 2018, Michael Hurlston agreed to sell Finisar to II-VI. It is tempting to imagine an optical-industry veteran recognizing a cyclical peak. Finisar’s business was already under pressure, however, and much of the consideration was cash. The lesson is subtler: optical scarcity creates extraordinary value, but what matters is who can preserve that value when capacity and technology catch up.

Eight years later, Hurlston has agreed to sell another optical company under very different circumstances. On November 2, Lumentum reported the strongest quarter in its history and agreed to exchange each of its shares for 4.5 Marvell shares. The deal valued Lumentum at about $115 billion of enterprise value when announced. Instead of an earnings call, investors received a merger call, which began with an instruction not to ask about current business conditions.

The numbers were no longer the principal story. They were the evidence against which to judge the decision to sell.

The Earnings Call That Wasn’t

Revenue increased 136.5% year over year to $1.26 billion. Non-GAAP gross margin reached 53.2%, operating margin 42.3%, and December-quarter guidance pointed to $1.575–1.625 billion of revenue and 45–46% operating margin. August’s operating tests were passed: gross margin above 50%, operating margin above 40%, optical circuit switching exceeding its first $100 million quarter, and another NPO/CPO design win supported by more than $5 billion in purchase commitments. The full durability test—whether tightness survives new industry capacity—cannot be settled in one quarter.

Even the margin mix was remarkable. Systems rose to 40% of revenue without preventing gross margin expansion. Yet the cash picture was less spectacular: roughly $461 million of adjusted net income produced just $38 million in free cash flow, with capital expenditure near $230 million. Lumentum has built extraordinary pricing power in factories that remain expensive to expand.

What Hurlston Is Actually Selling

The most revealing evidence is a change in Hurlston’s description of his own business. In August he said:

For us, again, we see ourselves as largely a component supplier. We want to supply components. We also see ourselves as a great packaging house... We can apply that packaging capability to the whole optical engine.

On the merger call he argued:

While customers can try to source these independently from multiple suppliers, Lumentum and Marvell believe the performance of the solution will be dramatically better with a faster time to market if it’s co-architected and delivered by a single supplier. This is why we believe now is the time for this combination.

The shift is not from components to packaging: he already contemplated complete optical engines in August. It is from acquiring electronics from partners to owning the electronic expertise beside the optics. Marvell works with hyperscalers years earlier on processors, switching, SerDes and connectivity architecture; Lumentum has historically supplied critical lasers after more of those decisions were made. As light moves from connecting data centers to linking racks and eventually sitting beside processors, the electrical and optical design choices become harder to separate.

Consider a laser that improves optical yield. As a merchant component, Lumentum can charge a premium for the laser. But if the entire engine is co-designed: laser, photonics, driver electronics and packaging, the supplier involved at the architectural stage may influence both specifications and how much content it supplies. Marvell is purchasing the chance to move Lumentum from a component chosen after the design to a participant in defining the design.

This is a credible reason to merge, but not proof common ownership is required. Lumentum won that $5 billion-plus CPO commitment independently. Customers already coordinate electronics and photonics through contracts. Hurlston’s stronger claim is that integration makes the collaboration faster and more reliable, not that independence made it impossible.

The form of payment matters too. This is not Finisar’s largely cash exit. Lumentum holders will own around one-third of the combined company, and Hurlston will join its board. They are exchanging control of one optical bottleneck for a claim on a broader system. That could express confidence in the future of optical scarcity, not fear of its imminent collapse. It could also diversify a highly concentrated manufacturing bet.

The price is not obviously generous either. On estimates for the year ending June 2028, Lumentum contributes roughly 31% of the two companies combined net income and receives approximately 33% of the shares. Marvell expects more than $1 billion in annual cost synergies, but those are forecasts, not value already earned. Lumentum shareholders are surrendering independent upside in exchange for participation in a potentially more valuable architecture, without a large cash premium protecting them if that promise disappoints.

A sale is evidence of management’s preferences, not privileged access to the cycle’s endpoint.

Nvidia’s Interest and Coherent’s Alternative

Nvidia’s role makes the decision more consequential. In March it committed $2 billion each to Lumentum and Coherent, and another $2 billion to Marvell, alongside arrangements supporting optical capacity and the NVLink Fusion ecosystem. Bloomberg reports that Hurlston and Marvell’s Matt Murphy sought Jensen Huang’s views before signing and received his support. That is evidence of influence, not proof that Huang originated or controlled the deal.

For Nvidia, the potential return extends beyond its investments. Hyperscalers building custom accelerators need not build an entirely separate connectivity system if Marvell can deliver chips and interconnect compatible with Nvidia’s architecture. Securing optical capacity and encouraging compatible integration can protect Nvidia’s position even when a rack contains someone else’s accelerator. An independent Lumentum purchased by a rival or hyperscaler might have constrained that flexibility. This is a possible motive, not a disclosed condition of Huang’s endorsement.

Coherent offers the competing organizational model. It has secured a roughly $20 billion five-year PhotonLink agreement and is expanding indium phosphide output, with capacity projected to more than double again by the end of 2027. Its management emphasizes the value of independence: a photonics provider can serve multiple system architects without making its customers depend on a competitor. Coherent’s contracts also demonstrate that optical suppliers can win integrated-system commitments without owning a large merchant DSP franchise.

The two models each have a cost. Combining electronics and optics may shorten development cycles and improve performance but could make competing module makers wary of sharing roadmaps with their own supplier. Remaining independent preserves neutrality but may leave the richest architectural decisions to semiconductor companies. On the merger call, Hurlston said Lumentum supplies lasers to nearly all its competitors and intends to continue. Customers will decide whether that promise is credible after ownership changes.

That is the deeper uncertainty. The optimal boundary for designing an optical engine need not be the optimal boundary for owning the companies that build it.

What Shareholders Own Now

The all-stock structure has already transformed the investment. Marvell closed November 5 at $133.55, making the fixed consideration worth about $601 per Lumentum share; Lumentum closed at $509.70, a 15% discount to that floating value. This is not free arbitrage. Completion is expected in the first half of 2028, and investors face regulatory approvals, time, and Marvell’s share-price risk. The announced 24% premium against unaffected prices is no longer a cash amount anyone can collect.

The previous standalone valuation framework must therefore become a transaction-and-industry framework. My illustrative end-2029 values are:

The probability-weighted value is roughly $820, although the precision is deceptive: earnings, exit multiples and completion probabilities are our assumptions, not contractual economics. A regulatory termination payment could add value in a qualifying break; it is not a floor in every failed-deal scenario. In successful-deal cases, the share-price outcome increasingly depends on Marvell’s entire semiconductor business, not merely Lumentum’s lasers.

There are now two tests, and investors should not confuse them. The operating test is whether Lumentum’s December gross margin reaches approximately 54% or higher as revenue accelerates; 52% or lower would strengthen the case that scarcity economics are peaking. Coherent’s 2027 capacity ramp gives that test a deadline. The transaction test is whether the merger proxy supports the claimed need for common ownership, whether merchant customers remain, and whether integrated optical designs generate enough incremental profit to justify surrendering independence.

For now, I would not treat the stock’s fall as a reason to add mechanically. Lumentum is no longer simply a concentrated bet on optical pricing. It is a discounted, contingent claim on Marvell, with the independent business as an alternative if the deal breaks. A better operating quarter cannot remove those transaction risks.

In August, the thesis was that the optical boundary would keep moving closer to compute, making Lumentum’s lasers more valuable. November validated that movement emphatically. What we did not anticipate was management’s conclusion that the next portion of its value might be captured more effectively by the company designing the entire electrical-optical system.

Hurlston has not told us the window is closing. He has told us that, in his judgment, the toll booth belongs inside a larger architecture. Whether that architecture earns more than an independent Lumentum could have earned is now the investment thesis.

Subscribe now

• The marketplace is getting better at utilization: Uber is shifting incentives toward scarce supply, expanding products like Wait & Save and premium rides, and says autonomous partners complete about 30% more trips per vehicle per day than first-party fleets. The operating machine is improving.

• But supplier choice is becoming more important than matching efficiency: Waymo wants customers to request Waymo specifically, while Amazon can put Uber Eats fulfillment behind its own front door. Uber can be the marketplace, or merely the infrastructure behind someone else’s customer relationship.

• Uber One may be the bridge: membership and supplier-funded benefits can give a customer or an AI agent a reason to select Uber even when the interface sits elsewhere. But those benefits only matter if they improve the all-in offer and do not simply transfer margin away from Uber or fleet owners

In 1927, Henry Ford halted Model T production to retool. His assembly line still worked; what had changed was the customer. Alfred Sloan’s General Motors offered cars for different budgets, regular model updates and financing through GMAC. Ford had perfected the economics of making a car. GM had built a relationship that could accommodate the customer’s next purchase. By the time Ford responded, Chevrolet had passed him.

In February 2025’s Beyond the Ride, I argued that autonomous vehicles would not eliminate Uber’s value: removing the driver does not remove the challenge of keeping expensive transportation capacity productive. August’s Proved the AV Thesis. Now Comes the Bill asked who would finance that capacity. Q3 strengthens the operating argument. Uber says its US autonomous partners complete roughly 30% more trips per vehicle per day than first-party fleets; gross bookings grew 20% in constant currency and non-GAAP operating income 38%. Yet at $68.78 on November 5, the stock was almost unchanged from its price when the February article appeared.

The week’s developments suggest why the next question is different. Waymo wants riders to ask for Waymo, not just a driverless car. Amazon now sells restaurant delivery fulfilled by Uber without requiring an Uber account. AI agents may eventually compare both offers on a customer’s behalf. Uber has proved that its network can make transactions more efficient; it has not established how much of the value it will retain when others influence which transaction gets made.

The Marginal Dollar

Mobility bookings grew 17% in constant currency, a little below management’s plan, while trip growth slowed from 18% to 16%. International driver availability and fuel costs constrained supply; the anniversary of the Trendyol Go acquisition also affected consolidated trip comparisons. Dara Khosrowshahi explained Uber’s response:

Then we have incentives... These are more flexible dollars that we target... whereas last year, more of the incentives went to riders to reduce price, now this year, we’re taking down rider incentives, and we’re using those dollars to fund driver incentives. Now that does mean sometimes higher prices for riders, which is something that we watch out for.

Uber shifted money toward the scarce side of the marketplace, tolerating some higher rider prices while Mobility operating margin reached about 8% of bookings. The result was not simply better matching. Uber continuously decides whether another dollar should attract a driver, persuade a rider or create an order a merchant is willing to subsidize. The company can also change the product to make supply easier to use.

Wait & Save exceeded an $8 billion annualized bookings run rate, doubling over the year, while premium products grew roughly 35%. Uber has largely deprioritized UberX Share. The prepared remarks describe the broader design:

The same rider may choose Wait & Save for an everyday trip and a premium product for a business journey. Our advantage is the ability to serve those different needs on one platform, rather than offer the same experience for every occasion... A broader mix of trips improves driver utilization and marketplace efficiency, creating more room to reinvest in affordability and reliability.

Both cheaper products ask riders to accept something in return for a discount but Wait & Save purchases flexibility in pickup time rather than tolerance for a shared journey. My inference is that the first concession is easier to turn into efficient dispatch than the second; Uber has not published a controlled comparison. Together with premium services, the choice reveals a network that can offer different combinations of time, price and certainty to the same person on different occasions.

That is also the economic logic behind autonomous utilization. If a fleet completes 30% more trips with unchanged fixed costs, fixed cost per trip falls about 23%, before variable costs and marketplace fees. Uber’s reported advantage is not a controlled comparison of identical fleets, so it should not be mistaken for a measured profit uplift. But within each service tier, efficient dispatch depends on treating qualified vehicles as substitutes. A rider requests an UberX or an autonomous option; the network wants freedom to assign the best available car. A supplier with its own brand may want exactly the opposite.

The Refusal and the Login

On November 4, Waymo co-CEO Tekedra Mawakana described requesting a Waymo through Uber as something of a “crapshoot”: customers wanting a robotaxi might get a human driver. Waymo plans its own app in Austin and Atlanta from January 2028, while continuing its existing Uber distribution arrangements. The plan was reported in July, but November brought the customer-relationship argument into sharper focus. Uber says it offered a dedicated robotaxi option that Waymo declined; Waymo has not corroborated the account or its motivation.

The tension exists regardless of that disputed detail. Uber’s utilization advantage comes from allocating interchangeable capacity. Waymo’s direct app and $29.99 Premier membership encourage customers to choose a particular supplier. That does not make Uber dispensable: Waymo can value incremental marketplace demand and still cultivate direct riders. But it corrects our February assumption that AV developers would generally need Uber more than Uber needed them. The leader, with its own capital and consumer recognition, has a different bargaining position from smaller fleets.

Amazon presents the mirror image. On November 5 it introduced Uber Eats restaurant ordering inside Amazon’s website and app. Prime members get delivery-fee waivers on qualifying orders, pay with Amazon credentials and need no Uber account. Uber fulfills the transaction, but Prime supplies the conspicuous membership proposition. Compare Uber’s April Expedia partnership: Uber is the customer-facing destination, Expedia supplies hotel inventory, and Uber One provides benefits. Uber can be the front door to someone else’s supply, or the supply behind someone else’s front door.

Neither arrangement is inherently bad. An externally originated order can be profitable without becoming a direct customer; owning the front door may require substantial subsidies. Amazon’s restaurant integration also excludes Uber’s fast-growing grocery business, which Amazon has reason to serve itself. The partnership validates the delivery network without establishing that Uber will capture the wider purchasing relationship.

What matters is the bargaining position created by discovery, benefits and the ability to influence a future purchase. As Khosrowshahi said when asked about AI agents:

I used to work in the travel business where we were dependent on the Googles of the world. I don’t want to go back there... That said, we will have discussions with third-party AI agents... every single AI system company is talking to us, and we’re really now negotiating the terms of trade. And the good news here is, in most markets, you have two mobility, big mobility providers, two big delivery providers. I think I’ve counted now there are eight different assistants, and I think there’ll be more than 10 assistants.

The negotiating case is credible: there are few scaled physical-service networks and many potential AI interfaces. Yet scarcity does not settle how much each party gets. Agent-originated activity remains negligible, Uber cited roughly 500 Eats orders and four rides in a Halloween week encompassing hundreds of millions of transactions. The risk is prospective: comparison becomes easier, a differentiated supplier can favor its own channel, and a powerful distributor can ask for more of the surplus. Uber needs to be selected for reasons beyond having inventory available.

Uber One’s Next Job

Uber One has approximately 55 million members, accounting for more than half of combined Mobility and Delivery bookings and around 70% of Delivery bookings. That restores the comparable membership disclosure we criticized in August. It establishes considerable engagement, not how much spending the subscription caused. We also overstated its defensive power in May: a member can take a Waymo ride without canceling Uber One. Membership creates a reason to return, not an obligation.

Agents make this distinction explicit. The fee a customer has already paid is sunk; the benefit available on the next ride is not. A useful assistant should recognize applicable credits, pickup priority, service quality and future rewards alongside the fare. A program becomes durable when those advantages improve the offer being compared, rather than relying on customers failing to compare at all.

Uber is testing elements of that shift. A limited-time Platinum and Diamond tier pilot rewards qualifying rides and orders over a rolling three-month period, while management has discussed richer benefits linked to frequency. This is not proof of increased incremental retention. It is a move from a flat prepaid discount toward rewards whose value can vary with usage. The more interesting question is who funds them.

Booking’s Genius program suggests an answer. A hotel with an empty premium room may fund an upgrade that is worth far more to a traveler than it costs to provide. Uber already has a related mechanism: merchants spend more than $10 billion annualized on advertising and offers across its platform. That figure is not advertising revenue or a loyalty pool. It does show that suppliers are willing to pay to influence incremental demand. If Uber can connect that willingness to benefits members actually value, the program might improve both customer preference and supplier returns.

Autonomous fleets make the logic more consequential. Human drivers can log off when demand disappears; a fleet’s capital cost persists while its vehicles sit idle. An off-peak perk could create a ride that would not otherwise happen and leave its owner better off, although energy, maintenance, insurance and repositioning still cost money. A discount on a ride that would have occurred anyway simply transfers profit. Uber’s advantage is its ability to distinguish those occasions through pricing, dispatch and demand data, not an ability to make every subsidy productive.

Finally, benefits must be recognizable beyond Uber’s own interface. Costco’s membership integration on Uber Eats shows that two relationships can coexist inside a transaction; Amazon shows that the external relationship can instead dominate the visible offer. An assistant might eventually compare an Uber One benefit with a supplier’s direct perk, but Uber has not demonstrated universal portability into outside agents. Nor can an impeccable loyalty offer promise a driverless ride unless Waymo makes that product bookable. The agent needs a credible all-in proposition, price, pickup window, service type, usable rewards and reliability, not a discount on something Uber cannot deliver. The opportunity is to give the next chooser a real economic reason to select Uber, with the supplier helping fund that reason.

Who Keeps the Gain?

August’s capital question remains unresolved. Khosrowshahi offered a possible mechanism when discussing financing for autonomous fleets:

If there’s a fleet that for some reason fails in terms of their operations, we can take that AV vehicle and put it under separate ownership, and it will still earn on the Uber network, so to speak. So we think that the financing on the Uber network can have a real advantage in terms of cost of capital versus alternatives.

If a vehicle can continue earning after its operator fails, lenders may be financing access to a durable demand network rather than betting entirely on one fleet company. That could reduce the cost of capital and Uber’s need to own vehicles. But the conditional words matter: software licenses, permits, maintenance arrangements and residual-value risk may constrain transferability. External financing transfers value and risk only according to its actual terms; it cannot make a fleet owner’s required return disappear.

The core operation can support the transition, though not without limits. Q3 free cash flow reached $2.54 billion, up about 14%, below non-GAAP operating-income growth. Buybacks increased to $750 million, and management is limiting additional M&A while it integrates acquisitions. Some early autonomous vehicles and related obligations will nonetheless remain on Uber’s balance sheet. The same utilization gain cannot entirely fund cheaper rides, superior member benefits, Uber’s marketplace fee and the capital owner’s return. This is why financing terms and the AV take rate belong in the same investment discussion: each determines how the improvement is shared, and which risks shareholders retain.

At $68.78, Uber traded near 15.6 times FactSet FY2027 non-GAAP EPS consensus of approximately $4.40. I would retain August’s three outcomes rather than assign a premium to loyalty features or financing arrangements that have not yet demonstrated their returns.

Their probability-weighted value remains approximately $112, 63% above the November 5 close. These are our previously published outcomes, not freshly recalculated forecasts. The next evidence that could justify greater conviction is a disclosed AV take rate, financing that genuinely transfers retained obligations, and Uber One benefits that increase incremental contribution rather than merely reward existing heavy users. Without those proofs, the share price’s discount cannot simply be called a mistake.

Ford’s factories had not stopped working when General Motors overtook him. The source of advantage had moved from producing efficiently to offering customers something they preferred. Uber’s utilization machine is getting stronger, but Waymo and Amazon reveal how that same machine can operate for someone else’s customer relationship. Its next task is not preventing choice. It is creating an offer worth choosing and retaining enough of the benefit to reward the suppliers and shareholders who make it possible.

Subscribe now

• The strategic position is stronger than we thought: data-center royalties doubled, Arm expects roughly 60% server CPU share by year-end, and independent licensing surged despite Arm’s move into its own CPU. The evidence suggests customers may tolerate Arm being both platform supplier and chip vendor better than feared.

• But the economic mix is deteriorating: royalties grew only 8% while licensing rose 50%, reducing royalties to 46% of revenue. Licensing is less recurring, and silicon is materially more capital-intensive, with initial AGI CPU gross margins expected at just 30–40%.

• That makes valuation the real problem: consensus now expects Arm to generate substantially more revenue than our old bull case, but less EPS. The company can become strategically more important while the incremental dollar of revenue becomes less valuable.

In 2015, activist Jana Partners urged Qualcomm to separate its patent-licensing business from the chips it sold into the same mobile market. Qualcomm declined, arguing that the combination was worth more than the parts. The disagreement was not about whether making chips was useful. It was about whether revenue earned competing for hardware sockets deserved the same valuation as royalties collected across an industry.

Arm is moving toward a similar structure from the other direction. It collects license fees when chips are designed and royalties when they ship; from the March quarter it expects to sell its own AGI CPU. Qualcomm’s patent licensing and Arm’s processor architecture are different businesses, but the valuation problem is recognizable. At the $191.80 pre-results close, Arm traded around 64 times FY2028 consensus earnings. The question is whether an increasingly hybrid business can keep earning a licensor’s multiple.

What May Got Right and What Changed

In May’s AI Upside, Platform Risk, I argued that Rene Haas was choosing to monetize Arm’s architectural position more directly, at the possible cost of neutrality. We wrote:

The risk is asymmetric in time. The capture benefit is realized first, in the FY27 and FY28 AGI CPU revenue ramp. The neutrality cost is realized later, in the FY29 through FY31 royalty growth that disappoints because licensees have spent the intervening years quietly hedging.

The capture is now more credible, but it has not yet reached the income statement. Management sees up to $2 billion of AGI CPU revenue across FY2027 and FY2028, supported by purchase orders extending through calendar 2027; first commercial deliveries are expected in the quarter ending March 2027. That is evidence of customer commitment, not yet proof of profitable silicon sales. Meanwhile, our February test for data-center royalty growth at least 70%, was passed as management reported a doubling. Consolidated royalties rose just 8%, however, well below our approximately 20% expectation. Rising memory prices and weak lower-end phones, rather than demonstrated licensee retaliation, explain the immediate shortfall.

I had been too quick to assume that an Arm chip and an independent Arm chipmaker must be enemies. The quarter offers evidence that customers may want both. It also exposes a second error: my earlier price target depended on earnings Arm has not yet generated. The shares reached roughly the price I expected years later before the silicon ramp had begun.

A Larger Opportunity Than Server Share

Haas described the data-center expansion on the call:

I think there is two vectors independent of the Arm AGI CPU on the royalty. One is Arm taking market share, which I mentioned before. Now we think it is going to be 60% by the end of the year. That market share gain is driven both in terms of units going up, agentic AI driving increased design starts for Graviton, Cobalt, Axion, et cetera. In addition, we are seeing a transition away from x86 to Arm for the head node CPU that connect to accelerators. We already know about Vera Rubin, but Google has already announced the conversion of x86 to Axion, connecting to TPU version 8.

The second vector changes the size of Arm’s opportunity. The AI server is not simply an accelerator with a few attached CPUs. It needs hosts to prepare inputs, move data, coordinate tools and execute the software between accelerator calls. More agentic inference can mean more CPU work per accelerator, while x86-to-Arm conversion adds new royalty-bearing designs. Even a hyperscaler market in which Arm is already dominant can therefore produce additional royalty growth through a larger amount of CPU work.

The timing is less attractive. Arm earns a license when designs begin, but royalties when silicon ships, often 12–24 months later. Data-center royalties can double while the larger smartphone base still determines consolidated growth. Cloud AI was just 14% of royalties in FY2026, and management says data center is approaching one-fifth today. The market for Arm architecture is changing faster than the royalty mix, which is why a stronger roadmap has not yet become stronger recurring earnings.

The most telling new evidence on neutrality comes from the 6-K. Independent-customer license and other revenue increased from $222 million to $490 million, while related-party licensing declined from $293 million to $283 million. Three Arm Total Access agreements, including a major customer’s first, and four Compute Subsystem licenses reinforce the point: customers are committing more resources to Arm despite knowing it plans to compete in processors. Oracle’s intention to offer both Arm’s own CPU and NVIDIA’s Arm-based Vera shows how buying Arm silicon and designing with Arm IP could coexist. The better analogy may be Microsoft’s Surface strengthening the Windows ecosystem, rather than Windows Phone alienating it.

Here lies the paradox of Arm’s success. Graviton, Axion and other custom CPUs can make the architecture ubiquitous while leaving hyperscalers to capture most of the value from the finished systems. The AGI CPU is an attempt to claim more of that surplus. But Arm cannot both demand the economics of the finished chip and assume its customers will continue investing in their own implementations on unchanged terms. The licensing surge suggests it can offer both paths; it does not mean it can force customers down the more profitable one.

That does not settle the neutrality question. Arm’s position depends on licensees believing they can obtain competitive access to architecture, roadmaps and commercial terms while Arm sells against them. That belief will be tested at renewals, not during a single strong licensing quarter. Arm is becoming a more useful platform; whether it can remain a trusted one is still open.

Three Currencies

In July, CFO Jason Child said that licenses could offset weaker royalties:

I would expect similar trends for the rest of the year, and that is any softness or weakness that we experienced in royalties will be at, I would say, at least as much license revenue to overcome that. So, I don’t really expect overall to be any sort of a slowdown.

That is what happened: total revenue rose 27% to $1.44 billion, but the kind of revenue Arm collected changed.

Royalties are the premium currency. Once a design is shipping, Arm receives a claim on production without financing the chip.

Licensing is a payment for engineering access today and a possible source of royalties later. Its 50% revenue growth looks less extraordinary beside just 12% growth in annualized contract value and deferred revenue declining from $1.05 billion in March to $875 million in September. When asked whether $770 million was now the quarterly licensing baseline, Child replied:

We have not talked about next year, whether or not that is the new number every quarter. I do not know if I would say that.

Arm has also dropped RPO and active license counts from its shareholder letter, saying the metrics are less relevant as it moves into production silicon, although the filing still shows $2.42 billion of RPO. These are reasons to investigate the quality of licensing growth, not dismiss it.

Silicon is the third currency and changes the company’s obligations. The first AGI CPU is expected to earn 30–40% gross margin, versus nearly 98% for Arm’s current IP-heavy business. Future versions may reach conventional chip margins above 50%, but that improvement must be earned through product performance, volumes and supply-chain execution. The 6-K describes $672 million of non-cancellable chip purchases matched by a related-party reseller agreement; it does not identify that arrangement as AGI CPU backlog. Separately, Arm faces minimum-volume capacity obligations and anticipated $200 million of advances. An IP licensor can earn across many customers without taking these manufacturing risks. A chip vendor cannot.

Cash conversion illustrates the distinction. Free cash flow was $110 million against $566 million of adjusted net income this quarter, though first-half conversion was much stronger. Stock compensation remains about 26% of revenue, and $407 million of first-half employee share-tax payments sat in financing cash flows outside Arm’s free-cash-flow definition. The point is not to treat one weak quarter as permanent. It is that the new revenue engine will increasingly demand cash and capital from a company valued for licensing economics.

The paradox appears in consensus. Our February bull case assumed FY2029 revenue of $9–9.5 billion and EPS of $4.50–5.00. FactSet now expects about $10.9 billion of revenue but only $3.93 of EPS. Arm is forecast to sell far more than our most optimistic case anticipated, while earning less from those sales. The market is getting the growth story; it has not yet settled what kind of growth it is buying.

What the Hybrid Is Worth

I would retain our three outcomes, valued at end-2029 on forward FY2031 earnings. They differ less in whether Arm participates in AI than in how much of that opportunity remains high-margin IP, how much becomes silicon, and what investors should pay for the blend.

The base case is more ambitious than its name suggests: silicon supplies over half of sales, yet the IP business must still support much of the profit. The probability-weighted value of approximately $215 leaves only modest annualized appreciation from $192. Even if Arm more than doubles forward earnings, a lower multiple can consume much of the gain. The issue is not whether the AGI CPU sells; it is whether selling it expands the high-return architecture franchise or gradually changes the kind of company investors own.

I am more confident in Arm’s computing position and less concerned about immediate licensee abandonment than I was in May. I am also less willing to give the new revenue a licensor’s multiple. At $192, the stock appears fair rather than cheap; nearer $160–170, the prospective payoff becomes more attractive if royalty recovery and customer commitments remain intact. Next quarter’s royalties, subsequent ACV growth, AGI CPU shipment economics and licensing renewals will reveal which revenue currency is becoming dominant.

Qualcomm’s board decided that licensing and chips belonged together. Arm may reach the same strategic answer, and its customers have so far given us reason to believe coexistence is possible. But building a more important computing company does not automatically create a more valuable share of its earnings. The crucial distinction is no longer whether Arm can win more of the AI market, but how it gets paid when it does.

Subscribe now

• The logistics thesis is working operationally: Brazilian items per buyer rose 16%, cost per shipment is 14% below its pre-threshold level, and lower-value free shipping has become contribution-positive. The network is delivering the efficiencies MercadoLibre intended.

• But competition is capturing more of the surplus: shipping economics have weakened, Brazil’s revenue growth barely translated into higher direct contribution, and Shopee now has enough scale and funding to force MercadoLibre to reinvest more of its efficiency gains into the customer proposition.

• The moat is migrating from logistics toward the integrated customer relationship: customers who shop and use Mercado Pago are substantially more valuable, while social commerce and AI increasingly threaten to own discovery. The key question is whether commerce, finance, loyalty and data together remain more valuable than the components rivals can match individually.

MercadoLibre’s third-quarter shareholder letter contains perhaps the most revealing statement in our five quarters of coverage:

Across Commerce and Fintech, the pattern is the same. Our users are doing more with us, and the cost of serving each of them is falling or holding flat. That efficiency creates a choice: how much to reinvest in the value proposition, and how much to let through to the bottom line. Today, we are choosing to invest, because the habits we are building are worth more than the margin we are forgoing in the short term to build them. That is the trade-off. As those habits set and the market matures, that balance will shift.

The choice sounds familiar. In August, we concluded that MercadoLibre’s integrated system was working and that the price of its flywheel was time: management was deliberately sacrificing present margins for future earnings power. We also established a condition under which that explanation would no longer suffice. Operating margins below 6.5%, accompanied by continuing investment requirements, would impair the proposition that management could restore profitability when it chose.

Third-quarter revenue rose 45% to $10.75 billion, but operating income declined 7% to $676 million, missing the $729 million FactSet consensus. Operating margin fell to 6.3%, its third consecutive quarterly decline. The stock initially dropped approximately 8% after hours.

The uncomfortable finding is that MercadoLibre’s operating advantages are improving exactly as expected. What has changed is our confidence that it can retain the economic benefits. Management still controls the intensity of investment, but competition increasingly determines the minimum. The margin dial has acquired a second owner.

The Warehouse Advantage Is Working. Its Economics Are Not.

In July 2025, we described MercadoLibre’s reduction of Brazil’s free-shipping threshold from R$79 to R$19 as the weaponization of logistics density. More shipments through an established network would lower unit costs, allowing MercadoLibre to offer economics that competitors could not easily match. Q3 provides compelling evidence that the mechanism works.

Brazilian items per buyer increased 16%, even after the threshold change had been fully lapped. Cost per shipment is 14% below its pre-threshold level, and in September, free shipments in the R$19–79 range became variable-contribution positive in aggregate. The infrastructure is generating the efficiencies we anticipated.

Yet net shipping margin has deteriorated by approximately 200 basis points of GMV since early 2025. Applied illustratively to this quarter’s $22.8 billion GMV, that represents about $456 million of quarterly economics transferred away from MercadoLibre. This is not recoverable profit: much of today’s additional volume might not exist without cheaper shipping. It nevertheless establishes the scale of the trade-off.

Brazil provides an even starker illustration. Revenue increased from $4.0 billion to $6.0 billion year over year, while direct contribution barely moved, from $475 million to $479 million. Across the company, $3.34 billion of additional revenue produced $48 million less operating income.

Some pressure is temporary. MercadoLibre expanded fulfillment capacity nearly 20% sequentially ahead of peak season, incurring costs before the associated volume arrives. But management acknowledges that continued growth requires continuing capacity additions. An individual warehouse can mature; a perpetually expanding network always contains another immature warehouse.

We have seen this story before. Between 2016 and 2019, MercadoLibre’s operating margin collapsed from roughly 23% to negative 7% as it subsidized shipping and built infrastructure. Investors ultimately funded the strategy because the incumbent Brazilian retailers lacked both its marketplace economics and the resources to respond. We were too quick to invoke that precedent when, in August 2025, we suggested competition had effectively ended.

Shopee is a different challenger. Its Brazilian distribution network has expanded dramatically, its VIP program exceeded one million members within months, and Sea can support investment using profits generated elsewhere. Shopee has not replicated MercadoLibre’s fulfillment density or merchant relationships, but it no longer needs to be dismissed as an underfunded competitor. It needs only to match enough of the proposition to prevent MercadoLibre from charging for its full advantage.

The distinction changes our thesis. MercadoLibre’s logistics network remains superior, but superiority no longer guarantees the ability to harvest the resulting surplus.

The Wallet and the Missing Denominator

If warehouses are increasingly contestable, MercadoLibre’s stronger advantage may lie in customers who both shop and bank with it.

Our Q2 article identified the ecosystemic user as the relevant economic unit. A subsidized delivery might be unprofitable in isolation but worthwhile if it produces a customer who subsequently saves, pays and borrows through Mercado Pago. Such users increased 33% in Q3, generating 70% more merchandise value than commerce-only customers, 90% more payment volume than fintech-only customers, and three to four times the revenue of single-service users.

These relationships are harder to reproduce than delivery capacity. Commerce supplies Mercado Pago with distribution and purchasing information; Mercado Pago returns financing, savings products and loyalty benefits that make commerce more attractive. MELI+ increasingly bundles these advantages. A competitor can build warehouses, but recreating years of purchase, repayment and engagement history is considerably harder.

The credit portfolio, approaching $18 billion, illustrates both the opportunity and its cost. Brazilian credit-card cohorts issued during 2021–2024 are profitable on a net-interest-margin-after-losses basis and represent more than half the country’s card portfolio. New issuance, however, initially requires provisions before meaningful interest income develops. Consolidated NIMAL declined from 20.7% to 20.1%, while loans overdue more than 90 days increased to 19.3%. The relatively stable 7.0% early-delinquency ratio does not eliminate that concern.

More importantly, management’s disclosures have become less useful for establishing the economics of integration. Q2 emphasized that ecosystemic users generated multiples of the combined contribution profit of single-service users. Q3 emphasized revenue multiples instead. We still lack penetration, acquisition costs and incremental contribution profit. Fintech’s roughly $1 monthly cost to serve excludes credit losses and funding costs. Nine-month adjusted free cash flow fell from $718 million to $443 million despite extraordinary revenue growth.

None of this invalidates the integrated relationship. It means we cannot yet capitalize engagement as though it were cash earnings.

And Shopee is following MercadoLibre beyond commerce. Its Monee financial-services platform, formerly ShopeePay, is expanding wallet and credit capabilities using transaction data and Brazil’s Open Finance infrastructure. That cannot instantly reproduce MercadoLibre’s proprietary history, but it makes the financial-services moat contestable too.

The question is whether shopping, payments, savings, credit and loyalty together create customer value that exceeds what competitors can offer through their individual products.

The War Moves Upstream

The competition is now moving toward the beginning of the transaction itself. Asked about social commerce, CEO Ariel Szarfsztejn was unusually candid:

On the one hand, we can do as much as we can or want to design the way we want to capture demand, but in the end, we need to be where the demand and the attention is, and affiliate marketing is part of that playbook. There’s some broader shift in how consumers discover and potentially buy some products. Social media in general is becoming an additional channel for discovery, many times incremental, and that happens whether we like it or not.

Affiliate-generated orders increased 240%, while sales and marketing rose 40 basis points sequentially to 11.6% of revenue. Management described the quarter’s spending intensity as tactical, but the channel itself as strategic.

This creates a second pressure on the flywheel. MercadoLibre may reduce the cost of completing a transaction through greater logistics density while spending more to attract that transaction in the first place. Shopee and social platforms do not need to replicate its entire infrastructure to claim part of the economic surplus.

AI agents introduce a more consequential choice. Szarfsztejn explained management’s position:

For now, we have decided not to provide our inventory to any external platform. Our priority is still to build the best agentic experience inside MercadoLibre, taking advantage of the 27 years of first-party data that we have, our unique combination of selection, pricing, logistics, financing, loyalty, and so on. E-commerce is much more than just search and find.

The distinction matters. An affiliate sends customers into MercadoLibre, where checkout, financing, delivery and future engagement remain inside its ecosystem. An external AI agent could control product comparison and purchasing decisions, reducing MercadoLibre to one inventory provider among several. The first arrangement raises customer-acquisition costs; the second potentially transfers customer ownership.

Management is defending the latter boundary, even at the risk of missing demand originating elsewhere. Its own shopping agent is available to a limited portion of users in its largest markets, but there is insufficient evidence to establish whether this strategy will succeed.

Mercado Ads offers a partial counterweight. By using purchase data to measure advertising effectiveness across social and streaming platforms, including Netflix, MercadoLibre can sell advertisers something external discovery channels lack: visibility into completed purchases. Whether commerce-media economics can offset rising distribution costs remains unproven.

The company’s most valuable asset may therefore be neither the warehouse nor the wallet individually, but a customer relationship that makes the combined offering preferable even when rivals match particular components.

What Is the Integrated Moat Worth?

Our earnings expectations must change. August’s published 2029 base case assumed operating margins of 10–12% and EPS of $105–120. With margins now at 6.3%, Shopee investing aggressively and customer-acquisition costs rising, that recovery looks too optimistic. I now expect margins around 7% in 2027 and 8.5% in 2029, closer to the profitability previously embedded in our downside case.

For a consistent three-year investment horizon, the scenarios below value FY2030 forward earnings at the end of 2029.

These are our estimates, not consensus. They start from approximately $41.8 billion of FY2026 revenue, assume $450 million in annual net non-operating expense, a normalized 29% tax rate and 50.7 million shares. The integrated case requires sustained commerce growth, credit maturation and a recovery in operating leverage; the dominant case additionally requires the customer relationship to preserve superior monetization despite competition. Cash conversion must support the multiples, particularly as credit growth absorbs capital.

The probability-weighted value is approximately $2,560, against the indicative $1,604 post-results price. That implies roughly 60% potential appreciation through end-2029, before discounting for time and uncertainty. These targets are not directly comparable with August’s because the valuation now uses FY2030 rather than FY2029 earnings. The underlying margin and earnings expectations have been reduced.

The investment case is attractive but no longer rests on assuming the margin dial can simply be turned upward. The next two quarters must establish whether Brazil’s contribution margin recovers as new capacity fills. During 2027, operating margins moving sustainably above 7.5%, improving cash conversion and stable credit quality would support our base case. Continued 6–7% margins despite strong growth would suggest the competitive investment burden is structural.

MercadoLibre won its first free-shipping war by building infrastructure competitors could not afford to reproduce. Shopee has made that advantage contestable, while social commerce and AI threaten to move customer decisions outside the marketplace. We were right that MercadoLibre could create extraordinary operational efficiencies, and wrong to assume those efficiencies would inevitably accrue to shareholders.

The next phase depends on whether the integrated relationship across commerce and finance remains worth more to customers than the individual components rivals can match. The infrastructure still works. What MercadoLibre must now prove is how much of the value it creates it can keep.

Subscribe now

• The first proven AI economics are internal: subscription gross margin expanded 320 basis points, with management explicitly citing AI-driven support automation as a contributor. The invisible manager is already working; its first customer is Toast itself.

• The bigger opportunity is customer profit, not another software module: Toast wants agents to coordinate marketing, staffing, inventory and margin management around a restaurant’s contribution profit. If that works, the payoff should show up in retention, payment volume, product adoption and possibly lower restaurant failure not just standalone AI subscription revenue.

• But value capture must be earned, not imposed: the FTC case puts contractual switching friction under scrutiny just as Toast asks investors to tolerate slower margin expansion and more reinvestment. The next phase only deserves a premium if restaurants stay because Toast makes them better businesses, not because leaving is difficult.

Toast and Block reported on the same evening and offered investors different futures. Block projected roughly 30% adjusted operating-income growth in 2027. Toast’s incoming CFO, Michael Senno, promised only “very modest margin progression.” Block shares climbed about 15% the following day; Toast rose 5.5%. The businesses and earnings measures are not comparable, but the contrast is instructive: one proposed harvesting more profit, the other reserving the right to spend it.

Toast earned that choice. Third-quarter recurring gross profit rose 27% to $642 million, above management’s $615–625 million guidance, while adjusted EBITDA reached $240 million, beating expectations near $220 million. Full-year EBITDA guidance increased to $840–850 million. In August, we argued that restaurants were starting to pay for the operating system rather than merely the payment terminal. Q3 reveals a more interesting distinction. The invisible manager is already producing results, but its first customer was Toast itself.

The Extraordinary Becomes Ordinary

The underlying thesis has not changed. The register brings Toast into the restaurant, payments make it economically relevant to every transaction, and software makes the relationship more difficult to replace. What has changed is the pace at which those mechanisms create incremental profit.

Payments ARR outgrew payment volume by seven percentage points in 2025, but only one point this year. Departing CFO Elena Gomez attributed improved payment economics to product adoption, processing efficiencies and “small targeted price moves,” qualifying our August claim that the benefit required no higher pricing. The extraordinary monetization spread is normalizing. So is the 80 cents of incremental EBITDA per dollar of gross profit produced during the earlier restructuring period. At 47 cents, leverage remains substantial: EBITDA grew 36% while recurring gross profit grew 27%.

We also need to correct two August judgments. We described three consecutive earnings beats followed by falling shares, but only the first-quarter 2026 report fits that description. And we characterized 2026 as a margin-sacrifice year; adjusted EBITDA will instead reach roughly 35.5% of recurring gross profit, about two points above 2025. Toast has been investing and expanding margins simultaneously. The question is whether its next investment cycle can deliver returns comparable to the gains already captured.

The Invisible Manager’s First Job

Gomez gave the clearest evidence of what AI has accomplished:

Subscription gross profit grew 36% and SaaS gross margin increased 320 basis points from a year ago. AI tools are handling a growing portion of support tickets, providing a better customer experience and are an important contributor to that margin expansion.

Subscription revenue climbed from $244 million to $317 million, while its reported cost of revenue stayed at $67 million. Not all the improvement belongs to AI, but management explicitly identifies automation as a contributor. The first visible agent economics are therefore not additional fees paid by restaurants; they are lower costs incurred by Toast to serve them. That matters because Toast’s thesis increasingly assumes it can automate work that customers currently buy from people and specialized software vendors.

CEO Aman Narang described the ambition:

We’re building towards a team of agents that drive results for our customers, crafting marketing campaigns, forecasting demand with staffing to match, flagging low inventory before an out of stock, and protecting margin from swings in costs... Restaurants already pay for marketing, scheduling and payroll, and bookkeeping services, often several times what they spend on software.

The opportunity is not a collection of digital assistants. A marketing campaign can fill tables while destroying profit if discounts, overtime and food costs exceed the extra revenue. Cutting labor can improve today’s margin while worsening service and future demand. Restaurants need to optimize contribution profit, not individual tasks. Toast IQ Grow can stimulate demand; IQ Team, launched this quarter, can align staffing with expected sales; and Toast’s transaction data can reveal what actually happened. Inventory and supplier information could eventually close the loop between demand and cost.

That is an attractive architecture, not yet a demonstrated profit engine. Grow is on course for roughly $10 million of annualized revenue, less than half a percent of Toast’s $2.5 billion ARR. Even $50 million at a hypothetical 40% margin and 16 times EBITDA would be worth only $320 million to a company valued near $19 billion. The prize lies in better retention, higher payment volume, additional product adoption and perhaps fewer restaurant failures. Narang says Toast customers process more volume and are less likely to close than average restaurants, but stronger operators may simply be more likely to adopt Toast.

The evidence needs similar care elsewhere. Management’s 41% reduction in overtime hours concerns users of established Scheduling Pro, not a controlled incremental benefit of IQ Team. The nearly 50 million Online Ordering accounts are not 50 million active Toast Local users. None of this invalidates the opportunity; it establishes how much of the customer-outcome thesis remains to be proved.

When the Moat Is Put on Trial

The October 29 lawsuit from the FTC and state and local authorities introduces an uncomfortable test. The complaint alleges hidden or unauthorized fees, unwanted add-ons, weak support, difficult cancellation, early termination penalties and loan terms that may impede switching. Toast disputes the allegations. Narang told investors:

We’ve been engaged with them and cooperating with them really over the past year, and fundamentally, believe their application of ROSCA is unprecedented for custom contracts in B2B, which is what we negotiate for every single customer.

His defense raises a legitimate question about applying a consumer-protection statute to negotiated business contracts, but the suit matters beyond potential penalties. We have argued that Toast’s moat lies less in hardware than in operational muscle memory: once ordering, staffing, payroll and reporting depend on the platform, replacing it is disruptive. Contractual penalties are another kind of switching cost. If some apparent retention reflects cancellation friction rather than operational value, the quality of that moat is lower than we assumed. A remedy could also change the economics of selling new agents into the installed base.

The support allegation must not be treated as proof that AI-driven efficiencies compromised service; no such causal link has been established. Nonetheless, it creates a test of whether the 320-basis-point SaaS margin improvement is compatible with the customer experience Toast promises. The stronger Toast’s product-driven retention, the less its economics should depend on disputed contractual restrictions.

The Right to Reinvest

The larger financial tension is between earnings and cash. Third-quarter free cash flow was $181 million, 75% of EBITDA, but through September it was $426 million versus $430 million a year earlier, even as nine-month EBITDA climbed from $470 million to $640 million. Inventory built ahead of hardware and memory costs is a material reason: it reached $244 million, up from $114 million at year-end. That cash may return when stocking normalizes, but shareholders cannot spend EBITDA.

Toast Capital creates a second distinction. Credit-related costs reached $35 million, above our earlier $30 million comfort line, while Toast now carries approximately $70 million of loans directly. Retaining loan economics can improve reported profitability while tying up capital and adding credit risk. Horizon 2, enterprise, international and retail, offers more encouraging evidence of reuse: management expects about $200 million of ARR with 25% less direct R&D than the core required at that scale. That proves some engineering leverage, not lower total selling, onboarding or servicing costs.

Senno framed the coming trade-off explicitly:

And the pace of margin expansion over the past couple of years has actually been faster than we had anticipated. We’re sitting on what we believe is a generational opportunity... And as a result, we expect very modest margin progression next year.

This is a defensible choice if Toast can translate its existing transaction, staffing and customer data into outcomes independent point solutions cannot reproduce. Core margins already exceed 40%; management is choosing to direct incremental capacity toward new markets and agents. But August’s warning remains appropriate: “Investment is not the conclusion. Return on investment is.” The departure of Gomez and the planned retirement of the chief revenue officer add execution risk just as the company asks investors to tolerate slower near-term margin expansion.

Better Business, Fuller Price

The valuation should not capitalize pass-through payment revenue or treat adjusted EBITDA as cash available to shareholders. I would retain August’s three economic outcomes but value FY2030 owner earnings, free cash flow after cash taxes, less stock compensation, at end-2029. Starting from roughly $2.38 billion in FY2026 recurring gross profit, the model allows tax losses to run down, loan funding to consume cash, and buybacks principally to offset dilution.

These are our scenarios, not management guidance. With approximately 590–595 million diluted shares and net cash added only once, the probabilities imply about $50 per share, roughly 41% above the November 4 price of $35.37, or around 12% annualized to end-2029. That is less compelling than August’s $53 weighted value from a $32 stock price. The company has strengthened, but the stock is fuller.

The margin trap assumes slowing location additions, weaker monetization and regulatory or credit friction. The quiet compounder depends on roughly 34,000–37,000 annual net additions, improving profit per location and disciplined reinvestment. The expanding-wallet case requires agents to improve merchant economics sufficiently to increase retention and share of wallet, not merely sell another subscription. The base valuation also assumes that Toast can sustain a premium for future owner earnings; attractive growth alone does not justify any multiple.

The next evidence should be concrete. SaaS ARR growth below 26%, after 28% in Q3, would weaken the software thesis. Cash conversion should recover as inventory normalizes; persistent conversion below 70% would force a lower owner-earnings estimate. The FTC outcome will test contractual versus operational switching costs. Most important, management needs to show whether Grow and Team produce incremental merchant profit and whether Toast can capture part of it without proportionate support and acquisition expense.

The register was the wedge, the operating system became the moat, and the invisible manager is the next bet. Q3 demonstrated that AI already helps Toast run itself more efficiently. The next and far larger opportunity is to make restaurants better businesses because they run on Toast, and to retain a meaningful share of that improvement on terms restaurants freely choose.

Subscribe now

• The core consumer business is still improving: room nights grew 7%, Connected Trip activity expanded, merchant penetration increased, and B2C marketing showed leverage. The original “marketing treadmill” concern has not been triggered.

• But Booking increasingly operates through two different doors: direct consumer distribution creates the possibility of cheaper repeat demand, while B2B distribution creates profitable volume but leaves the partner owning the customer. AI agents could accelerate both models.

• Genius is becoming more strategically important than a loyalty program: if an external agent recognizes and values a traveler’s Genius benefits, Booking can remain the preferred supplier even when it no longer owns the interface. That is the clearest path to retaining bargaining power in an agent-mediated world.

From Booking Holdings’ third-quarter earnings call, CFO Ewout Steenbergen:

Marketing expense increased 11% year-over-year, slightly faster than gross bookings, driven primarily by faster growth in B2B affiliate marketing expenses. Excluding our B2B affiliate business, our B2C business had marketing leverage in the quarter as we continue to aim to grow our topline metrics faster than marketing investments.

In August’s Fewer Nights, More Money, I argued that Booking had become better at extracting value from each trip but had not demonstrated that completing one trip made the next customer cheaper to acquire. Steenbergen’s distinction changes how to test that proposition. Booking is increasingly operating two distribution businesses: one in which it attracts and retains the traveler, and another in which a partner owns the relationship while Booking provides inventory and fulfillment. AI agents could make both businesses larger. The investment question is which one gets paid for the second booking.

Revisiting August’s Test

August established a measurable threshold: marketing above 5% of gross bookings in any of the next four quarters would force a reassessment of the thesis; a quarter below 4.5% would suggest improving customer-acquisition economics. Q3 came in at 4.75%, versus 4.71% a year earlier. Neither condition was met, but the original interpretation needs refinement.

The underlying quarter was strong. Room nights rose 7% to 345 million, accelerating from 5% in Q2 and beating management’s 3–5% guidance. Gross bookings increased 10% to $54.5 billion, adjusted revenue 9% to $9.79 billion, and adjusted EBITDA 9% to $4.61 billion. Adjusted EPS grew 17%, partly because repurchases amplified adjusted net-income growth of 8%. The reported revenue decline reflected a $900 million regulatory adjustment, not falling transaction volume.

Still, more activity does not automatically mean richer economics. Merchant bookings reached almost 75% of gross bookings and Connected Trip transactions grew nearly 20%, yet adjusted revenue as a share of bookings slipped from 18.14% to 17.98%, reflecting flight mix and booking-versus-travel timing, partly offset by payments. I also overstated the earlier marketing deterioration: shifting merchandising discounts from contra-revenue to performance marketing can raise both reported revenue and marketing expense without increasing profit. Part of August’s apparent spread and treadmill was the same spending decision. The better evidence this quarter is B2C marketing leverage, not an unqualified claim that each transaction has become more lucrative.

The Owned Door and the Rented Door

Steenbergen attributed B2C leverage to improving returns on paid channels and increasing direct traffic. But B2C direct mix remained in the mid-60s. Booking is getting more direct traffic without yet demonstrating a sustained increase in its share of consumer bookings, much less disclosing the cost of an incremental repeat customer. The other part of the story is more revealing. Asked about B2B’s increasing importance, Steenbergen explained:

...that is a business that is getting more focus and attention internally. I think, you know about our plans that we’re launching on January 1, a combined B2B business. So, you will also see then more details about the numbers, because most likely we’ll break that out going forward. What is happening there is if the business is growing and they are reporting more revenues, usually the commission structures are based on a rev share. So, therefore also the rev share and the commissions are going up, and that is flowing through the marketing line.

Booking’s combined B2B business begins operating in January 2027, with additional disclosure likely, though not promised. The economics are straightforward: a partner brings demand, Booking supplies inventory and execution, and the partner receives a share of revenue. That can be attractive incremental business, as management insists, but the partner, not Booking, may control the next transaction. If the rounded direct-mix figures for B2C and total room nights are comparable, B2B already represents roughly one room night in seven. This is an inference, not a disclosed figure.

In the consumer business, loyalty can reduce reliance on paid distribution. In B2B, volume can grow even while commission costs remain attached to every transaction. Consolidated marketing expense blends the two, obscuring whether Booking’s customer relationship is becoming stronger or simply being rented more efficiently. It is therefore possible for the company to become indispensable to travel fulfillment while having less control over how travelers reach it.

When the Agent Chooses

The supply side favors Booking. The ten largest hotel chains represent only a low-double-digit percentage of Booking.com room nights; the remainder includes fragmented suppliers that cannot maintain commercial and technical integrations with every new AI interface. Booking already assembles availability, pricing, payments, service and reliability at global scale. An agent that wants to complete a trip, rather than merely propose one, has strong reasons to connect to it.

But useful infrastructure need not command attractive distribution economics. An agent can refer the customer to Booking, leaving checkout and repeat engagement largely intact. It can instead own the conversation and checkout while Booking remains responsible for fulfillment. Or it can treat Booking as wholesale inventory and control the customer outright. In the first arrangement Booking may retain the relationship; in the others the intermediary can demand a lasting share of its economics. Expedia’s integration with Meta’s Muse illustrates the second route, while Booking’s existing B2B business illustrates the third.

We should not confuse strategic possibility with present financial impact. LLM-originated room nights remain significantly below 1% at Booking.com. Management’s reported roughly twofold conversion for AI-assistant users is encouraging but cannot establish incremental profitability, because those users may be different to begin with. The agents’ importance lies less in today’s bookings than in the bargaining structure they could create as distribution migrates.

Genius as a Reason to Choose Booking

Booking has an answer that is more interesting than simply protecting its website. When discussing Genius, CEO Glenn Fogel described a program whose benefits are often funded by hotels because those hotels receive attractive returns:

We are going to improve Genius. It’s going to be an even better program down the road... Genius works right now very well because the partner is funding a lot of it, and it’s getting a good ROI out of that. And of course, the traveler’s getting a great deal too, because again, a lower cost for their holiday. So, it works both ways. But there are a lot of other ways we can improve upon that tremendously. There are other things that travelers want, and maybe the partner will be willing to give it up either at no cost because there’s very little marginal cost to them, or perhaps they’d be willing to do it at a lesser cost that we can help work with.

The economics deserve attention. A late checkout, flexible cancellation privilege or upgrade can be worth considerably more to a traveler than it costs the hotel to provide. Booking can coordinate such benefits across thousands of suppliers. Genius Levels 2 and 3 already represent more than 30% of active Booking.com customers and a high-50s share of room nights, concentrating travelers who are especially valuable to hotels. That concentration is not evidence that Genius caused their loyalty, but it gives Booking a network of entitlements to build on.

Now suppose a traveler asks an independent agent to book a hotel. If the agent can recognize the traveler’s Genius benefits, Booking’s offer may be preferable even when another supplier provides the same room at the same headline price. The agent owns the conversation, but Booking supplies a reason for the agent to choose it. Fogel’s discussion of extending Genius to travel and non-travel partners, including in response to a question about loyalty when an external agent controls the interface, suggests management understands the opportunity.

This remains an inference, not an announced capability. Booking has not shown that Genius benefits will be portable to third-party agents, that they change agent selection, or that the resulting commissions leave better margins. Hotels can fund rival programs, and agents may develop loyalty systems of their own. Nevertheless, Genius offers a plausible route between two extremes: controlling every customer interaction and becoming interchangeable fulfillment infrastructure. Booking might lose the interface without surrendering all the bargaining power behind it.

What Is the Second Booking Worth?

The near-term evidence argues against complacency. Q4 guidance implies only 1–3% room-night growth and roughly flat adjusted EBITDA, against prior Street expectations of substantially higher profit growth. International demand has softened, and Expedia’s cautious outlook suggests a common travel headwind rather than proof that agents are already taking share. Booking is also reinvesting in AI and adjacent travel products, so its $650 million transformation program cannot simply be treated as permanent margin expansion.

Regulation adds risk. A $900 million revenue reduction and a separate $100 million accrual relate to an FTC matter involving Priceline and a B2B affiliate; additional costs remain possible. Repurchases also fell from $3.7 billion in Q2 to $2.1 billion in Q3, undermining my earlier treatment of buybacks as an unambiguous management signal. Neither observation disproves the franchise’s strength, but both belong in the underwriting.

At approximately $171, Booking traded near 14 times revised FY2027 consensus EPS of $12.25. The valuation looks more consistent with a future in which intermediaries take an increasing share of transaction economics than one in which Booking retains meaningful customer preference. Our base case does not require beating the limited FY2029 consensus, which is near $17 per share. It requires the franchise to preserve its earnings power and recover a more appropriate multiple.

The scenarios start with about $29.1 billion in FY2026 adjusted revenue and 750 million diluted shares, use net share reductions consistent with the table, and allow approximately $1 billion annually for depreciation, net financing and other items below EBITDA. Normalized tax rates are 21–22%. The Durable case assumes roughly 5–6% room-night growth, additional pricing and payments contribution, and enough cash generation to fund repurchases. Challenged reflects weaker demand and greater distribution tolls; Preferred requires stronger growth, retention economics and margins. At end-2029, their probability-weighted value is approximately $257, about 50% above $171 before dividends. The $274 base outcome depends on both EPS growth and a rerating from 14 to 17 times; the upside is not purely an earnings story.

These outcomes are below August’s published ranges of $290–320 for the base case and $400–460 for the bull case. The reduction reflects greater weight on distribution costs, regulatory exposure and the uncertainty of an agent-driven rerating, rather than a newly discovered weakness in Booking’s operating franchise.

What Would Change My Mind

The next results should distinguish travel weakness from deteriorating economics. I want B2C marketing leverage to persist, 2027 EBITDA growth to recover relative to revenue, and any new B2B disclosure to show whether affiliate growth brings attractive contribution rather than merely purchased volume. August’s 5% consolidated marketing test remains in force through Q2 2027, but channel mix must inform its interpretation. More importantly, agent agreements should reveal who controls checkout and how commissions work; Genius integrations must demonstrate actual benefit recognition, incremental repeat business and profitable selection, not simply partnership announcements.

I would retain August’s conditional willingness to add toward $175 and add around $171, with sizing that acknowledges the $133 downside scenario. Booking has not proved it can preserve the customer relationship when an agent handles the conversation, but neither has the quarter supplied evidence that its role is being commoditized. The company already makes the first booking work. The investment opportunity is that its supply, execution and loyalty advantages may also give the next agent a reason to choose Booking and leave Booking with a meaningful share of the economics of the second booking.

Subscribe now

• The product thesis is stronger than before: Palantir is no longer just coordinating intelligence from external models; it is helping customers post-train open models on proprietary operational data, making the model itself more interchangeable while increasing the value of the Ontology and the surrounding learning system.

• The commercial model is becoming less contract-heavy: U.S. commercial revenue is accelerating far faster than remaining deal value, and shorter initial contracts are becoming part of the sales motion. Management argues this reflects faster proof of value followed by expansion; the missing duration-adjusted bookings disclosure makes that claim harder to verify.

• The real moat may be reconstruction cost, not contractual duration: if replacing Palantir means rebuilding permissions, workflows, evaluations and feedback loops, customers can remain economically locked in while still owning their models and signing shorter contracts. But at $178, investors need evidence that this dependence reliably becomes future spending.

In February, Palantir’s appeal rested on a useful separation of responsibilities. Model laboratories would spend billions producing intelligence; Palantir would make that intelligence useful inside institutions. Our description was deliberately blunt:

Palantir doesn’t train models. They don’t care which model wins. They sit above the model layer, buying cognition at declining prices and selling coordination at increasing prices.

The economic argument still holds, but the product description needs updating. Palantir does not build general-purpose foundation models; it increasingly helps customers post-train open models on their own operational information. We identified this model factory in August. What Q3 adds is its growing importance to both the product proposition and the way customers buy it.

The operating results were extraordinary. Revenue grew 96% to $2.31 billion, U.S. commercial revenue rose 148% to $984 million, and adjusted operating margins remained at 62%. Full-year revenue guidance increased by nearly $380 million to $8.53 billion. U.S. commercial revenue exceeded U.S. government revenue for the first time.

Yet August’s description of Palantir as “converting backlog without emptying it” now requires qualification. U.S. commercial bookings fell 31% sequentially, remaining deal value grew just 4%, and long-term remaining performance obligations declined. The issue is not that the commercial order book shrank it did not, but that forward commitments grew much more slowly than current revenue. The business looks stronger; the evidence supporting its duration has become less straightforward.

The Compiler Now Trains

Palantir’s product argument begins with the workflow, not the individual worker. A procurement manager who analyzes prices faster creates limited value if the resulting decisions remain disconnected from production, inventories and logistics. The Ontology is intended to connect those activities; the model factory adds a learning loop that improves the intelligence operating within them.

CTO Shyam Sankar described the mechanism on the call. His explanation places the durable value in a continuously renewed process rather than a particular model:

It is clear that post-training on your specific operational data in Ontology enables better-than-frontier performance on weights you control, protecting your alpha. Yes, it is cheaper, but the dominant variable is that more use cases can actually get to the threshold of viability for machine intelligence to generate economic value when elegantly integrated into your enterprise. Our product investments in AIP to create a closed-loop model factory are making this easy and fast. The specific post-trained model is a commodity item. The compounding value is in the post-training pipeline that is constantly fed from your operational exhaust, captured in the Ontology, that drives automated post-training runs, model checkpoints, evaluations, and deployment across yesterday’s, today’s, and tomorrow’s open-weight ecosystem.

“Better-than-frontier” is management’s claim about performance on particular operational tasks, not evidence that Palantir has built a superior general-purpose model. Nevertheless, the strategic implication is important. Palantir no longer merely benefits when model laboratories reduce the price of intelligence; it can help customers create task-specific alternatives that make the underlying models more interchangeable.

NVIDIA’s supply-chain application illustrates the intended mechanism, using operational information to improve specialized Nemotron models. AIP Evolve is designed to select and optimize models against the customer’s desired accuracy, latency and cost. Palantir therefore still does not need a single model provider to win. Its opportunity is to make the system that chooses, adapts and deploys models more valuable than any individual model passing through it.

This also makes Karp’s description of the laboratories as “symbiotic” less surprising. They establish that AI matters; Palantir sells to institutions discovering that general intelligence does not automatically produce useful business outcomes. The laboratories can expand Palantir’s addressable market without owning the operational system through which that intelligence becomes valuable.

A Faster Sale, a Shorter Commitment

A more demonstrable product can change the sales process. Instead of negotiating a long commitment before showing an outcome, Palantir can demonstrate value, secure a shorter initial contract and expand once the customer sees the results. An industrial materials company moved from a roughly $3 million pilot to a $29 million annual contract commitment within two quarters. That is a selected example, not a disclosed conversion rate, but it illustrates management’s proposed mechanism.

The aggregate evidence is less conclusive. Across selected quarters, revenue has accelerated much faster than the remaining commercial contract base:

These measures need care: total contract value includes options, and remaining deal value is not guaranteed revenue. Nevertheless, revenue’s 29% sequential increase against roughly 4% growth in remaining deal value raises a legitimate question about forward visibility.

CFO Dave Glazer addressed the contract change directly. His explanation was that customers were beginning relationships differently, rather than becoming less interested in the product:

We signed multiple eight-figure deals in the quarter that were shorter in duration than usual as customers rapidly adopted initial Sovereign AI capabilities in their enterprise. [...] As we continue to deliver value with Sovereign AI, over time, we expect both the duration and size of these contracts to increase.

Shorter contracts can produce lower bookings without lower annualized spending. Certain contracts with original terms of twelve months or less are also excluded from reported remaining performance obligations. Glazer’s comments therefore offer a plausible explanation, although several unusually short contracts do not establish a company-wide reduction in average duration.

The problem is that Palantir omitted the disclosure best suited to testing that explanation. In Q2, U.S. commercial bookings grew 153% year over year, but 271% on a dollar-weighted duration basis. In Q3, when duration became central to management’s explanation, the release continued to define that measure without reporting it. That is not proof of deteriorating demand, but it makes the distinction between faster adoption and weaker forward visibility unnecessarily difficult to establish.

Sovereignty Cuts Both Ways

The sovereignty proposition contains a tension we identified in August. Customers are being told to own their intelligence, retain their proprietary knowledge and avoid dependence on model providers. Yet the more operational logic they encode into the Ontology, the more difficult replacing Palantir could become.

Karp’s explanation is that demonstrated value resolves that tension. He described a commercial process in which operational usefulness develops faster than contractual duration:

What Ryan and I are seeing on the front line in commercial is this much quicker interaction, much quicker to deployment, shorter duration of the contract, but then an ability to expand that contract dramatically. Because as soon as you go in there and you show them that you can, for example, train open-weight models in our Ontology, those models are valuable, and you have beyond Frontier capacity without losing the alpha, you just are hooked. What we have to do internally is figure out how to re-resource our company to focus exclusively on those people, because we don’t have even the bandwidth to do that.

There is another possible interpretation: customers persuaded of the importance of independence may want to preserve flexibility with Palantir too. That remains our hypothesis, not an observed explanation for shorter contracts. Management’s faster-adoption account has direct support from the call; both explanations could nevertheless coexist.

The strongest version of this business does not need contractual protection because replacement is economically irrational. Leaving would mean rebuilding the permissions, workflows, evaluations and feedback loops that make customer-owned models useful. But ownership of those models does not automatically give Palantir unlimited pricing power. It must remain the best place to operate the learning system, not merely the place where the system began.

Two different things must therefore compound. The customer’s operational knowledge must improve, while Palantir’s ability to deploy and extend the platform must become more repeatable. Strong growth and profitability are encouraging, but Karp’s capacity constraint leaves distribution unfinished. Microsoft and ServiceNow do not need to reproduce every element of the Ontology; becoming sufficiently capable for enough ordinary workflows could limit Palantir’s opportunity. Partner-led deployments delivering comparable outcomes and economics would help answer that challenge.

Net dollar retention of 165% supports the case that existing relationships become more valuable. It does not settle what today’s shorter contracts will become, because it is a trailing-twelve-month measure. The investment question is whether operational dependence consistently converts into sustained spending as both customers and competing platforms become more capable.

What the Stock Already Knows

The revenue outlook remains more attractive than the contract debate might suggest. Using the $12.4 billion consensus cited in our materials, and assuming Q4 revenue reaches approximately $2.83 billion, 2027 needs only about 4% sequential growth per quarter. My expectation remains $13–14 billion, although the Q4 starting point is an estimate rather than guidance. Revenue can beat expectations even while the durability of the relationships generating it remains imperfectly visible.

Valuation also requires correcting our earlier treatment of earnings quality. GAAP and adjusted EPS both reached $0.44, but unusually low GAAP taxation helped produce that equivalence, while stock-based compensation was $319 million. May’s assertion that the earnings-quality debate was effectively settled was too strong. The operating profitability is real; favorable tax accounting does not eliminate the economic cost of equity compensation.

I retain August’s approach: value the shares at the end of 2029 using FY2030 earnings. The revenue starting point increases, the projected 2030 share count is lower than in August, and the growth trajectories and terminal multiples remain unchanged. However, applying Palantir’s 23% normalized tax rate reduces the base and upside net margins. I also shift five percentage points of probability toward the downside scenario because forward visibility weakened, not because the operating thesis reversed.

These scenarios assume approximately 2.6–2.7 billion diluted shares and add no separate cash premium. Even the downside case assumes 26% annual revenue growth and a 35% net margin. The risk is not confined to business failure; an excellent business can produce a poor investment when its valuation becomes less exceptional.

The probability-weighted end-2029 price is approximately $203. Better operating prospects are largely offset by the tax correction and revised scenario weights, while the reference share price rose 42%, from $125.65 to the November 2 pre-earnings close of $178.34. This does not establish that the entire share-price increase was multiple expansion. It does establish that the return available under our own framework has narrowed considerably.

What February Must Show

Management’s guidance implies U.S. commercial revenue above $1.216 billion in Q4, but another revenue beat would not answer every question. I would look for remaining deal value above $7 billion, the restoration of duration-adjusted bookings disclosure, and bookings above 1.6 times quarterly commercial revenue. Missing $7 billion would warrant reassessment alongside contract duration and expansion evidence, not automatically invalidate management’s explanation. Adjusted gross margin, down from 88% to 85% over two quarters, also deserves attention as the product and delivery model evolve.

Palantir has a credible opportunity to become more valuable as the models beneath it become more interchangeable. The strongest evidence is the combination of rapid growth, expanding customer relationships and extraordinary operating profitability; the unresolved issue is how reliably those relationships translate into future economic commitments. Customers need not surrender ownership of their intelligence for Palantir to become indispensable. But at this valuation, investors need evidence that customers will keep choosing it not simply evidence that the technology works.

Subscribe now

• The CPU recovery is real: DCAI revenue rose 78%, operating margin reached 41.5%, Granite Rapids is margin-accretive, and capacity not demand is now the constraint in Intel’s server business. Pasted markdown

• But vertical integration is still economically misaligned: Intel 3 is supply-constrained while 18A has underutilized capacity, forcing Intel to reallocate manufacturing toward the products customers want today. Node-level progress is real, but progress on one process does not automatically improve utilization or returns on another. Pasted markdown

• The foundry thesis remains unproven: manufacturing capability is improving, but external foundry revenue is still small, packaging bottlenecks remain, and 14A requires substantial capital before external customer commitments are validated. The investment case now depends on synchronizing demand, process technology and capital, not merely proving that Intel can build competitive chips again.

“What we’re taking offline to put Intel 3 into Oregon is 18A wafers, which we’re generally going to build Panther Lake products. We see the opportunity with some of that capacity freed up to go produce more data center wafers.”

- CFO David Zinsner, Intel Q3 2026 earnings call

I think that was the most revealing passage in Intel’s third-quarter earnings call. Intel has more demand than it can satisfy for its older Intel 3 process, and underused capacity on its newer 18A process. It is rearranging production to make the chips customers want today while funding the technologies they may want tomorrow.

The headline numbers obscured that predicament. Revenue increased 21% to $16.5 billion, non-GAAP earnings reached $0.52 per share, and the Data Center and AI business grew 78%. Yet fourth-quarter guidance of $15.0–16.0 billion was $1.6 billion below pre-earnings FactSet consensus at the midpoint, and guided gross margin fell from 44.1% to 38%. The selloff was understandable. The puzzle is why extraordinary server demand produces such an uneven earnings trajectory.

In April, I argued that Intel’s CPU thesis had arrived while its foundry thesis had not. In July, the argument advanced: TSMC’s capacity constraints and the complexity of AI systems were creating a market for a credible second manufacturing platform. Q3 doesn’t overturn either conclusion. It exposes the missing link between them: demand for Intel processors does not necessarily fill the factories Intel needs to justify.

The CPU Clock Accelerates

Intel’s CPU revival is no longer a hypothesis. DCAI revenue rose from $4.1 billion a year ago to $7.3 billion, while operating profit increased from $964 million to $3.04 billion. Roughly 65% of incremental annual revenue became incremental operating profit. Granite Rapids, Intel’s main server family, is now gross-margin accretive; capacity, rather than customer appetite, is the immediate constraint.

Lip-Bu Tan has given this recovery a much larger destination: he estimates that agentic workloads could expand the server CPU market from $40 billion to more than $200 billion in five years. The mechanism is plausible. GPUs handle accelerated reasoning; CPUs handle orchestration, tool execution and persistent workflows. As AI moves from answers to actions, the work surrounding each model call may expand.

Still, an expanding CPU market is not an Intel-specific moat. AMD competes within x86; cloud providers are developing Arm processors; Nvidia is selling Vera CPUs specifically for agentic workloads. Oracle illustrated the industry’s actual buying behavior in October: it announced a planned multiyear Xeon collaboration with Intel while also making Nvidia’s Vera CPUs available. Large customers want supply, flexibility and negotiating leverage, not necessarily allegiance to one architecture.

It matters inside DCAI, too. Intel’s purpose-built silicon revenue grew more than 50% sequentially and more than doubled year over year, but management acknowledged gross margins below the corporate average. Zinsner said he and Tan had decided to focus on growing absolute gross profit, even if the mix pressures margins. I agree with that priority: selling more economically valuable silicon is preferable to protecting a percentage. But Intel has not disclosed enough about custom-silicon revenue, associated design expenses or capital requirements to prove its returns will match Xeon’s.

The CPU clock is moving faster than anyone expected. What matters is how much Intel captures, and at what cost.

The Wrong Capacity Is Constrained

Intel spent years demonstrating the downside of vertical integration. Delayed manufacturing nodes weakened products; weaker products reduced fab loading; poor utilization raised costs further. Our July article argued that the cycle was beginning to reverse: better CPUs would fill factories, improve yields, lower unit costs and provide cash for the next process.

Q3 shows why this loop is more complicated than it looks on a diagram. Granite Rapids uses Intel 3 and sells as quickly as Intel can produce it. Panther Lake and other newer client products use 18A, where yields are improving but weaker PC demand leaves some capacity underutilized. Intel is therefore preparing to put Intel 3 into Oregon by taking some 18A capacity offline. It is a rational allocation decision, and a reminder that manufacturing learning, tooling and demand do not transfer effortlessly between nodes.

Node-level improvement is real. Intel 3 manufacturing costs have declined about 15% since Q1; 18A costs are down about 40%. Nearly one-third of client CPU revenue now comes from 18A, up from less than 5% in Q1. Yet progress on one node doesn’t fill another, nor prove external foundry economics.

The PC downturn makes the mismatch worse. AI infrastructure spending is contributing to memory shortages and higher memory prices, which are making conventional PCs more expensive just as OEMs liquidate inventory. Intel expects PC industry units to decline roughly 12% in 2026 and another 10% in 2027. Q4 client revenue is expected to fall by more than 20% sequentially, overwhelming another projected mid-teens increase in DCAI. The AI boom is simultaneously filling Intel’s server order book and emptying part of its client production schedule.

The resulting earnings leverage works both ways. At Q4 guidance midpoint, $15.5 billion of revenue and 38% adjusted gross margin imply $5.9 billion of gross profit. Deduct the guided $4.4 billion of adjusted operating expenses and operating profit falls to roughly $1.5 billion, from $3.2 billion in Q3. The Israel fab incident explains about $300 million of revenue and 200 basis points of margin pressure; it does not explain the remaining collapse in profitability.

Intel’s planned transition to 18A-P server products, including Diamond Rapids, may eventually bring demand and advanced-node capacity back into alignment. But much of the incremental Intel 3 and substrate supply is not expected until the second half of 2027. Intel is investing to exploit demand already visible while relying on future products to fill newer assets. The risk is not simply a late technology roadmap; it is that several correctly executed roadmaps still produce poor consolidated returns if their timing is wrong.

A Second Source Is Not Yet a Second Business

Our July argument was that Intel need not beat TSMC to find customers. A hyperscaler facing insufficient allocation at the world’s dominant foundry may value a credible alternative even if moving a design creates additional cost and risk. AI makes that case stronger because bottlenecks increasingly sit between chips, in substrates, memory interfaces, packaging and power delivery, not merely in wafer supply.

Q3 validates the existence of the bottlenecks, but not yet Intel’s ability to monetize them externally. Advanced substrates are restricting shipments of Intel’s own profitable server processors. Tan is working with suppliers to expand capacity; management expects meaningful relief mainly in late 2027. Intel is also improving EMIB and EMIB-T packaging and discussing external customer opportunities. Turning existing good dies into saleable processors may create value faster than adding wafer capacity.

The foundry evidence is simultaneously encouraging and incomplete. Intel delivered its 14A 0.9 process design kit on schedule, reported lower costs and better yields on 18A, and described external interest in 18A-P. But external foundry revenue was just $411 million in Q3; the $118 million sequential increase was primarily attributable to a non-recurring program. Most of the $6.7 billion of reported foundry segment revenue represents internal transactions. Manufacturing for Intel is becoming more capable; the merchant-foundry business remains largely unproven.

Nor is the second-source opening permanent. TSMC is expanding leading-edge capacity over 2027–28, and customers choosing an alternative foundry must still fund design-porting, qualification and ecosystem work. Intel needs contracts that justify those costs and its capital commitments. An interested customer, a PDK milestone and a guaranteed wafer purchase are three different forms of evidence.

Tan’s ambition to offer modular x86 IP for customer-specific designs, and to manufacture CPUs regardless of architecture, is strategically more interesting than a simple “Western TSMC” pitch. Intel could supply processing IP, custom silicon, packaging and manufacturing within systems it does not control. But Oracle’s simultaneous embrace of Xeon and Vera illustrates the challenge: being a useful component supplier does not mean owning the architecture or the customer. Breadth becomes an advantage only when customers pay enough for the combination.

The Price of Time

Zinsner’s clearest financial statement was that Intel probably will not produce positive free cash flow in 2027. Gross capex is expected to be around $20 billion in 2026 and rise substantially next year as Intel adds equipment for Intel 3, 18A and 14A. The $23 billion equity offering has bought time, but it has also increased the shares over which any eventual success must be divided.

I don’t regard negative free cash flow during an investment cycle as inherently bad. Intel’s capital spending on profitable, supply-constrained server processors may earn very attractive returns. The economics of expanding substrates and packaging may also be compelling. A new 14A production network built ahead of signed external volumes is a different risk. Management’s aspiration of 20% returns on node investments needs to be tested node by node, not averaged across the company.

That is the change Q3 makes to our two-clock thesis. The CPU clock is running exceptionally fast, but Intel does not have one foundry clock. It has several: Intel 3 is constrained, 18A is improving but partly underloaded, and 14A is consuming capital ahead of commercial validation. External funding can keep all of them running; it cannot make them economically synchronized.

The selloff improves potential returns but cannot settle the underwriting question. At around $72, $4.50 of sustainable 2029 earnings valued at 25 times would imply roughly $113 a share; $3 valued at 20 times would imply $60. Those outcomes require very different assumptions about server share, process utilization, external contracts and future dilution. Neither follows automatically from a $200 billion CPU market.

I am more convinced than I was in April that Intel has a valuable, recovering product franchise, and more convinced than I was in July that its manufacturing engineers are making tangible progress. What I am less willing to assume is that those successes combine seamlessly into one reinforcing business. The next evidence is not another CPU-demand forecast. It is whether Intel 3 capacity starts shipping, 18A-P earns good margins, packaging commitments become revenue, and 14A attracts customers prepared to pay for the factories required.

Intel has proved that customers want its chips again. It now must prove that the right factories can make them at the right time and at a return that belongs to shareholders.

Subscribe now

• The operating model is changing in the right direction: Sandisk realized less price than peers, shipped more bits, and still expanded margins. That is exactly what a more contracted, supply-disciplined NAND business should begin to look like.

• The important shift is from cyclicality toward a floor: management is refusing to accelerate bit growth despite demand exceeding supply, while multi-year commitments and fixed or semi-fixed pricing reduce dependence on spot NAND. The hard-drive industry shows how valuable that discipline can become if it persists.

• But duration is still the unresolved variable: no new contract closed this quarter, prepayments slowed, and management is talking about customer “signals” beyond FY30 rather than signed commitments. The rerating case depends on extending contracted coverage beyond FY28 on attractive terms.

The Most Valuable Word on the Call

MST’s David Gibson asked whether Sandisk would speed up its move to newer chip generations to meet demand. David Goeckeler did not hedge:

No, we’re committed to the bit growth we talked about, mid to high teens. Those nodal transitions require a lot of planning and tool purchases and where lines are set up. So, we have that fab plan set up for the kind of bit growth we’ve been talking about for a long time. So, no plans to change that at this point.

For twenty years, every NAND cycle ended when someone said yes to that question and flooded the market. Sandisk said no, in a quarter when its CFO said the customer pipeline “exceeds our supply throughout the end of the decade.”

The quarter was strong: revenue of $11.03 billion, up 23%, an 85.1% gross margin, EPS of $49.51 against a $44-46 guide, and a next-quarter guide of $12.2-12.6 billion at about 85%. The numbers matter less than the trade they reveal.

Right on Direction, Too Small on Size, Wrong on One Thing

In August, at about $1,350, I argued Sandisk was “trading upside for duration” and that I would own it. It is $1,330 today, after touching $1,887 in September. First, a scorecard.

My May tracking items mostly passed. I said Datacenter above 35% of revenue would mark Sandisk as a core AI supplier; it is now 43.5%. Most important, I wrote that bit-led growth was “needed for durability.” This quarter, for the first time, growth came mostly from volume. Contract coverage, which crossed 50% of FY27 bits in August, did not move. And one warning I set in January is flashing: I said inventory above 125 days of supply would signal softening demand. It is about 166 days, down from 178 last quarter but well above my line. The likely reason is stock built for datacenter shipments, but that is a question for management, not an answer.

Two admissions. First, I was wrong about why customers sign. In May I wrote that customers “will not pre-fund the same commitments with Micron or SK Hynix,” because Sandisk’s joint venture with Kioxia makes its promises more credible. They do pre-fund them. Micron has $32 billion of customer commitments, mostly in cash. SK hynix has long-term agreements with 14 customers. Samsung wants 60% to 70% of its capacity on multi-year terms. That argument is withdrawn. Second, in May I wrote that “the multiple is roughly right”; in August I argued for a higher one. I changed my mind, and this quarter is the evidence for the later view.

The bigger pattern is magnitude: every bull case in this series became the next floor. My August base case put FY30 earnings at $151 a share; Sandisk will earn about $105 in this fiscal year’s first half alone. That was a downturn case, corrected below.

Less Price, More Bits, Higher Margin

The most revealing number came from CFO Luis Visoso. Asked by Bernstein’s Mark Newman about price and volume, he began, “We normally don’t break this, Mark, into its pieces.” He already had. Revenue growth was “approximately 60%... attributable to volume growth and 40% to higher pricing.” On a 23% revenue increase, that is roughly 13% more bits at roughly 9% higher prices.

Compare that with the rest of the industry in the same three months:

In August I described Sandisk as “two books, one income statement”: a contracted book at mostly fixed near-term prices, and an uncontracted book that floats with the market. The pattern above is what that structure should produce. If half the bits were priced roughly flat under contract, a 9% blended increase implies the spot half rose about 18%, close to Samsung. Treat that as an illustration, not a finding. The contracts have variable components, Sandisk’s product mix differs from its peers’, and its contracted share of revenue may differ from its contracted share of bits. What the numbers show is that Sandisk realised less price and more volume than its peers. What they suggest is that contracts are part of the reason.

Either way, margin rose. Operating profit grew $1.77 billion on $2.07 billion of extra revenue, an incremental margin of 85.5%, because Datacenter went from a third of revenue to 44%, cost per bit fell, and Consumer was allowed to shrink 29%.

Hard Drives Already Ran This Experiment

My first piece on Sandisk, in January, argued that NAND was following hard drives “twenty years behind,” and that “profitability came from market structure, not innovation.” I also wrote that even the hard drive duopoly “hasn’t generated spectacular returns.” That line has not aged well.

Hard drive makers stopped adding factories and sold to hyperscalers on build-to-order contracts. Seagate’s Dave Mosley, last year: “We believe BTO contracts are also beneficial for our customers by providing predictable assurance of supply.” Seagate’s gross margin went from a record 37.9% in mid-2025 to 57.8% this week. The market pays about 19 times earnings for Seagate and 16 times for Western Digital.

Sandisk runs the same playbook: multi-year commitments, supply growth set by technology rather than new fabs, and a “No” to chasing demand. Goeckeler to Melius’s Ben Reitzes:

In general, we’re hearing from our customers that they have underestimated the amount of NAND they need to run their infrastructure. We have multiple customers coming back for a second round, and then balancing that against bringing new customers into the portfolio.

The honest limit is structure. Hard drives have three makers and no Chinese entrant; NAND has five, plus YMTC, whose revenue share has risen from 8% to 13%. And my January warning that Samsung would respond if Sandisk pushed for share still applies: Sandisk grew bits about 13% this quarter, and Samsung guided fourth-quarter bits up in the mid-teens.

A Signal Is Not a Contract

No new contract closed. Customers (eight), floor revenue ($93.9 billion) and guarantees ($16.5 billion) are unchanged, contract obligations slipped to $88.9 billion from $91.1 billion, and new prepayments and deposits fell from $1.94 billion to $465 million. Asked by Cantor’s C.J. Muse about the pipeline, Goeckeler was candid: “But we’re just not there yet, right? We got to go through negotiations on all of those.”

Management’s word of the night was “signals”: “customer signals now extending beyond fiscal year 2030.” A signal is not a contract. I think Sandisk is waiting on purpose. Prices are rising, half its bits still float, and Visoso said the right share under contract is “likely higher than our current levels.” Smart, but it means duration is still a forecast. Even more coverage would not settle it alone: more bits under contract raise the floor only if the floor prices are good, and Sandisk has never disclosed where they sit.

On funding, the worry behind October’s sell-off, Visoso was specific: 95% of contract partners are investment grade, and demand is “if anything, strengthening.” Receivables, up from about 48 to 54 days of sales, are the number to watch.

The Variant Perception

The market prices Sandisk at about 6 times this year’s earnings, level with Micron, as if it still rides spot prices. But Sandisk realised less of the price spike than its peers, shipped more bits, and still lifted its margin. That looks like a lower-swing earnings stream. The hard drive makers show what investors pay once they believe in supply discipline. The gap closes when Sandisk’s contracts extend past fiscal 2028. If they do not, the market is right.

Three Prices for 2029

A three-year price is what the market will pay in late 2029 on fiscal 2030 earnings: revenue × operating margin × (1 − 14.3% tax) ÷ shares × multiple. Fiscal 2026 revenue was $20.25 billion and fiscal 2027 is about $50 billion. Prices are pre-split; divide by eight after 20 November.

Two Books Hold sits below management’s model on both revenue and margin, and its 78% gross margin is under the “around 80%” Visoso says contracts earn even at floor prices. Getting from 154 million to 120 million shares needs about $61 billion of buybacks at $1,800; fiscal 2027 free cash flow alone is about $29 billion, at an adjusted 52% margin. One caution: before stripping out contract deposits, free cash flow covered only 83% of earnings this quarter as receivables grew, and capex rises through the year. The Hard Drive Path caps margins at management’s own targets and needs a reason to grow 18% a year. That reason is NAND moving into more inference tiers: cache for long contexts, QLC data lakes, and possibly SLC in place of some DRAM, where Goeckeler cited a “20x-40x” cost gap. Paper Floor is wrong if Sandisk discloses fiscal 2029 coverage of at least 40% of bits by mid-2027, if gross margin stays at or above 80% through fiscal 2028, or if Samsung and SK hynix keep 2027 NAND bit growth at or below 20%.

The probability-weighted value is about $2,840, up 114%. In August my scenarios were $250, $1,815 and $3,270, weighted to about $1,790. The $250 case assumed the contracts failed outright; I now treat that as a small risk inside Paper Floor’s weight, not a scenario of its own. August’s middle case was a downturn; now the downturn is Paper Floor. The variable that matters most is operating margin: every five points moves Two Books Hold by about $195.

What Would Change My Mind

One threshold, no excuses: by the fiscal second-quarter report in late January, Sandisk either signs a new contract or extension, or discloses coverage beyond fiscal 2028. If it does neither, the existing contracts still stand, but the duration premium in my numbers is unearned, and I will move my weights toward the market’s view.

In June 2023 this same business ran a gross margin of negative 11.9%. The old Sandisk said yes to every upcycle and paid for it in every downturn. The new one is saying no, and charging customers for certainty instead. This quarter proved it can do that and still earn 85%. January will show whether customers keep paying.

Subscribe now

• The operating opportunity expanded materially: AWS grew 39.3%, margins reached 38.7%, backlog rose to $578 billion, and management now believes AWS can grow above 40% in both 2027 and 2028. That implies a much larger future earnings base than pre-results forecasts assumed. Pasted markdown

• The strategic advantage is broader than scarce GPU capacity: Amazon’s case rests on customers running applications, data, inference and tools in the same environment. If that platform relationship persists after capacity shortages ease, AWS can keep earning from the operating system around AI rather than merely renting accelerators. Pasted markdown

• But the financing burden is now impossible to treat as secondary: free cash flow is deeply negative, long-term debt has risen sharply, and infrastructure plus strategic investments are consuming enormous capital. The thesis works only if that spending ultimately produces a larger and sufficiently profitable earnings base.

“With our continued pace of innovation, new core and AI functionality, enterprises wanting to run their core and AI workloads in AWS, and the amount of capacity that we’re adding each year, we believe that AWS can grow north of 40% in each of 2027 and in 2028 and be a trillion-dollar revenue business in the next five to 10 years. We’re in the very early days of this inflection. Our strategy is working, and I expect it will for the foreseeable future.”

— Andy Jassy, October 29 earnings call.

My first reaction was that Amazon had materially increased the earnings opportunity investors should be underwriting. AWS is already a $184 billion annualized business. Growing that business more than 40% for two further years is not a routine extension of the current quarter.

Yet Amazon shares fell approximately 7% in extended trading. Revenue of $201.9 billion narrowly missed expectations; advertising disappointed; and the $31 billion midpoint of Q4 operating-income guidance fell below the $33.1 billion in our October 26 consensus snapshot. Those concerns are legitimate. But Q3 operating income of $28.1 billion exceeded that snapshot by 7%, and AWS grew 39.3% while earning a 38.7% operating margin.

In July, we concluded that the Foundry worked but might never finish being built. Q3 confirms that argument while enlarging the opportunity. I think the lower share price offers a better entry into a stronger franchise not because financing no longer matters, but because the potential earnings increase deserves more weight.

A Much Larger Earnings Base

The quarter largely delivered what our last published article anticipated. We expected AWS to add $2.5–4 billion sequentially; it added $3.75 billion. We wanted margins above 35–36%; they reached 38.7%. The new information is management’s expectation that this much larger business can accelerate again in 2027 and sustain that growth in 2028.

The distinction becomes clear against the forecasts retained in our pre-earnings preparation. Using approximately $175 billion of estimated 2026 AWS revenue, exactly 40% growth produces the following:

At a 35% operating margin, $343 billion of revenue would produce approximately $120 billion of AWS operating income. That exceeds Amazon’s entire current trailing operating income of $104 billion. Even allowing margin compression, the potential change in earnings scale is substantial.

The revenue difference alone would contribute roughly $13 billion of additional operating income at the same assumed 35% margin. That is a sensitivity, not a claimed beat against segment-profit consensus: it assumes the additional revenue earns that margin and says nothing by itself about the capital needed.

Jassy’s confidence rests partly on customers reserving hundreds of megawatts, sometimes a gigawatt, years ahead. Backlog reached $578 billion, up $82 billion sequentially. Those reservations improve visibility, but backlog is not revenue already earned, cash already collected, or proof of attractive pricing. Customer concentration and financing remain material unknowns.

The quarterly pattern also deserves attention. AWS’s sequential addition slowed from $4.65 billion in Q2 to $3.75 billion in Q3. Management attributed the variation to uneven capacity additions. That is plausible, but it cannot become a permanent explanation for missing the new growth trajectory.

My conclusion is nevertheless constructive: Amazon has supplied a credible reason to raise our expectations for future operating earnings. Whether those earnings deserve a premium valuation depends on why customers keep choosing AWS and what Amazon must spend to serve them.

The Advantage After Scarcity

The strongest strategic claim was not the trillion-dollar ambition. It was Jassy’s explanation of why AWS should remain advantaged when computing capacity becomes less scarce:

“In AWS, having the leading functionality, security, and operational performance can’t help but drive customers to run their apps and store their data in AWS, which can’t help but drive AWS AI demand because customers prefer their production inference to be near the rest of their data and apps, which can’t help but drive core AWS demand because AI uses core for work like reinforcement learning, post-training, and tool usage, which can’t help but drive reinvestment back into our broader AWS functionality. This is a different model than primarily hosting GPUs or renting short-term data center capacity. This is a long-term platform strategy that broadly helps customers invent and evolve their customer experiences, operate stably, and build a reliable partnership they can grow for many years. It also yields very attractive long-term economics for us, regardless of supply scarcity.”

This is the mechanism our July article identified, now articulated more explicitly by management. An enterprise agent does not merely generate text. It retrieves information, calls applications, checks permissions and records actions. The model supplies intelligence; the surrounding infrastructure makes that intelligence useful. Amazon can earn revenue from both. Past Articles on Amazon

The important customer relationship is therefore not simply access to an accelerator. It is the operating environment containing the customer’s applications and data. Moving inference between providers may be relatively easy; recreating every integration, permission and operational dependency is harder. That is the potential source of durability after capacity shortages disappear.

Management’s evidence is encouraging. Core compute and storage grew at their fastest rates in several years. Inference represents well over half of AWS AI revenue, and most Bedrock workloads run on Trainium. Fortune 500 spending on Bedrock increased nearly sevenfold. These disclosures support the platform argument, although none isolates its incremental return on capital.

The silicon matters because Amazon can lower the cost beneath a service customers already use. Customers choosing Bedrock need not choose their accelerator separately. If several models become sufficiently capable for most workloads, as Jassy expects, model choice can become a feature of AWS rather than a reason to leave it.

That does not guarantee Amazon retains the savings. Competition can pass them to customers. Olsavsky also acknowledged that AI margins remain below core AWS margins, creating dilution as AI grows. Our valuation therefore should not assume today’s margin simply survives unchanged; the base case below uses 36%, not 38.7%.

External Trainium sales, planned for late 2027, add another way to monetize the capability. Selling equipment recognizes revenue sooner and transfers much of the deployment financing to the customer; operating it inside AWS can yield years of service revenue but requires more capital. Jassy explained the trade-off without quantifying comparable returns. I treat external sales as an option, not a separate profit pool to add on top of cloud earnings from the same chips.

The advantage I am underwriting is consequently broader than cheaper inference. It is Amazon’s ability to combine cost-efficient computing with services customers need to operate enterprise applications. Q3 strengthens that case without proving every part of it.

The Financing Error We Must Own

The strongest objection is that attractive operating margins can coexist with disappointing returns on a rapidly expanding capital base. Our July article anticipated the mechanism:

“Every capacity cohort may earn a good return, reveal still more demand, and justify a larger cohort before the cash from the previous one appears in consolidated results. Operating cash flow rose 33% to $161.4 billion. Net property spending rose 64% to $169.0 billion. Free cash flow turned negative. Long-term debt nearly doubled in six months.” Past Articles on Amazon

Q3 made that imbalance larger. Trailing operating cash flow reached $168.3 billion, but net property spending reached $210.7 billion, leaving negative free cash flow of $42.4 billion. Long-term debt reached $173.1 billion, exceeding the $150 billion warning threshold we retained in July.

We should acknowledge that miss without rewriting the original thesis. Debt financing was already part of our April analysis; Q3 changes its magnitude, not its existence. Nor can consolidated operating cash flow be attributed entirely to retail and advertising. AWS itself contributes to the company’s cash generation. Past Articles on Amazon

There is a second financing demand outside the headline capex figure. Amazon spent $63.8 billion during the first nine months in the category covering acquisitions, nonmarketable investments and other items. That is excluded from its reported free-cash-flow measure. Not all borrowing finances data centers, and strategic investments in AI customers must be evaluated separately from infrastructure spending.

The operating evidence still argues against treating this as indiscriminate expansion. AWS added approximately $6.3 billion of year-over-year operating income on $13 billion of revenue. But that roughly 49% incremental margin is not the return on this quarter’s capex: current earnings reflect previously deployed assets, mature services and accounting depreciation, while much current spending supports future revenue.

Retail remains relevant to the funding capacity. Units grew 12% against 11% shipping-cost growth, restoring the favorable spread we flagged in July. Advertising is less reassuring: constant-currency growth slowed to 12%, below our threshold, despite credible explanations involving Prime Day and other comparisons. Management’s promised high-teens Q4 recovery now needs to happen.

The evidence favors productive reinvestment over wasteful expansion, but fully burdened returns remain unproven. My response is to lower assumptions about cash conversion and allow for financing not to ignore the debt, and not to assume negative free cash flow disproves the business.

What the Stronger Business Is Worth

Our variant perception is not that AWS is growing quickly. It is that Amazon can sustain a larger earnings base than the pre-earnings forecasts implied, while preserving enough profitability to compensate shareholders for the capital required.

That is more demanding than accepting Jassy’s revenue outlook. It requires lower-cost silicon, utilization and operating efficiency to offset some AI margin dilution. It also requires the investment programme eventually to generate cash beyond its continuing expansion. The share-price decline offers a better starting valuation; it does not prove that the market is wrong.

I value 2029 operating income after depreciation and stock compensation, rather than capitalize investment gains or add back the cost of running the infrastructure. The starting assumptions are approximately $829 billion of 2026 group revenue, comprising $175 billion of AWS and $654 billion of everything else.

The calculation is terminal operating income multiplied by the enterprise-value multiple, plus the net financial-asset bridge, divided by diluted shares. The bridge is an assumption, not a forecast balance sheet; an additional $100 billion of net liabilities would reduce each value by approximately $9.

Cash flow is a separate test, not something mechanically inferred from operating profit. The scenarios assume 2029 operating cash flow/cash capex of $285/$320 billion, $360/$285 billion and $440/$305 billion, respectively. These assumptions are demanding, especially the recovery in the base case, and are not independently verified forecasts.

The base case lets AWS’s margin decline and its growth moderate in 2029. Its $367 valuation sits within July’s published $345–425 range, although this model uses operating income rather than the earlier table’s P/E approach. The bull value remains within the earlier $540–690 range. The bear value falls below the previous $200–240 range because financing and cash-conversion risks have increased. Past Articles on Amazon

Assigning subjective probabilities of 25%, 50% and 25% produces a weighted value around $377. I would not confuse that arithmetic with precision: terminal multiples, asset returns and financing outcomes drive the distribution. Nevertheless, approximately 17% annualized upside in the base case makes selective accumulation around $220–230 attractive to me. The bear value is not a guaranteed floor.

The thesis now has concrete tests. AWS should deliver approximately 40%-plus growth in 2027 while maintaining underlying margins around 36–39%; growth repeatedly below 30% despite available capacity, or margins below 35% for several quarters, would challenge the operating case. Advertising should recover to high-teens growth in Q4, while shipping costs should continue growing more slowly than units.

The financial test must not keep moving. July called for operating-cash-flow growth to begin overtaking capital-spending growth during 2027. I retain that as an important signpost, alongside broader paid Trainium adoption beyond strategically financed customers. More backlog without better conversion, or continued rapid debt accumulation without improving capital returns, would push the valuation toward the downside case. Past Articles on Amazon

The next major checkpoint is the Q4 report and a quantified 2027 spending plan. My judgment today is that Amazon’s operating opportunity has improved enough to outweigh the greater financing burden at the reduced price. We underestimated the funding required, but we also underestimated how large AWS could become. The right response is to incorporate both not let the cash-flow disappointment erase the stronger earnings evidence.

Subscribe now

• The near-term operating story has improved materially: revenue rose 29%, free-cash-flow conversion recovered to 83%, and December guidance points to another sharp step-up as optical supply constraints ease. What looked like sequencing was partly a capacity-planning problem.

• But the structural thesis is more demanding than a bigger WFE market: KLA’s 2030 framework assumes its share of wafer-fab equipment spending rises from roughly 7.5% toward 9%. The industry has advanced faster than expected; KLA now has to prove that inspection content per fab rises too.

• That matters because the stock is not priced for obvious failure: at current levels, consensus earnings are already broadly embedded. The variant is not simply “WFE is strong,” but that process-control content, margins and cash conversion outperform already elevated expectations.

In July’s The Market Arrived First, I wrote that KLA’s structural thesis was advancing while the stock was running ahead of the financial evidence. The most uncomfortable number was free-cash-flow conversion: just 59% in June. September brought that back to 83%, revenue rose 29% to $4.14 billion, and management guided December revenue to $4.70 billion, nearly $200 million above the pre-earnings consensus.

It also demands harder questions.

KLA’s 2030 model assumes that its share of overall wafer-fab equipment spending rises from about 7.5% in 2025 toward 9%. Meanwhile, the wafer-equipment market itself is growing much faster than management expected. The industry has advanced the model; KLA has yet to demonstrate the entire share gain. That is the investment question now: can KLA turn a bigger market into a bigger claim on each equipment dollar, or is the complexity tax still mostly a metaphor for a good cycle?

We mistook the constraint

Our May article offered a plausible explanation for why Lam Research and Applied Materials were growing faster than KLA: the builder arrives before the inspector. First fabs need capacity; later they need to make that capacity yield. In July, I argued that the market had capitalized the eventual handoff before KLA’s cash flows arrived.

The October call complicates that story. CFO Bren Higgins said the weak first half of 2026 was not principally about business mix, but KLA’s ability to supply its customers. CEO Rick Wallace said KLA had planned for a relatively flat 2026, only to discover that demand was arriving sooner than expected. What I called sequencing was partly a capacity-planning error.

The bottleneck is specialized optics. Tool lead times average roughly twelve months; certain advanced inspection platforms take longer. KLA has spent a year expanding critical supplier output, and that output is now increasing month by month. Management still says demand exceeds supply. It has not become unconstrained; it is better positioned to deliver against commitments it could not previously fulfill.

The financial evidence fits. Wafer-inspection systems revenue rose 37% year over year to $2.11 billion in September, versus approximately 1% growth in June. Wafer inspection accounted for roughly $330 million of the quarter’s $478 million sequential increase in consolidated revenue. KLA’s biggest product category, not just its smaller patterning business is accelerating.

Lam’s September revenue still grew approximately 54% year over year, against KLA’s 29%. The builder has hardly surrendered the lead. But Lam’s December guidance implies about 5% sequential growth, while KLA’s implies nearly 14%. This is an encouraging forward convergence, not proof of a permanent handoff. Equipment mix and shipment timing complicate comparisons.

Customers do not need to discover inspection; some demand already exists. What remains to be established is how much of the growing backlog reflects a lasting increase in inspection content rather than shipments that were simply delayed.

The market is bigger. Is KLA’s share?

KLA’s March Investor Day envisioned approximately $26 billion of revenue and $8.40 of split-adjusted earnings per share by 2030, supported by wafer-equipment spending around $220 billion, plus or minus $20 billion. KLA now expects industry spending to exceed $200 billion in 2027. During the call, Wallace said the assumptions behind the 2030 model would be exceeded before the target year.

Bernstein’s Stacy Rasgon asked the obvious question: why should the 2030 revenue target not arrive in 2027? Higgins identified what the industry forecast leaves out. KLA has not yet reached its targeted overall equipment-market share; its services business needs years of installed-base growth, and businesses outside core process control have their own trajectory. Industry spending can arrive early without every part of KLA’s financial model arriving alongside it.

UBS’s Tim Arcuri framed the tension more pointedly, describing KLA as having lost roughly 20 basis points of WFE share in 2026. That was an analyst characterization, not a published company measurement, and Higgins did not independently quantify the figure. Still, he directly committed to the direction: KLA expects to outgrow WFE in 2027.

That is the distinction worth underwriting. At $200 billion of equipment spending, a one-percentage-point change in market participation represents approximately $2 billion of industry expenditure before differences in definitions, timing and revenue recognition. The scale explains why share gains matter more than another industry forecast increase.

The technical evidence is promising. Management says some advanced DRAM applications now require inspection intensity comparable to leading-edge logic. HBM’s stringent performance requirements, base-die logic, reduced redundancy and expensive yields make earlier defect detection more valuable. New DRAM fabs add equipment volume; more demanding processes could add KLA tools per fab.

But this cannot be inferred from quarterly memory mix alone. Memory represented 38% of semiconductor process-control systems sales to semiconductor customers in March, 21% in June, 28% in September and is guided at 40% in December. Conventional DRAM contract prices were forecast by TrendForce to rise 58–63% in the June quarter, when KLA’s memory mix was lowest; they are forecast to rise 10–15% in December, when its mix is highest. These are unlike quantities with different lead times, so the absence of same-quarter correlation disproves neither the scarcity premium nor rising technical intensity. It does show why a strong memory quarter cannot be equated with a permanent content gain.

Advanced packaging supplies the more distinctive possibility. Hybrid bonding is pushing certain inspection requirements toward 100-nanometer sensitivity. KLA is selling more sophisticated laser-scanning tools and even some Gen4 inspection into packaging, reusing technologies developed for advanced front-end manufacturing. The customer is moving into a problem KLA already knows how to solve. That expands the economic return on accumulated R&D and qualification expertise. Its approximately $1.1 billion calendar-2026 packaging outlook, however, was reiterated rather than increased; higher technical ambition is not yet evidence of a new revenue surprise.

The moat is difficult inspection physics, production learning, qualified tools, engineering integration and services, not a consumer-style network effect. Those advantages become more useful as device complexity rises, but KLA must demonstrate that they translate into a larger fraction of industry spending.

The reported quarter conceals two different cycles

September’s numbers are impressive, but their composition matters.

The first complication is geography. China represented 34% of September revenue, approximately $1.41 billion, versus about $967 million in June. On rounded disclosures, that increase accounts for roughly 90% of KLA’s entire sequential revenue gain. This does not mean the wafer-inspection acceleration was driven by China: KLA does not provide the product-by-country cross-tab needed to establish that. But it does mean we should not describe September as though all incremental revenue came from leading-edge foundry deployments outside China.

Management expects China’s percentage to decline toward the low 20s in 2027, even as absolute China sales may grow modestly. If that happens, next year’s outperformance must come from the promised leading-edge logic, new DRAM facilities and packaging demand. This makes the test more revealing, not easier.

The second complication is profitability. The same DRAM scarcity that raises the value of yield for memory makers raises the cost of DRAM inside KLA’s image-processing systems. KLA cannot readily reprice fixed-price backlog. It expects new product agreements and engineering changes that use less memory to improve gross margin later in 2027 and into 2028. Fast-growing packaging and adjacent businesses may also dilute the consolidated gross-margin percentage.

There is a potential partial offset in a future memory downturn: falling DRAM prices could reduce KLA’s system costs just as memory makers cut equipment spending. But the sizes and timing are not disclosed; calling this quantified earnings hedge would be false precision.

For now, volume is doing the work. September operating margin reached a record 46.1%, and December guidance implies approximately 47.4%, even with gross margin broadly flat. That is why the timing matters: supply-led revenue growth is arriving first; better product pricing and memory efficiency may improve profit conversion later. Services, $851 million in September, up 14%, should benefit as more systems enter the installed base, although hardware is presently growing faster.

Cash offers a more tangible improvement. KLA produced $1.34 billion of free cash flow while inventory consumed $434 million of operating cash and capital expenditures reached $195 million. The recovery from June’s 59% cash conversion therefore did not depend on starving the business of investment. It answers our July concern favorably, though one quarter cannot establish the working-capital cost of the entire expansion.

The stock is not priced for failure

One tempting valuation argument is that at $158.77 before earnings and a 25-times terminal P/E, KLA’s price implied just $6.35 of FY29 EPS, about a quarter below consensus. That arithmetic divides today’s price by a future multiple but omits the return investors require while waiting for FY29.

At a 10% annual required return over three years, $158.77 today at 25-times future earnings implies about $8.45 of FY29 EPS. Updated consensus as of October 29 is approximately $8.47, alongside FY29 revenue of $26.05 billion. On those assumptions, the price was broadly consistent with consensus being achieved, not with a 25% earnings collapse. Around $165 after hours, the same hurdle requires roughly $8.78 of FY29 EPS. Dividends and different multiples change the result, but the point holds: the stock is not demonstrably priced for failure.

Our variant must therefore be specific. Better shipment conversion and rising process-control content may push FY29 earnings above an already ambitious consensus. Greater confidence in those earnings may also sustain KLA’s premium multiple. Neither proposition is free. Our central $9.05 EPS estimate is only about 7% above the newly revised FY29 consensus.

The next tests are straightforward: December shipments must support the $4.70 billion guide; the subsequent quarter should preserve sequential growth; KLA’s comparable process-control systems growth should exceed WFE during 2027 and remain strong after the supply catch-up; gross margin should move toward 63–64% as pricing and product redesign take effect; and trailing cash conversion should remain close to 80%. China and fab-readiness timing need to be reconciled rather than used as permanent excuses.

In April, we identified the complexity tax. In May, we tried to explain when it would be collected. In July, we learned that the stock market had anticipated the collection. October shows KLA finally shipping more equipment and generating more cash, but a larger tax base is not the same as a higher tax rate.

2027 is when KLA must demonstrate the difference.

$KLA

Subscribe now

• The control-layer thesis is getting more concrete: enterprises may need governance before they deploy agents at scale, giving ServiceNow a chance to sell visibility, permissions and safety first and then participate in the workflows and consumption that follow.

• ServiceNow does not need to own the AI interface: ChatGPT or another assistant can initiate the request while ServiceNow executes the governed workflow underneath and still monetizes the activity. That makes the company potentially more valuable as infrastructure beneath multiple front doors, rather than as the exclusive front door itself.

• But the economic proof is still incomplete: cRPO guidance weakened, gross margins remain under pressure, deployment support is absorbing resources, and management has not yet shown how much incremental customer spending AI creates. The strategy is clearer than the financial acceleration.

ServiceNow reported $3.988 billion in subscription revenue, growing 20.5% in constant currency, and raised its year-end AI annual contract value target to $1.7 billion. But fourth quarter contracted-revenue guidance disappointed, and shares fell approximately 18% in after-hours trading. Investors were being asked to believe in accelerating AI adoption while the principal forward indicator suggested slower growth.

I think the quarter deserves a better explanation than either “the market missed the strategy” or “AI has failed to accelerate the numbers.” Our July article had already distinguished the two: ServiceNow’s operating context was becoming more valuable to agents, but strategic relevance had not established better economics. Q3 explains how management intends to connect them. Past articles on ServiceNow

The most revealing exchange was not about the revenue beat. An analyst asked whether enterprises were purchasing Control Tower before deploying agents at scale. Amit Zavery answered that governance was coming first. That sequence matters: ServiceNow is not simply waiting for AI adoption to generate demand for its platform. It is trying to sell the capabilities that make adoption possible, then participate in the activity that follows.

Control Before Consumption

Enterprise AI creates a distinction between being capable of an action and being authorized to perform it. An agent may know how to change a firewall configuration without knowing whether the change is permitted, which applications depend on it, or who must approve it. Greater intelligence does not eliminate those requirements; more autonomous activity can make them more consequential.

This was the foundation of our control-layer thesis. What Q3 added was management’s description of how that need becomes a customer purchase. Zavery told Evercore:

It’s coming in first before AI goes in production and the volume of usage just starts going up because as you start building out more AI, as you get confidence, we are really helping them manage it. It’s not about orchestration really. It’s really the visibility, the control, and the safety. Then we can orchestrate many things underneath the covers depending on what systems customers might be using.

The implication is more interesting than another governance product. A customer can purchase control capabilities before its agents perform much work. Successful deployment can then generate demand for execution, additional workflows and more control. Governance is potentially both a revenue opportunity and an enabler of subsequent consumption.

Security belongs inside this mechanism. Armis contributes asset visibility; Veza contributes identity and permission context; ServiceNow’s existing workflows provide a route from discovering a problem to resolving it. Management’s ambition is to connect discovery, authorization, execution and remediation rather than sell disconnected tools. Security and Risk appeared in sixteen of the twenty largest Q3 deals, although acquisitions prevent that figure from establishing organic growth on its own.

This also explains why treating Armis only as an adjustment to the growth calculation misses half the investment question. The acquired revenue must be separated when evaluating demand, but the capabilities must be assessed when evaluating the platform. Strategically coherent acquisitions can still earn poor returns; neither their coherence nor their cost settles the matter.

The resulting advantage would be accumulated operating context within each customer: more workflows, more dependencies, more permissions and greater difficulty replacing the system. That is not proof of a universal network effect or an exclusive control point. It is a plausible mechanism through which serving more enterprise work makes ServiceNow harder to displace.

The Interface Can Belong to Someone Else

Our earlier writing sometimes put too much weight on ServiceNow owning the AI front door. The stronger thesis is that it can remain valuable behind several front doors. Q3’s OpenAI and Google integrations make that distinction concrete: an external assistant initiates the request, while ServiceNow executes the relevant governed workflow. Past articles on ServiceNow NOW 3Q26 PR

Zavery explained the commercial arrangement in the final question:

What we’re doing with OpenAI is allowing ChatGPT as a way to interface into ServiceNow. But all of the work happens inside ServiceNow. It’s just a call coming from ChatGPT. Then the work which we were doing, typically through our user experience, now happens through this interface, and we get full value for it from the monetization perspective because you are burning down assist for every work you’d ask ServiceNow to do.

This is not an abandonment of its own interface. EmployeeWorks reached $100 million in ACV in fewer than three quarters; management argues that its planning, reasoning and specialized agents offer capabilities beyond simply calling a workflow. ServiceNow wants both opportunities: a differentiated employee experience and distribution through other assistants.

The important upside is not merely replacing a lost seat with an equivalent consumption charge. McDermott described a chance to increase revenue per user fivefold once customers become productive consumers. That is an ambition, not an achieved result, but the underlying proposition is intelligible: a person supported by agents can initiate substantially more work than that person previously completed manually.

More activity does not automatically create another invoice, because assists can be consumed from existing entitlements. The economic expansion arrives when customers purchase additional capacity or renew at higher spending levels. Here the evidence is incomplete but not absent: Zavery described customers expanding from initial use cases into broader deployments and renewing with higher token allowances.

AI net-new ACV growth of 65%, Control Tower net-new ACV growth above sixfold and thirteenfold growth in production agentic consumption belong at different points in that progression. Together, they show adoption and commercial demand. They do not establish consolidated acceleration, but dismissing them because the final profit bridge is missing would be as mistaken as treating them as proof of it.

Nor must ServiceNow defeat every competing control layer. Microsoft, cloud providers and application vendors can retain important roles while ServiceNow earns attractive returns from the processes it handles. The relevant competitive question is whether its context and execution remain sufficiently valuable to support pricing not whether every enterprise action becomes its exclusive property. Past articles on ServiceNow

Paying to Put the Agents to Work

The missing link is deployment. Under a contracted subscription model, an application can generate revenue while the customer implements it slowly. If future expansion depends on consumption, getting the customer into productive use becomes more directly connected to growth. Implementation support starts to resemble an investment in future spending rather than merely a service obligation.

ServiceNow is therefore including deployment expertise in selected large AI contracts. Management says supported customers advanced twice as quickly during their first five weeks and consumed more assists. Gina Mastantuono described the intended economics:

We do not expect to include this delivery expertise over the long term. As with other new technology adoption cycles, it is about optimizing the first wave of adopters and leveraging the know-how in a repeatable and scalable way. As the AI gets better, the ability to autonomously deploy gets easier.

Our earliest articles anticipated a renewed role for services in enterprise AI. Q3 makes that role financially tangible. Some contract value is allocated to professional services rather than subscriptions, reducing subscription growth by approximately 25 basis points in Q3 and an expected 50 in Q4. A later return of that allocation to subscription revenue would improve the reported category; it would not, by itself, prove higher total customer spending. Past articles on ServiceNow NOW 3Q26 Slides

There is a revealing tension here. The complexity that makes ServiceNow difficult to replace also makes new deployments difficult to activate. Management must make implementation repeatable without commoditizing the operating knowledge customers value. Its self-service Flow product attacks the adoption problem from the opposite direction: a lightweight service desk for customers not ready for a full enterprise implementation.

July’s article established the economic test before these deployment disclosures. It remains the right one:

Trusted execution is scarce only if ServiceNow can capture more value from each governed action than it spends delivering it. Inference, hosting, third-party models, acquired technology, and cloud infrastructure all sit above the old cost structure. Past articles on ServiceNow

Our expectation that gross margins would stabilize around 82–83% was wrong on timing. But recovering the old percentage is not the only route to better economics. Renting more infrastructure can reduce capital requirements, while internal automation can reduce operating expenses. Management says quota-bearing sales headcount increased by double digits even as support functions contracted; total headcount fell from 30,457 to 28,855 during Q3.

ServiceNow need not finance frontier-model development to benefit from it. Cheaper intelligence could lower its input costs while increasing the work customers automate—the strongest reading of McDermott’s capital-light claim, rather than an assertion that compute is free.

These are plausible offsets, not permission to ignore cash quality. Nine-month operating cash flow declined 3%, while lower capex helped conventional free cash flow grow about 7%. Company-adjusted FCF grew 20%, benefiting additionally from acquisition-related cash-cost add-backs rising from $28 million to $371 million. The higher 36% full-year adjusted FCF-margin target therefore needs to be read alongside how the cash is generated.

The economic test is broader than gross margin: can additional consumption, less capital spending and internal productivity produce more cash per share after deployment costs and dilution? Q3 supports parts of that operating model, but not the complete return.

The Contracts Still Have a Vote

None of this makes the weak forward indicator disappear. Current remaining performance obligations measure committed revenue expected over the next twelve months; they are not simply a measure of the old seat business. The Q4 growth guide of 18.5% includes approximately 175 basis points from Armis, leaving an arithmetic 16.75% before other acquisition, timing and classification adjustments.

Management identified early renewals pulled from Q4, more on-premise business, service allocations and cancellable contracts excluded from cRPO. These effects are legitimate, but their different denominators prevent a clean synthetic “underlying growth” calculation. Moreover, $12 million of the $15 million full-year subscription-guidance increase came from currency.

The useful conclusion is not that every accounting explanation is an excuse. It is that the evidence still does not support raising our organic growth assumption. The omission of Q2’s headline count of million-dollar transactions and the earlier quantified non-seat mix also leaves less continuity in disclosures precisely where the transition needs measuring.

William Blair asked the right question: how had total spending changed among the most advanced AI customers? Management offered usage expansion and higher token renewals, but no comparable spending figure. That is encouraging evidence short of the information needed to establish how much incremental business the model creates.

What the Price Requires

My view has changed, but not because the company has discovered another name for its platform. July showed that agents were creating demand for ServiceNow’s context and controls. Q3 explains how governance can enable deployment, external assistants can distribute its workflows, and deployment investment can feed consumption. The strategy is more complete; the path to consolidated acceleration remains slower than we anticipated.

The potential variant is consequently not a governance monopoly. It is that ServiceNow can become a larger business by monetizing more work across an entrenched customer base, without owning every assistant or restoring every point of historical gross margin. Its 98% renewal rate and 680 customers above $5 million in ACV provide a substantial foundation, although neither substitutes for incremental growth. These are the three-year outcomes I would underwrite.

The central case requires neither fivefold spending per user nor universal control. It assumes high-teens growth, manageable delivery costs and moderate cash-margin expansion, valued at 20 times FCF. The upside requires consumption to create a meaningfully larger spending pool; the adverse case assumes a useful platform whose new activity earns disappointing returns. The adverse price is not a guaranteed floor.

The next evidence should connect adoption to money: higher customer renewal spending net of seat reductions, organic contracted growth holding in the high teens and eventually exceeding 20%, and conventional cash generation catching up with adjusted results. A persistent organic trajectory below 17%, or subscription margins below 79% without compensating cash returns, would undermine that judgment.

ServiceNow is trying to sell enterprises permission to deploy AI and the execution that follows. Q3 made that commercial sequence more credible while exposing the investment required to activate it. The opportunity is not that the company has already become the control tower for everything; it is that making more enterprise work safe and executable may be enough to build a materially more valuable business.

Subscribe now

• Cloud has arrived as a genuine second engine: revenue grew 103% to $30.7 billion, operating profit reached $11.1 billion, margins hit 36.2%, and backlog rose to $561 billion. Google is also extending its TPU architecture beyond its own data centers, potentially allowing customers to finance part of the infrastructure required to use Google technology.

• But Cloud’s success does not automatically improve Alphabet’s consolidated economics: shared AI research costs are rising sharply, Search growth slowed to 12%, and Alphabet-level activities absorbed enough expense that Cloud contributed more than 100% of the group’s incremental normalized profit. The laboratory is increasingly eating the gains before shareholders see them.

• The most interesting new thesis is capital-light expansion of the TPU platform: selling systems lets Google earn from its architecture even when workloads run in someone else’s facility and potentially when a competitor’s model wins. That could broaden the return on Google’s research while shifting some power, datacenter and financing requirements to customers.

In July’s The System on Trial, I argued that Alphabet’s advantage was the combination of its models, silicon, infrastructure and distribution, rather than any individual product. That system was beginning to work, but I ended with a qualification:

Search no longer carries the whole strategic argument, but it still carries most of the bill. Alphabet can tolerate Search settling into the low teens if Cloud becomes a second profit pool, which it visibly is becoming. It cannot tolerate Search weakening before the rest of the system is self-supporting.

Three months later, Cloud has become that second profit pool more quickly than I expected. Revenue grew 103% to $30.7 billion, operating profit reached $11.1 billion, and the margin increased to 36.2%. Backlog reached $561 billion, while Gemini 4 Argon and Carbon strengthened Google’s position at the technological frontier. These were better outcomes than the tests I set in July, yet the shares fell roughly 10% after hours.

The qualification, it turns out, mattered more than the milestones. Search growth slowed to 12%, shared AI research expenses rose sharply, and free cash flow was negative $9.1 billion. I had expected a Cloud business exceeding $100 billion in annualized revenue to justify a re-rating; at $123 billion, it instead revealed the weakness in that argument. Cloud’s success does not automatically establish the economics of the system supporting it.

Why the Lab Comes First

The most revealing exchange concerned how Google allocates computing capacity. Asked by Goldman Sachs about competing demands across the company, Sundar Pichai explained:

Each of our business areas submit long-range plans with a very disciplined ROIC framework. I would broadly think of all of this as a long-range planning process, not a short-term allocation exercise. Obviously, our AI models underlie everything we do, from Search and YouTube to Cloud to Waymo and so on. It is the foundation, and we view it as a highly leveraged approach. The full stack approach drives a lot of efficiency as well as gives us a lot of optionality across a wide variety of opportunities.

This is the rationale for putting the laboratory first: its output improves multiple businesses, whereas allocating another unit of capacity to Cloud produces a more immediate but narrower return. The opportunity cost is real. Thomas Kurian acknowledged that Cloud could serve substantially more demand with additional capacity, while Google’s agreement to access approximately 110,000 NVIDIA GPUs through SpaceX carries a contractual cost of $920 million a month, subject to termination provisions. Capacity is valuable enough to rent at an annualized rate exceeding $11 billion even while Alphabet builds its own infrastructure.

Pichai’s argument becomes more consequential when combined with his comments about recursive self-improvement. AI is now helping Google increase its training and launch velocity, and he expects further acceleration over the next twelve months. The proposed loop is that better models improve research productivity, which produces better models faster; those models then improve Search, enterprise software and the tools used to build the next generation.

There is evidence of progress, although not yet proof of that entire loop. Google reports that teams using Argon shipped 35% more code with 30% fewer errors, while developer usage ramped three-and-a-half times faster than Gemini 3. Both Argon and Carbon arrived after quarter-end, however, so they cannot explain Q3’s financial results. What they explain is management’s willingness to spend ahead of those results: the return might extend across Google, and falling behind could be more expensive than sacrificing near-term Cloud revenue.

A Platform Beyond Google

The problem with this strategy is that the laboratory’s ambitions are not bounded by the cash available to finance them. Google’s response is becoming more interesting than simply building additional data centers. When MoffettNathanson’s Michael Nathanson asked why Google would supply companies competing with its own products, Pichai distinguished between products and platforms:

We have long viewed what we build as categories of things which we build as products, and then there are categories of things which we build as platforms. You need to have a clear view on which is which. When you’re building things as platforms, you really work hard to make the platform great and make sure it works for everyone.

Pichai explicitly placed TPU systems in the platform category. Historically, Google’s custom accelerators made its own infrastructure more competitive; the commercial opportunity was largely bounded by workloads Google hosted. Selling complete systems removes that boundary. An AI laboratory can use Google’s architecture without renting all its capacity from Google, while enterprises and other infrastructure providers can deploy it where their own data, power and facilities are available.

This is already broader than a chip-development experiment. Apple and Anthropic use TPUs, Oracle and Crux AI extend distribution, and TorchTPU brings the architecture to developers using PyTorch. More customers increase production volumes and spread development costs; better software compatibility broadens adoption further. Google can potentially benefit from a competitor’s growth even when Gemini does not win that competitor’s customer.

Kurian explained why this matters financially, not just strategically. His answer connected customer commitments, market expansion and infrastructure ownership:

First of all, when Anat mentioned that we procure components to build inventory, we only do that after we have signed contracts. All of the inventory for TPU systems are associated with committed contracts. Additionally, by putting TPU systems in third-party facilities and in customer data centers, it allows us to broaden our addressable markets. For some industries, they have data in their premise that they cannot move to a cloud, and it allows us to bring accelerated compute to where the data sits. Additionally, it also offsets the need for us to have to pre-build facilities to host these TPU systems.

When Google rents compute to a customer, it finances infrastructure and recovers that investment over time. When it sells a system, the customer finances the deployment, while Google finances production and inventory ahead of delivery. Research, software and support still cost money, but Google no longer needs to own every facility in which its technology earns a return. The same strategy expands its market and reduces the long-term infrastructure investment required to serve those additional deployments.

That is the strongest addition to the investment case this quarter. The laboratory can remain expensive while its output becomes available to more buyers, including buyers supplying their own scarce power and capital. The question is no longer simply whether Google can build a better integrated system; it is whether it can sell enough of that system to make the integration more valuable.

What Cloud’s Growth Contains

The complication is that Cloud now combines recurring consumption with system deliveries, and Alphabet does not disclose the revenue or profitability of the latter. Illustratively, external TPU sales of $2 billion, $4 billion or $6 billion would leave the remaining Cloud business growing approximately 89%, 76% or 63%, respectively. Those are assumptions, not disclosures, but they show that substantial hardware revenue need not invalidate the underlying growth story.

There is evidence of demand beyond system deliveries. Just over 55% of the $561 billion backlog is expected to become revenue within two years, equivalent to approximately $309 billion. Kurian also described how Gemini adoption generates spending on compute, storage, databases and security: customers adopting the model have reasons to adopt more of the platform. Hardware and recurring services can reinforce each other without having identical economics.

That distinction changes how I interpret margins. A system sale could earn a lower margin than a recurring Cloud service while requiring less long-term capital, making it an attractive addition despite diluting the reported percentage. Conversely, Cloud’s 36.2% margin excludes shared research costs and barely reflects the third-party capacity that arrived at quarter-end. The business is stronger, but treating every dollar of its revenue or profit as economically interchangeable would overstate what the quarter proved.

Search’s Unfinished Transition

Search presents the reverse problem: usage is accelerating, but revenue is not keeping pace. Query growth was the fastest in five years, yet revenue growth slowed from 19% in Q1 to 17% in Q2 and 12% in Q3. Asked about the disconnect, Philipp Schindler gave an explanation whose order matters:

First, there was FX. It was the single largest driver of Search deceleration relative to Q2. Second, there were vertical-specific dynamics. In retail, some of our APAC advertisers were impacted by the E.U. eliminating its de minimis exemption. In media and entertainment, spend normalized after an unseasonably strong Q2, driven by the FIFA World Cup, which had more matches and therefore more advertising budgets allocated in June than actually in July. Also to the next part of your question, it’s worth noting that the strong growth in AI Mode queries had an impact because of the lower monetization we were seeing in AI Mode.

The final sentence matters, but it does not erase the preceding explanation. Currency was the largest contributor to the slowdown, and AI Mode is not AI Overviews, whose monetization Google previously described as approximately comparable to conventional Search. The evidence establishes a lower-yield product growing within Search; it does not establish that AI has broken the advertising franchise.

The economics depend on whether those conversations replace existing commercial searches or create additional ones. Management says most AI Mode queries are incremental, is expanding advertising internationally, and sees opportunities in agentic commerce. I find that opportunity credible, but query growth alone cannot settle the question. Google is deliberately building the experience before fully monetizing it, just as it invests in research before collecting the resulting revenue. Both choices can be right; making them simultaneously increases the financing burden.

What Reaches Shareholders

The segment reconciliation makes that burden visible. After excluding the previous year’s European Commission fine, Cloud added approximately $7.5 billion of operating profit, while Alphabet added only $5.9 billion. Cloud therefore supplied 128% of the group’s incremental normalized profit: everything outside it, combined, reduced the increase. Alphabet-level activities, primarily shared AI research and development, lost $7.7 billion, up 71% and approximately $1.9 billion worse than consensus.

Those expenses do not belong entirely to Cloud, because the research supports other businesses too. They do belong entirely to Alphabet shareholders, which is why the group’s normalized margin fell from 33.9% to 31.8% while Cloud’s margin expanded. My mistake was assuming that a second profit pool would necessarily improve consolidated economics; the cost of maintaining the technology behind both pools is rising faster than I allowed.

Cash generation tells the same story with an important qualification. Operating cash flow of $42.3 billion fell short of $51.4 billion in capex, but the quarter also included a $5.2 billion fine payment and inventory accumulation for contracted TPU deliveries. Those items make one quarter’s negative free cash flow an imperfect measure of underlying profitability. The structural issue is the $195–205 billion capex programme, a further substantial increase expected in 2027, and a financing mix that now includes common equity, preferred securities and additional debt rather than buybacks.

My valuation therefore remains anchored in consolidated earnings and cash conversion. In the “Platform works” scenario, 2029 revenue reaches $875 billion, comprising $515 billion from Services and $360 billion from Cloud. A 33% operating margin, 19% tax rate and 12.6 billion diluted shares produce approximately $18.56 of normalized earnings per share. Applying 21 times earnings and adding $15 per share of net financial assets, after financial claims and reserves, produces a value near $405. The accompanying cash-flow assumption is $95 billion after $350 billion of capex.

The alternatives remain $220 in “Capital-heavy,” where investment outpaces cash generation, and $629 in “System compounds,” where stronger growth and margins absorb the infrastructure bill. Weighting those outcomes at 27% and 23%, against 50% for the central case, produces approximately $407 over a three-year horizon. At a $300 purchase price, the implied annualized appreciation is about 11%, before dividends, with a central value around 35% higher and a downside scenario about 27% lower. Around that price, I would accumulate: the business has acquired more ways to earn a return, even though each share’s claim on those returns faces a larger capital bill.

The next evidence must connect those two sides. Search needs to turn additional usage into revenue, Cloud needs to absorb higher capacity costs, and TPU deliveries need to convert inventory into cash. My warning remains two consecutive quarters below 12% Search growth, with growth below 8% indicating a more serious problem; neither threshold substitutes for examining what is happening to monetization.

July’s question was whether Google could repeat its technological success. The stronger question now is whether that success can earn money beyond Google’s own products without requiring Alphabet to finance every deployment. By selling systems that other companies finance and operate, Google is creating a way for competitors’ growth to support its own research advantage. The frontier still eats first, but the number of businesses that can help pay for it is expanding.

Subscribe now

• Azure’s economics are improving at the operating level: Microsoft grew Azure 48% while extracting more output from installed GPUs and CPUs, showing that software engineering and workload allocation can raise revenue without proportionate new capacity.

• The more interesting strategic position may sit above the infrastructure: Microsoft is increasingly routing tasks across different models and separating intelligence from identity, permissions and governance. If that works, it can benefit from competition among model providers rather than needing to own the best model itself.

• But the capital burden is still rising faster than cash generation: free cash flow fell despite strong operating growth, capex is surging, and large future lease commitments remain. The thesis now depends on Microsoft turning routing, governance and consumption into durable earnings not simply carrying more AI traffic.

In July, after Satya Nadella told analysts to read 1873, I described Azure as the track, frontier-model companies as the cargo, and Microsoft’s identity and governance services as the right-of-way. At $386, roughly 19 times forward earnings, the attraction was that Microsoft did not need to own the AI operating system to earn an attractive return. It could carry the traffic while retaining the enterprise relationships that made the traffic valuable.

Three months later, Microsoft has demonstrated something more interesting than another increase in demand. It is extracting more revenue from infrastructure already installed, while increasingly choosing which model performs each task inside its products. But the shares closed at $531 before results, and the infrastructure business is consuming substantially more cash. The investment has therefore changed: the discount has largely disappeared, leaving a question about whether Microsoft’s emerging software advantages can justify the capital required underneath them.

A more productive railroad

The headline results were excellent. Revenue rose 19% to $92.3 billion, operating income increased 21% to $45.9 billion, and Azure grew 48%, beating guidance of 44–45%. More revealingly, Azure’s $33.2 billion of revenue represented a record $3.75 billion sequential increase despite delays to third-party capacity deliveries. Its $10.8 billion year-over-year increase accounted for approximately three-quarters of Microsoft’s incremental revenue.

Amy Hood identified three contributors beyond simply installing more equipment. Engineers improved GPU efficiency, discovered and sold underutilized CPU capacity, and benefited from a mix of higher-quality models that generated better realized token pricing. Foundry revenue more than tripled, while Copilot workload throughput increased more than 50% sequentially. Microsoft did not merely buy additional capacity; it made the capacity it already owned more productive.

This matters because the weakest version of the Azure thesis resembles an equipment-financing business: buy expensive machines, rent them out, and repeat when they become obsolete. Engineering improvements offer a better mechanism. More useful work from installed infrastructure can increase revenue without a proportionate increase in investment, while allocating workloads across a large platform creates opportunities unavailable to a provider selling undifferentiated capacity.

That does not establish a permanent improvement in capital returns. Favorable model mix is not the same as repeatable engineering progress, and delayed hardware still needs to arrive. December’s 49–50% Azure growth guidance implies roughly $2.8–3.0 billion of sequential additions, less than September’s, and management is counting on further efficiencies to reach it. Nevertheless, July’s description of Azure as “stable, not gaining” was too pessimistic: the quarter supplied concrete evidence that operating the infrastructure well matters alongside owning it.

The switchyard

The more consequential development sits above the machines. Asked about the division of token value between frontier models and Microsoft’s own, Nadella described competition happening not between annual software contracts, but between individual tasks:

Which is you can now get frontier performance 45% faster, 75% cheaper. Who would want to not take that? If you are a coding agent user, you want, in fact, Anthropic and OpenAI, by the way, to compete for your workload, not like once a quarter, but for every task. That to me is what the value of these routers and Auto, Fusion stuff routing does, and we are very excited about it.

Microsoft says 90% of Copilot queries already run in auto mode, selecting the model task by task. It also described Unilever moving from a frontier model to a Microsoft MAI model trained on its traces. The strategic implication is straightforward: Microsoft can sell a useful outcome while changing the intelligence used to produce it. Competition among suppliers potentially improves the product, lowers its delivery cost and reduces dependence on any single model company.

This is the switchyard: the layer that determines which model performs the work. The evidence today is strongest inside Microsoft’s own products, not across every enterprise AI workload. Even there, however, the distinction matters. A model provider must repeatedly justify its place in the workflow; the company controlling the customer relationship and the routing decision can benefit when several suppliers compete successfully.

There is a useful tension here. Azure benefited this quarter from higher realized token prices, while model routing promises cheaper intelligence. Those need not be contradictory strategies for a company selling both infrastructure and finished products. Microsoft can potentially benefit from scarce capacity today and lower delivery costs inside its applications tomorrow. Its long-term interest is not necessarily an expensive token, but a valuable customer relationship through which it can sell useful work.

Governance is the second control point. Nadella described Microsoft’s approach as separating “intelligence from authority,” with identity, permissions, monitoring and containment enforced outside the model. A company might welcome a better model tomorrow without wanting to rebuild who can access its records, approve a payment or act on its behalf. Microsoft’s established identity, security and workplace services give it a plausible route to providing that continuity.

The activity is substantial: Agent 365’s registry passed 50 million agents, and Purview audited 7.6 billion agent actions. But activity is not yet an independently demonstrated governance business. Microsoft Execution Containers are free and open source; Codex supports them, while further integrations have been announced. Adoption does not oblige customers to buy Agent 365, and Microsoft has not disclosed paid third-party governance economics. The opportunity is to monetize the surrounding management system, not to collect a proprietary toll on an open format.

From seats to work

The reason these control points matter becomes clearer when considering who currently buys the capacity. Microsoft disclosed $10 billion of quarterly revenue from commercial arrangements with OpenAI, including revenue sharing, equivalent to 10.8% of company revenue. That is not all Azure hosting revenue, but it establishes meaningful concentration. Nadella’s answer to a question about enterprise demand explained why he wants a different balance:

But like everything else, the enterprise stuff is not like a quarter-over-quarter. It is a decade-long thing, and that is what I have been waiting for, as opposed to four customers who buy a lot of capacity from you, and you can show a few quarters of growth, and that is not a business as far as I am concerned. So I want the breadth.

Breadth is not simply a preference for smaller customers. A frontier-model company principally needs enormous quantities of compute; an ordinary enterprise also needs applications, organizational context, permissions, security and integration. Microsoft can potentially earn across several layers of that workload, while becoming more embedded in how the customer operates. Routing and governance are therefore not distractions from Azure: they could help turn capacity demand into a more durable enterprise relationship. Nadella outlined the accompanying business-model change in his prepared remarks:

Our business model is expanding from per-seat subscriptions to include consumption, expanding our TAM and ARPU. The subscription brings increasingly capable everyday intelligence to every user. Typically, IT funds this, consolidating frontier advances for everyone across the organization. As agents take on higher-value work, consumption opens up business-funded demand, with spending tied directly to the value these agents deliver.

The important shift is the budget being addressed. Selling another subscription competes for an IT allocation; completing valuable work potentially competes against the cost of doing that work. Subscriptions provide the installed base and everyday access, while consumption allows spending to expand with useful activity rather than employee count alone. As deployment grows, common permissions and organizational context could make the platform more useful, supporting further deployment.

The quarter showed progress, not completion. Copilot daily users increased nearly 50% sequentially, and customers consuming paid credits grew more than 75%, while normalized Microsoft 365 Commercial Cloud revenue rose 19% in constant currency. However, Microsoft 365 commercial seat growth slowed to 5%, Hood acknowledged weaker execution in E3 and standalone products, and Microsoft did not update the paid Copilot seat count. Consumption monetization in core Microsoft 365 will also develop more slowly than in GitHub. These disclosures support growing usage and revenue per seat; they do not yet establish that Copilot has become the operating system for work.

Nor has enterprise diversification been demonstrated conclusively. Commercial remaining performance obligations were $677 billion, almost unchanged from June’s $678 billion, and Microsoft did not update comparable growth excluding OpenAI. That missing measure cannot be scored as a failed threshold, but it leaves July’s breadth test unresolved. The distinction is important: the switchyard thesis has become more credible without yet becoming a separately measurable profit stream.

Financing the track

The income statement makes this transition look easier than the cash-flow statement does. Operating margin reached 49.8%, and operating cash flow increased 29% to $58 billion. Yet cash additions to property and equipment more than doubled to $39.4 billion, taking free cash flow down from $25.7 billion to $18.6 billion. Better utilization is helping the business grow; it has not stopped the infrastructure bill from growing faster than cash generation.

leases went into shorter-lived assets, principally CPUs and GPUs. Next quarter’s capex is expected to exceed $50 billion, while $364.6 billion of additional leases have yet to commence. This is not solely a one-time construction programme followed by decades of inexpensive operation. A substantial portion of the equipment must eventually be refreshed.

Procurement arrangements add another funding requirement. Server-component-related receivables reached $35.7 billion, and Microsoft disclosed $12.6 billion of restricted investments under a supplier agreement. After $12.9 billion of other investing outflows, partly associated with procurement, reported free cash flow left $5.7 billion before $10.2 billion of shareholder distributions. Those figures are not a complete cash reconciliation, and supplier advances are not necessarily permanent costs. They do show why conventional free cash flow does not capture every current demand on Microsoft’s funding. Asked about financing, Hood said:

We are lucky. We have a lot of flexibility in terms of our capacity to raise financing. I have talked about this before. There is a form of financing called leasing, and we have been using that one as a vehicle. We are fortunate to have access in the debt markets, given where our ratings are, and we have not yet, I would say, tapped that.

Access to debt is an option, not evidence of financial distress or a commitment to borrow. Nevertheless, the economic burden is real: Microsoft generally retains the assets, their utilization risk and their replacement requirements. That is the tension running through the quarter. Software can make each machine more productive and each enterprise relationship more valuable, but it must generate enough additional earnings to justify a much larger physical capital base.

What the price buys

At $531, Microsoft traded at approximately 25.6 times next-twelve-month earnings, versus roughly 19 times in July. The easy re-rating argument has therefore largely played out. The remaining opportunity depends principally on earnings growth, supported by the installed enterprise franchise and improving infrastructure economics, rather than requiring a newly proven monopoly over agents.

My central scenario starts with approximately $396 billion of fiscal 2027 revenue. Azure growth moderates to 36%, 28% and 22% over the following three years, while Microsoft 365 Cloud grows 17%, 16% and 15%. That produces roughly $630 billion of fiscal 2030 revenue and, at a 46.5% operating margin, approximately $31.9 of earnings per share. Its fiscal 2029 revenue of about $548 billion remains below consensus of approximately $575 billion.

Applying 25 times fiscal 2030 earnings gives a late-2029 value near $800, roughly 50% above $531, or about 14.5% annualized over three years before dividends. A capacity-led outcome, with approximately $23 of earnings at 20 times, yields about $460; a stronger enterprise-platform outcome, with $38 at 29 times, supports about $1,100. Weighting the capacity, central and stronger-platform outcomes at 20%, 60% and 20% gives approximately $790; these are investment judgments, not measured probabilities.

Cash conversion is the qualification that matters most. Assuming operating cash flow reaches 55.5% of revenue and cash capex falls to 35%, the central case produces approximately $129 billion of fiscal 2030 free cash flow. The $800 valuation represents roughly 45 times that cash flow, not 38 times. There is attractive prospective earnings growth here, but not a cheap cash-generating utility.

The next evidence should determine which business Microsoft is building: repeatable Azure efficiency, broader demand beyond OpenAI, consumption revenue that contributes profit, and enterprises paying to govern rival agents. A slower Azure growth rate alone would not invalidate the thesis, just as more registered agents alone would not validate it. The central question is whether Microsoft retains more economic value as the amount of work passing through its system grows.

The quarter strengthened the case for that outcome. Microsoft is demonstrating that it can operate the railroad better, and its routing capabilities suggest it can influence which intelligence delivers the work. The investment no longer rests on buying overlooked infrastructure cheaply. It rests on earnings compounding through a broad enterprise platform, with a valuable governance business providing additional upside rather than an indispensable assumption.

Subscribe now

• The booking slowdown is less bearish than it looks: new reservations fell sharply, but 13 GW of earlier reservations converted into firm orders and the delivery calendar now reaches 2032. The pace of bookings is increasingly constrained by remaining slots, not necessarily by weaker demand.

• The earnings cycle is finally showing up: Power EBITDA margin reached 21.4% even as services became a smaller part of revenue, suggesting better-priced equipment, manufacturing productivity and higher volumes are improving the economics of the hardware itself.

• The debate has shifted to valuation: much of the revenue path through 2029 is already visible. The key variables are now Power margins, contract-vintage economics, capital allocation and the multiple investors assign to those earnings. At $895, the prospective return is reasonable rather than exceptional.

In February, we proposed a warning signal for GE Vernova’s gas business:

If signings decelerate while existing SRAs convert to orders, the backlog may be peaking rather than compounding. That’s the difference between a cycle and a franchise.

Q3 delivered exactly that pattern. New slot reservations fell from 19 gigawatts in Q1 and 18 in Q2 to eight, while 13 gigawatts of earlier reservations converted into firm orders. It would be convenient to dismiss a test once it produces an uncomfortable result. But July’s Two Clocks had already offered a different interpretation: a factory sold out through 2030 must eventually report fewer new reservations because it has fewer slots left to sell.

“This quarter, we sold multiple slots for ’32 delivery as the demand environment remains strong and customers secure turbines further into the future.”

— Scott Strazik, GE Vernova Q3 2026 earnings call

Q3 makes that distinction more persuasive. GE Vernova’s delivery calendar now reaches 2032, while Power’s EBITDA margin has climbed to 21.4%, eight percentage points above last year. The company is beginning to earn the profits secured when customers competed for scarce capacity. Its order cycle can slow even as its earnings cycle accelerates.

That improves confidence in the business. It does not automatically improve the investment case at any price.

The Booking Paradox

GE Vernova reported $29.3 billion of orders, $12.6 billion of revenue, $1.94 billion of adjusted EBITDA and $5.4 billion of free cash flow. Behind those headline numbers, the gas backlog became more substantial.

During the quarter, GEV signed 11 gigawatts of new commitments, eight in reservations and three in firm orders, converted 13 gigawatts of previous reservations and shipped approximately five. Firm backlog rose from 53 to 64 gigawatts; reservations fell from 63 to 58. Firm backlog now exceeds reservations for the first time.

That matters because a reservation and a firm order are not equally valuable. The company’s $104 billion equipment remaining performance obligations already carry an expected recognition schedule: 60% within two years and 97% within five. Unconverted reservations sit outside that contracted revenue.

Morgan Stanley asked whether financing strain, cancellations or weaker customer commitments explained the slowdown. CEO Scott Strazik said there was no deterioration, including in the cash terms required to secure equipment. The customer mix provides further context: roughly 80% of gas customers are traditional power customers, with demand spanning around 100 customers in 25 countries. An AI infrastructure funding scare might hit the share price without immediately threatening the broader order book.

Management expects at least roughly 13 gigawatts of new commitments in Q4, bringing contracted capacity to 130 gigawatts or more. Strazik suggested that 2027 quarterly commitments could resemble the second half of 2026 rather than its exceptional first half. That makes 2026 a plausible peak in the rate of new signings, not necessarily a peak in demand or future earnings.

The relevant measures are now distinct: new commitments show how quickly the remaining calendar fills; conversions indicate project maturity; shipments measure execution; and the economics of each contract vintage determine the profits still ahead.

The Factory Starts Earning

The clearest evidence came from Power’s margin. EBITDA margin rose to 21.4% from 18.8% in Q2 and 13.4% a year earlier. More revealingly, equipment revenue climbed from $1.97 billion to $2.87 billion sequentially while services revenue fell from $3.51 billion to $3.20 billion. Services declined from 64% to 53% of Power revenue, yet the segment margin expanded.

Historically, services have been the more profitable business. This quarter’s improvement therefore cannot be explained by a richer services mix. Better-priced equipment, higher production, productivity and product mix are improving Power’s economics, although GEV does not disclose equipment margins separately.

When Goldman Sachs asked whether second-half margins were an appropriate starting point for 2027, Strazik answered:

There’s no reason to think that the second half of this year isn’t the right way to start to think about our projection into ‘27.

The next layer of profitability is not yet in the numbers. Management expects 2026 equipment orders to carry about 50% more revenue per kilowatt than 2025 orders, partly because of product mix. That is neither a pure price increase nor a 50% margin increase. Nevertheless, those later-vintage contracts, combined with improving factory output, support a meaningful earnings ramp through 2029.

GEV is increasing annual gas-turbine production from roughly 20 gigawatts toward 24 in 2028 and 30 in 2030. More shipments deliver equipment profits first and enlarge the installed fleet later. Management estimates roughly $500 million of service revenue per HA gigawatt over twenty years. April’s The Window and the Annuity emphasized that second revenue stream. Q3 demonstrates that the equipment sold during the shortage can also earn attractive margins before the service annuity develops.

The Price of a 2032 Slot

Slower reservations are not proof of weakening demand, but they should not be dismissed indefinitely. Scarcity induces supply. Siemens Energy and Mitsubishi Heavy Industries are expanding turbine capacity alongside GEV, creating the possibility that the price of future delivery slots eventually moderates.

Management expects Q4 order revenue per kilowatt below Q3, but Q3 benefited from a richer mix of aeroderivative and 60-hertz turbines. Without comparable pricing by product and delivery year, a lower blended figure does not establish weaker underlying prices. Nor does a sold-out calendar establish that prices will remain elevated forever.

The distinction between contract vintages is critical. Equipment already signed can produce attractive margins when shipped in 2028 or 2029 even if the economics of a new 2032 order become less favorable. Past turbine cycles ended painfully after manufacturers overexpanded; today’s customer advances and firm-order conversions offer some protection, but not immunity from oversupply.

For now, the most useful leading indicators are cancellations, conversion rates, available delivery slots and comparable new-contract pricing. A slowdown in bookings may precede any deterioration in reported earnings by several years.

Buying the Bottleneck

Electrification deserves a more measured interpretation. Q3’s reported EBITDA margin was 19.1%, but organic-adjusted margin was 17.9%, down from 19.1% in Q2. Organic revenue growth moderated from 29% to 16%, although book-to-bill stayed near 1.4. This is a strong business, not yet evidence of an inevitable 30% organic margin.

GEV’s $9 billion agreement to acquire Trench Group extends the logic of Prolec: control critical components that can delay the delivery of much larger systems. Bushings represent only 5–8% of a transformer’s bill of materials, yet their availability can determine when the whole project ships. Trench also supplies instrument transformers and coils used in high-voltage transmission.

Control has value; the question is the price. GEV expects Trench revenue to rise from $1.4 billion in 2026 to $1.9 billion in 2028 at around 36% EBITDA margins. The purchase represents roughly 13 times projected 2028 EBITDA and 22.5 times free cash flow before synergies. Expected annual cost synergies of $50–100 million by 2030 are relatively modest, leaving the investment case dependent on growth and wider supply-chain benefits.

A higher reported Electrification margin after Trench closes in late 2027 will not itself prove value creation. Organic margin improvement and returns on the capital invested are better tests.

Earnings Are Visible. The Multiple Is Not.

At the October 27 reference price of $895, the earnings ramp is hardly a secret. FactSet’s pre-revision 2029 consensus anticipates $69.6 billion of revenue, $16.5 billion of EBITDA and $45.4 in EPS. The Street expects Power to earn a 26.3% EBITDA margin on $39.1 billion of revenue.

Our positive variant is specific: newer turbine contracts and factory productivity can lift Power’s margin above that expectation. We are less optimistic than consensus about Electrification’s organic margins.

Cost Pressure assumes that contracted revenue arrives but cost inflation and execution dilute profitability. Backlog Delivers assumes better-priced equipment and manufacturing productivity flow through as expected. Pricing Power Persists adds stronger profitability across Power and Electrification. The revenue range is relatively narrow because much of the factory schedule is committed; margins and the exit multiple drive the differences.

Backlog Delivers implies a 30% Power margin versus 26.3% in consensus, but a 24% Electrification margin versus the Street’s approximately 24.6%. The resulting $47.1 EPS is only about 4% above consensus. At 24 times earnings, the $1,130 value offers 26% upside over three years, around 8% annualized before dividends. A 20-times multiple on identical earnings would yield $942; 26 times would yield $1,224.

Cash and capital allocation matter to that calculation. GEV generated $15.3 billion of free cash flow over nine months, aided by $19.5 billion of operating inflows from contract liabilities and deferred income. Customer advances finance growth but are not recurring profits. GEV has repurchased approximately $9.9 billion of stock through October 23, leaving roughly $10.1 billion authorized; at $895, that would retire about 11 million shares before dilution. The central scenario’s 255 million diluted shares therefore require continued repurchases, while roughly $4.5 billion of cash funding for Trench and foregone interest income constrain the benefit.

I would hold at $895 rather than chase the stock. Below $800, the same central valuation would offer approximately 12% annualized upside before dividends, a more attractive reward for the remaining execution and multiple risks.

The Next Test

Q4 should bring approximately 13 gigawatts or more of new commitments and continued conversions. More useful still will be the profitability added to equipment backlog: management expects the 2026 figure to be substantially greater than 2025’s roughly $8 billion. Through 2027, shipments near five gigawatts per quarter and Power margins above 20% would support the earnings ramp; sustained margin deterioration despite healthy volumes would challenge it.

Electrification must demonstrate organic improvement, and Trench must eventually earn its purchase price. Service economics also warrant attention: the 10-Q records $293 million of adverse changes in estimated long-term service-contract profitability over the first nine months. Finally, free cash flow must remain strong when customer advances moderate.

GEV is operationally stronger than it was in July. The backlog is firmer, the delivery calendar stretches further, and better-priced equipment is already improving earnings. But consensus recognizes much of that progress. At $895, the prospective return is reasonable rather than exceptional.

The order cycle may be maturing, but the earnings cycle is only beginning to deliver. That is the distinction Q3 establishes, and the one that matters for the next three years.

Subscribe now

• Anthropic has already demonstrated extraordinary commercial demand: quarterly revenue rose from $376 million in Q1 2025 to roughly $11.6 billion in Q2 2026, while revenue per dollar of compute improved from $0.40 to $2.08 and the company reached GAAP operating profitability.

• But those economics may partly reflect scarcity: compute is contracted 12–18 months in advance, so revenue recently grew much faster than available capacity. As the next wave of infrastructure arrives, Anthropic must prove that higher utilization and model efficiency not simply constrained supply are driving the operating leverage.

• The real 2027 test is flywheel versus treadmill: better models can expand usage and deepen customer relationships, but if OpenAI and Google improve just as quickly, productivity gains may simply fund a larger research race. The key question is whether each generation strengthens Anthropic’s capability, cost or customer advantage enough to make the next one more profitable.

In 2021 and 2022, container shipping enjoyed the most profitable period in its history. Pandemic disruptions had constrained capacity, freight rates multiplied, and shipping companies rushed to order new vessels. Maersk generated $30.9 billion of operating profit in 2022, an extraordinary result for an industry accustomed to considerably thinner margins.

The economics changed before much of that new capacity could arrive. By the second quarter of 2023, Maersk’s EBITDA had fallen 72% from a year earlier. Freight rates had normalized, companies were destocking, and the exceptional conditions of the pandemic had disappeared. The ships ordered during the shortage would arrive in a market that had already changed.

The lesson was not that ships were unnecessary or that global trade had stopped growing. It was that capacity committed during a shortage could arrive after the economics supporting that shortage had changed.

Anthropic’s October 26 S-1 raises a similar question about artificial intelligence, although with an important difference. Unlike container shipping, increasingly capable intelligence can create entirely new sources of demand rather than merely serve an existing market.

The filing reveals a business generating $11.6 billion of quarterly revenue, with extraordinary customer growth and its first quarter of GAAP operating profitability. It also reveals hundreds of billions of dollars in infrastructure commitments, much of which will become operational over the next several years.

On the same day, Broadcom and a consortium of financial institutions announced a $63 billion financing arrangement to acquire AI accelerators and networking equipment for lease to Anthropic. The transaction finances part of an existing compute agreement, not additional orders.

Together, these disclosures capture the central tension in the AI trade. Demand for intelligence is real and expanding rapidly. But the infrastructure required to produce it is being financed on the expectation that demand will become considerably larger.

My reading is that Anthropic has sufficient financial resources to make a near-term liquidity crisis unlikely. But 2027 will test whether its extraordinary operating leverage represents a durable economic advantage or the temporary benefit of revenue growing faster than available compute.

The distinction matters well beyond Anthropic. An intelligence business that compounds its economic advantage could support successive waves of infrastructure investment. An escalating research arms race could also generate enormous hardware demand while leaving the laboratories conducting that research with disappointing shareholder returns.

Efficient Because Compute Was Scarce

Anthropic’s historical financial results are remarkable.

Quarterly revenue increased from $376 million in the first quarter of 2025 to approximately $11.6 billion in the second quarter of 2026. Over the same period, compute and infrastructure expenditure rose from $944 million to approximately $5.6 billion.

Revenue per dollar of compute consequently increased from $0.40 to $2.08. At the end of 2025, the ratio had been just $0.69. It then increased to $1.40 in the first quarter of 2026 and $2.08 in the second.

The improvement was sufficient to turn a $1.7 billion GAAP operating loss in Q1 into a $974 million operating profit in Q2.

There has clearly been substantial improvement in commercial efficiency. But the mechanism is not simply that Anthropic learned to run its models at a fraction of their previous cost.

Compute is generally contracted 12 to 18 months before it becomes available. Infrastructure expenses are therefore relatively fixed in the near term, while revenue can change rapidly as new models and products attract customers. In the second quarter, revenue grew 144% sequentially while compute expenditure increased 64%.

Customers arrived considerably faster than the capacity needed to serve them.

Anthropic acknowledges having repeatedly underestimated its compute requirements, leaving the company capacity-constrained and affecting both product development and deployment. The shortage helps explain the recent improvement in utilization and operating leverage, alongside genuine advances in model efficiency, pricing and product mix.

It also identifies the risk ahead.

Anthropic is now securing much larger amounts of capacity. Once that capacity becomes available, expenses begin before the company necessarily generates the additional revenue expected from the investment.

The current economics provide a useful benchmark. In Q2 2026, operating expenses other than compute represented approximately 43.5% of revenue. Holding that ratio constant, Anthropic needs roughly $1.77 of revenue per dollar of compute expenditure to break even at the GAAP operating level. Only Q2 2026 clears that threshold among the six quarters disclosed.

That ratio is not a permanent break-even point. Personnel costs, distribution expenses and stock-based compensation will change as the company scales. Nor does revenue per compute dollar isolate technological efficiency from utilization and pricing.

Nevertheless, it demonstrates how quickly exceptional profitability can turn into operating losses when compute represents such a large portion of expenditure.

And it raises a deeper question: why must Anthropic keep buying so much compute even after its existing models become more efficient?

A Flywheel or a Competitive Treadmill?

Anthropic’s economics are more complicated than those of a conventional software company because compute is not used exclusively to serve paying customers. It is also used to develop better models.

The supplemental financial statements show $2.38 billion of cost of revenue in the first half of 2026 against $16.29 billion of revenue, implying an 85% functional gross margin.

That is extraordinary. But during the same six months, Anthropic spent $9.74 billion on research and development, while total compute and infrastructure expenditure reached $8.95 billion.

These expense classifications do not reveal precisely how compute is divided among inference, training, research and internal usage. The 85% gross margin should therefore not be interpreted as the economic margin on all customer-serving compute.

What is clear is that Anthropic is investing enormous resources to develop the next generation of intelligence, not simply to deliver the intelligence it has already developed.

The economics of those activities are fundamentally different. Serving an existing model becomes more profitable as utilization increases, inference improves and customers consume more intelligence. Developing the next frontier model is partly a competitive necessity.

If OpenAI or Google produces a substantially more capable model, Anthropic may need to accelerate its research regardless of how profitable its current products have become.

The S-1 reports that models released in the preceding six months generated more than 80% of Anthropic’s third-quarter revenue. That demonstrates rapid migration toward newer Anthropic models, not necessarily switching between competing suppliers. But independent transaction data also shows how competitive the market is. Among businesses tracked by Ramp, OpenAI narrowed Anthropic’s adoption lead during September even as Anthropic continued growing.

Anthropic does not simply own a technological lead and collect rent from it indefinitely. It must repeatedly defend that lead through better models, lower costs and deeper customer relationships.

Yet there is an equally powerful argument for compounding.

Better models make more activities economically viable. Customers consume more intelligence and integrate it into increasingly consequential workflows. That generates revenue to finance further research. Commercial deployment also provides information about reliability, costs and the capabilities customers value most, helping direct subsequent development.

Anthropic’s customer numbers support this possibility. More than 11,900 customers were generating at least $100,000 in trailing annual revenue by September, compared with approximately 1,500 at the end of 2025. The company also reported net revenue retention of approximately 900%.

That latter figure deserves careful interpretation. It means revenue from a qualifying cohort was approximately nine times its level in the corresponding quarter a year earlier. But customers need only have generated $1 in the base quarter to qualify, and consumer Pro and Max subscriptions are excluded. It is powerful evidence of expansion from small initial deployments, not proof of durable retention or switching costs.

The potential flywheel is nevertheless real. If customers embed Claude deeply into critical workflows, each generation of better intelligence could expand usage while making displacement progressively harder.

Anthropic is also using Claude to accelerate its own research and engineering. More than 90% of contributions to its codebase in September were written by Claude, and daily code contributions per engineer increased more than 45 times relative to 2024. Those figures do not establish equivalent gains in useful engineering productivity, but they demonstrate how extensively AI has entered the development process.

Now consider the alternative.

Suppose AI allows Anthropic to conduct twice as much productive research with the same resources. If OpenAI and Google achieve comparable improvements, all three laboratories may simply pursue more ambitious research.

The cost of reaching yesterday’s frontier falls, while expenditure on reaching tomorrow’s frontier continues increasing.

All three become more productive. None necessarily becomes more profitable.

That is the competitive treadmill: technological progress expands the market but competition prevents individual laboratories from retaining much of the value they create.

Both outcomes can coexist. Intelligence may transform software development, scientific research and engineering while competitive research expenditure absorbs much of the resulting surplus.

The question is not whether Claude improves its successors. It is whether those improvements create a widening advantage in capability, cost or customer relationships that competitors cannot readily replicate.

That distinction has enormous implications for semiconductor demand. A research arms race can support extraordinary accelerator and networking expenditure even when laboratory profitability deteriorates. A strong semiconductor cycle therefore does not automatically imply an equally strong profitability cycle for the companies purchasing the chips.

The Bill

The S-1 shows the scale of the next infrastructure wave.

Anthropic has committed to minimum spending of approximately $114.5 billion with Google, $110 billion with Amazon and $44.9 billion with Microsoft over multiyear periods. These agreements are generally non-cancelable, subject to defined exceptions, and include payment obligations when actual spending falls below agreed minimums.

Separate Broadcom equipment leases involve approximately $161 billion of payments excluding interest. The GPU arrangements include up to $120.5 billion with SpaceXAI and approximately $86 billion with Nscale and Lambda. Additional TPU arrangements for 2028–31 involve estimated future payments of $125 billion to $150 billion.

These amounts should not simply be added into one definitive liability figure. Contractual periods differ, some categories may overlap, delivery conditions matter, and not every commitment represents an unconditional payment obligation.

But the mismatch is unmistakable.

Anthropic’s customers had committed $10.8 billion of remaining performance obligations at June 30, while the company had secured infrastructure involving hundreds of billions in potential future expenditure.

That comparison does not mean customers will spend only $10.8 billion. Anthropic’s revenue is predominantly consumption-based, so much of its future demand is not reflected in contractual backlog.

It means something more important: customers largely buy intelligence as they use it, while Anthropic must secure much of the capacity required to produce that intelligence years beforehand.

Whether that mismatch becomes dangerous depends on how much demand materializes, how quickly the capacity arrives and who ultimately bears the financial obligations.

Who Pays Meanwhile?

There are four sources of funding: customer receipts, previously raised equity, additional financing and capital supplied by infrastructure partners.

Anthropic held approximately $131.7 billion of cash and short-term investments at June 30, 2026. That is a formidable liquidity position, although the money was not primarily accumulated through operating profits. During the first half, operating cash flow was positive at $3.9 billion while financing activities generated approximately $110.8 billion.

The cash balance was built overwhelmingly through outside capital.

Nor should Anthropic’s $80.7 billion first-half GAAP net loss be confused with operating cash consumed. Approximately $72.4 billion reflected fair-value changes in convertible instruments and related liabilities, principally driven by the company’s rising valuation.

The earlier financing arrangements established much of the underlying economic dilution when they were agreed. Subsequent accounting losses largely reflected the increased value of those conversion rights.

The forward question is how much of the next investment cycle customer revenue can finance.

Consider an illustrative 2027 scenario. If Anthropic generates $113 billion of revenue, incurs $95 billion of compute expense and spends another $34 billion on other operating expenses, the resulting operating shortfall would be approximately $16 billion.

Against its existing liquidity, that appears manageable. But it is not a complete cash-flow forecast. Compute expenses and cash payments do not necessarily coincide, and finance-lease principal, capital expenditure, working capital and other obligations can materially change the outcome.

Nor can we yet establish precisely how much cash will be required each year. The preliminary S-1 leaves important aggregate contractual-obligation schedules incomplete, while potential IPO proceeds remain unspecified.

My base case is therefore not an immediate funding crisis during the 2027–28 expansion. Anthropic has substantial financial resources and a plausible path to supporting a meaningful portion of its expenditure through revenue. That is different from claiming every obligation has already been funded.

The company also has approximately $77 billion of conditional financing capacity under arrangements involving Google, Amazon and Broadcom. Those facilities are not equivalent to cash already received.

The Broadcom arrangement is particularly revealing. Anthropic can potentially issue up to $42 billion of convertible notes to finance equipment-lease obligations associated with Broadcom.

In economic substance, that is a form of vendor financing. It may be a rational way to fund productive infrastructure before the associated revenues arrive. But it can also shift part of the burden to future shareholders.

A rising valuation makes this financing easier. More capital supports better models, which attract customers and potentially justify the valuation. If competition absorbs those improvements, however, repeated financing may become necessary merely to remain competitive.

Anthropic’s ability to finance the next buildout is stronger than its demonstrated ability to finance successive buildouts entirely from internally generated cash.

Where Does the Risk Go?

The financing arrangements also reveal that risk is distributed unevenly across the AI infrastructure supply chain.

Google, Amazon and Microsoft benefit from substantial minimum-spend protections. These improve revenue visibility but do not eliminate counterparty credit risk. Their value ultimately depends on Anthropic’s ability to meet its obligations.

SpaceXAI is different. Under agreements covering up to $120.5 billion of Nvidia-based capacity, fees are expected to reach approximately $3.1 billion per month once all capacity is delivered. Following initial non-cancelable periods ending in January and March 2027, either party can terminate the relevant agreements with 90 days’ notice.

That flexibility works in both directions. If Anthropic’s demand disappoints, it may be able to reduce a substantial portion of its compute bill. Conversely, if GPU rental prices increase materially, the provider may have an incentive to terminate rather than continue supplying capacity on less attractive terms.

Nscale and Lambda carry financing, delivery and acceptance conditions, but these should not be confused with unrestricted rights to cancel because customer demand is weak.

The $63 billion Broadcom financing illustrates another distinction. A special-purpose vehicle will acquire Broadcom equipment for lease to Anthropic, with senior financing arranged by Bank of America, Citi and Morgan Stanley and committed second-lien financing from Blackstone.

Broadcom is providing residual-value support equivalent to the first-lien portion. Its equipment revenue may be supported by firm demand, but the company retains contingent exposure to the value of the financed hardware.

The financing redistributes risk. It does not eliminate it.

My working hypothesis is that leveraged infrastructure providers with flexible customer agreements, substantial financial obligations and limited alternative demand for their capacity could experience stress before well-capitalized hyperscalers or Anthropic itself. Whether that happens depends on leverage, recourse, collateral values and actual contract terms.

For semiconductor investors, the broader implication is important. Nvidia, Broadcom and the networking and equipment supply chain can generate substantial revenue while the infrastructure owners have yet to demonstrate attractive returns on the capital deployed.

An escalating research competition can therefore remain excellent for hardware demand and much less attractive for laboratory shareholders. But the divergence cannot persist indefinitely if the laboratories lose access to financing or become unwilling to commit to successive infrastructure cycles.

The AI trade is not one homogeneous investment. The laboratories, hyperscalers, semiconductor suppliers and infrastructure financiers occupy different positions in the same economic chain.

The Conversion Test

The central question for 2027 is whether Anthropic’s expanding compute base generates enough commercial value to strengthen its business, rather than simply raising the cost of remaining competitive.

Anthropic is reducing the cost of useful intelligence. In September, Opus 5.5 lowered customer cost per task by approximately 40% relative to Opus 5, through a combination of model efficiency and pricing improvements.

That is not the same as a 40% price reduction. A more efficient model can complete a task with less compute even when its price per unit changes much less.

The distinction matters because lower costs can stimulate additional consumption without necessarily compressing provider margins proportionately.

Coding became an early commercial market partly because useful outputs can be verified. Similar principles could increasingly apply to scientific research, engineering and professional workflows. Autonomous agents performing lengthy sequences of reasoning, verification and tool use may consume vastly more intelligence than a conventional chatbot interaction.

The potential market is not restricted to substituting existing software or labor expenditure. Better intelligence could enable discoveries, products and activities that would not otherwise have existed.

That is the strongest argument for the scale of investment now underway.

But growing demand does not automatically establish attractive returns for Anthropic. The commercial value created must exceed both the cost of delivering intelligence and the continuing expenditure required to defend its technological position.

An early financial indicator is revenue per dollar of compute.

Consider a future quarter in which Anthropic incurs $15 billion of compute expenditure. At $1.80 of revenue per compute dollar, it would generate $27 billion of revenue. If non-compute operating expenses remain at Q2 2026’s 43.5% of revenue, GAAP operating profit would be only approximately $250 million.

At $1.40 per compute dollar, the same illustrative cost structure would produce an operating loss of approximately $3.1 billion. At $2.08, operating profit would approach $2.6 billion.

The same infrastructure expenditure can therefore support very different earnings outcomes depending on how quickly monetization catches up.

These calculations are deliberately simplified. They are not forecasts of cash flow after finance-lease principal and other infrastructure payments. They also do not establish an attractive return on invested capital.

My expectation is that revenue per dollar of compute will fall from Q2’s exceptional $2.08 as substantial new capacity becomes operational. I expect at least one GAAP operating-loss quarter during the first half of 2027.

That would not invalidate Anthropic’s commercial progress. It would demonstrate that the infrastructure investment cycle had temporarily moved ahead of monetization.

If Anthropic maintains revenue per dollar of compute around $1.80 or better in the first half of 2027, while quarterly compute expenditure reaches at least $15 billion, my expectation of meaningful operating-margin compression would be too pessimistic.

The more demanding test is whether the resulting profits translate into cash after infrastructure financing, and whether each investment in research and compute makes Anthropic more competitively differentiated.

Operating profitability is necessary evidence. It is not sufficient proof of a compounding economic advantage.

The Call

I take three conclusions from Anthropic’s S-1.

First, the commercial market for intelligence is real and expanding extraordinarily quickly. Customer growth, cohort expansion and improving operating efficiency provide substantial evidence that enterprises are finding economic value in Claude. The opportunity could become considerably larger as AI moves beyond coding into scientific discovery, engineering and increasingly autonomous work.

Second, Anthropic’s financial position is stronger than the headline infrastructure commitments alone suggest. Its accumulated liquidity, growing revenue and contractual flexibility in certain agreements provide meaningful protection against near-term disappointment.

My base case is therefore not a liquidity crisis as the next capacity wave arrives.

Third, that wave will begin revealing whether Anthropic is building a compounding economic advantage or an exceptionally well-funded competitive treadmill.

I expect a strong October run-rate revenue disclosure, followed by pressure on operating margins as new capacity enters service in 2027. An October annualized run rate above approximately $78 billion would strengthen the near-term demand outlook, although it would not settle the profitability question.

The first half of 2027 will provide the more important operating test. Sustained revenue per compute dollar around or above $1.80, at substantially higher compute expenditure, would challenge my expectation of margin compression.

By the second half of 2027, I would pay greater attention to cash generation after infrastructure payments, the growth of customer commitments, research expenditure and Anthropic’s dependence on partner-financing facilities. The strongest evidence for a genuine flywheel would be improving economics alongside customer retention and competitive differentiation, not simply another increase in revenue.

For the wider AI trade, the timing matters. Semiconductor suppliers can benefit from several quarters of extraordinary infrastructure demand even while the laboratories purchasing that capacity experience deteriorating margins. Eventually, however, the returns generated by the infrastructure determine how much additional capacity its owners are willing and able to finance.

The shipping analogy reminds us that capacity and demand operate on different timelines. AI introduces an additional possibility: improvements in the underlying technology can expand the market itself, potentially justifying far more infrastructure than current consumption suggests.

That possibility makes Anthropic fundamentally different from a shipping company ordering vessels during a freight shortage.

But it does not eliminate the economic question.

Anthropic has demonstrated that customers are willing to pay extraordinary sums for intelligence. It has also demonstrated that revenue can grow considerably faster than the compute infrastructure required to produce it. What it has not yet demonstrated is that the profits from one generation of intelligence can finance the next while strengthening its competitive position.

The question is not simply whether Anthropic can pay for the next generation of intelligence. It is whether each generation makes the following one more profitable to produce.

That is what 2027 will begin to tell us.

$ANT

Subscribe now

• The competitive position is strengthening: AI agents can generate far more design possibilities, but those designs still need simulation, verification, signoff and implementation. Cadence’s 16 full-flow wins suggest it is becoming more deeply embedded in that underlying engineering stack.

• The monetization thesis is less automatic than we previously assumed: an agent can call Cadence tools more often without creating equivalent revenue if the activity stays inside existing licences. Incremental value only becomes incremental revenue through new agents, more capacity, hardware or larger contracts.

• The stock already reflects a lot of quality: the quarter was strong and guidance rose, but underlying growth is still mid-teens, cRPO has stalled at $4.2 billion, cash conversion needs scrutiny, and the valuation requires evidence that agent-driven adoption accelerates earnings rather than merely customer productivity.

On September 30, Synopsys announced GPT-Synopsys, a specialized chip-design model being developed with OpenAI. Four weeks later, alongside its third-quarter results, Cadence introduced Maestro, a family of domain-specialized models integrated with its engineering tools and capable of routing tasks to whichever model performs best, including frontier AI models.

The announcements appeared to establish competing strategies. Synopsys was partnering with a frontier AI laboratory, while Cadence was developing more of the intelligence itself. Both, however, are pursuing the same objective: remaining central to the customer’s design workflow as engineers increasingly delegate work to autonomous agents. Asked why customers should choose Cadence, CEO Anirudh Devgan made his ambition explicit:

We want to be the primary interface to our customers.

The quarter was strong. Revenue grew 22% to $1.633 billion, backlog reached a record $8.2 billion, and management raised its 2026 revenue outlook for the third time, to $6.4 billion. Yet the more interesting development is what the announcements and earnings call reveal about the industry’s changing economics.

Cadence does not need to own every AI agent. It needs the proliferation of agents to increase customers’ paid dependence on its engineering capabilities. Q3 strengthens the evidence for the first half of that proposition. The second remains the investment question.

The Engine Beneath the Agent

The essential distinction in semiconductor design is between proposing a chip and proving that it works.

A sophisticated AI model can generate circuit descriptions, explore architectures and optimize design objectives. But those proposals must still satisfy timing, power, physical and manufacturing constraints. Engineers depend on specialized tools to translate promising designs into silicon that can be manufactured and operate reliably.

This creates an important distinction between the intelligence generating the work and the engineering infrastructure validating it. As agents become more capable, they may explore substantially more alternatives than human engineers could evaluate manually. Each proposed design requires simulation, optimization and verification. Better models can therefore increase demand for the underlying engineering tools even when Cadence does not supply the agent.

CFO John Wall acknowledged that customers have always built workflows around Cadence tools and will increasingly develop their own agents. Cadence’s strategy accommodates that choice.

This gives the company two opportunities. When customers adopt ChipStack, ViraStack, InnoStack or AuraStack, Cadence can sell agents and orchestration alongside its existing tools. When customers use their own agents, Cadence can still benefit from demand for implementation, simulation, signoff and verification capacity.

Maestro strengthens the first opportunity without sacrificing the second. Its specialized models can use Cadence’s engines to evaluate their outputs, while its router preserves flexibility to use frontier models when those perform better. Cadence owns the engineering tools supplying the feedback, but that is not the same as possessing a cross-customer data network effect. Confidential customer designs cannot simply be pooled into a shared learning system.

The competitive advantage is instead the depth of qualified engineering tools and their integration across workflows. Devgan emphasized that customers increasingly want synthesis, implementation and signoff to work together. An agent operating across a complete design flow makes that integration more valuable because the performance of one tool increasingly depends on its interactions with the others.

Cadence reported 16 new digital and signoff full-flow wins during Q3. That is the most persuasive evidence of competitive momentum. Core EDA revenue grew 20%, but this includes emulation hardware, so the growth rate cannot independently establish software market-share gains.

The opportunity also extends beyond conventional semiconductor companies. Devgan said approximately 45% of Cadence’s business now comes from systems companies, including businesses developing their own silicon. Easier design tools may encourage more companies to undertake custom-chip programs, creating additional demand for software, IP, hardware and services.

There are two possible growth mechanisms: Cadence earns more from existing design programs, or a broader population of customers begins designing chips. Neither requires Cadence to impose a tariff on every design iteration. Both depend on its engineering tools remaining competitive.

More Work Is Not More Money

This brings us back to the distinction we established in July.

In February’s Design Throughput Toll, we argued that AI agents could eventually break the relationship between Cadence’s revenue and the number of engineers its customers employ. If agents could run vastly more design iterations, Cadence might capture value from engineering throughput rather than primarily from conventional licences.

The technological logic remains plausible. The commercial logic was too compressed.

An agent can call a simulation engine a hundred times more often without producing a hundred times more revenue. If those calls fall within an existing licence, Cadence may earn nothing incremental. Greater usage creates revenue only when it requires additional contracted capacity, new products or larger agreements.

Wall clarified Cadence’s intended economic model this quarter. Customers can use their own LLM infrastructure and bear the associated compute costs. Cadence monetizes design agents and orchestration, while continuing to earn revenue from its engineering software and hardware.

This is a sensible division of the value chain. Cadence can participate in AI-enabled engineering without becoming a reseller of general-purpose inference compute. Synopsys’s GPT-Synopsys offering, by comparison, is intended to bundle models, compute and engineering licences. Neither model is inherently superior, but Cadence’s approach is designed to preserve its attractive software economics.

The distinction matters when non-GAAP gross margin is already 88.8%. Owning more AI infrastructure would not necessarily improve the quality of that business.

There are three realistic routes to incremental monetization: customers purchase agent subscriptions and orchestration capabilities; increased workloads require additional engineering licences, compute capacity or emulation hardware; and broader dependence on Cadence’s products supports larger contracts at renewal.

The company has demonstrated the products and described these routes. It has not disclosed the commercial contribution from agents or established how much incremental usage translates into incremental spending.

There is another uncertainty. Better agents might reduce wasted iterations per design even as they increase the number of designs attempted. The net effect on engine consumption is therefore not automatic.

The economic opportunity depends on whether faster, cheaper design expands the total amount of engineering work enough to outweigh greater efficiency within individual projects, and whether Cadence captures a meaningful share of that expansion.

A Strong Quarter, but Not a New Growth Regime

The headline results were excellent, although Cadence has made beating expectations something of a habit.

Revenue exceeded Bloomberg consensus by approximately $20 million and non-GAAP EPS beat by six cents. More importantly, management raised its full-year revenue guidance by $102 million to a midpoint of $6.402 billion. Only about $23 million reflected Q3 exceeding the previous guidance midpoint; the remaining $79 million increased implied fourth-quarter expectations.

This was a genuine forward raise, not simply the arithmetic of a quarterly beat. The historical comparisons nevertheless require some care.

Sources: Cadence, Bloomberg. Q4 2026 uses the guidance midpoint. Two-year comparisons include acquisition effects.

Management describes growth as accelerating and broadening. The description is reasonable against the weak second half of 2025, but the two-year figures tell a more restrained story. Q3’s two-year revenue CAGR was approximately 16%, while Q4 guidance implies about 12%.

Part of the apparent acceleration reflects easier comparisons, hardware-delivery timing and the acquired Hexagon design-and-engineering business. Wall characterized underlying revenue growth excluding Hexagon as mid-teens, with high-teens EPS growth and incremental margins around 54–55%.

Those are excellent economics. They are also broadly consistent with Cadence’s established business rather than proof of a new agent-driven growth regime.

Contracted revenue is similarly encouraging but inconclusive. Backlog reached $8.2 billion, yet current remaining performance obligations, representing revenue expected over the next twelve months, remained flat sequentially at $4.2 billion.

Wall attributed strong bookings during a relatively light renewal year to customer expansion and add-on opportunities. That is precisely where agent monetization could appear before major renewals. But without quantified add-on growth or revenue per comparable customer, we cannot separate agent-driven expansion from ordinary design activity and competitive wins.

Recurring revenue grew 18%, compared with approximately 24% in Q2. The earlier quarter included Hexagon’s contribution, while Q3 faced an unusual prior-year China comparison. Cadence has not provided enough comparable information to isolate underlying agent-driven growth.

Our July claim that decoupling was already becoming visible in recurring revenue was therefore premature. The business is growing strongly, but the evidence does not yet identify the source of that growth.

Profitability adds another qualification. Non-GAAP gross margin rose from 88.0% to 88.8%, while operating margin fell from 47.6% to 44.1%. Hexagon integration and investment explain much of the pressure, but operating leverage needs to recover as the acquired business matures.

Q3 free cash flow rose 31% to $362 million, although full-year operating cash flow guidance remains around $2 billion despite the earnings raise. Days sales outstanding have remained elevated at 67, 65 and 66 days across 2026, compared with 55 days in Q3 last year.

We had already identified persistent DSO elevation as a warning. Three quarters are enough to treat it as something requiring explanation rather than dismissing it as timing.

These figures do not undermine Cadence’s technological advantage. They determine how much of that advantage reaches shareholders.

Three Possible Cadences

The valuation question is whether stronger adoption produces enough incremental earnings and cash flow to justify today’s premium.

Bloomberg’s pre-release 2027 consensus stood near $7.18 billion of revenue and $9.55 of adjusted EPS. Against the new 2026 guidance, those forecasts imply roughly 12% revenue growth and 14% EPS growth.

At the October 26 closing price of $322.17, Cadence trades at approximately 38 times 2026 adjusted earnings and 34 times pre-release 2027 consensus.

An opening 2027 revenue guide near 12% would not itself signal slowing demand. Cadence has a history of conservative initial guidance followed by upward revisions.

I would retain July’s three operating scenarios and 25%/50%/25% probabilities, rebase the forecasts to the higher 2026 revenue outlook, and explicitly model the financial assumptions that were previously implicit.

Estimates, not company guidance. Assumes 16.5% tax, $80 million annual net non-operating expense, and scenario-specific free-cash-flow conversion of 78–87% of adjusted net income. Share assumptions reflect net dilution after repurchases.

In Essential Utility, agents improve customer productivity without materially expanding paid engineering capacity. Growth settles at 9%, hardware normalizes and the premium multiple contracts. Earnings rise, yet shareholders lose money because the valuation falls faster.

In Premium Add-On, agents, broader tool attachment, full-flow wins and system-company expansion sustain 13% revenue growth. Operating margins recover to 46%, and earnings compound without requiring a radical change in Cadence’s commercial model.

In Throughput Toll, a meaningful share of additional engineering work becomes paid capacity and larger recurring commitments. Revenue compounds at 17%, margins reach 48%, and the company earns a premium multiple on substantially higher profits. This scenario does not require literal per-call pricing.

The market already recognizes Cadence as an exceptional business. Our variant perception must therefore concern how much faster its earnings can compound; not simply how indispensable its software becomes.

What Would Change Our View

The February 2027 report provides the first meaningful test. Q4 is traditionally Cadence’s strongest bookings quarter, and year-end cRPO approaching or exceeding $4.4 billion would indicate a contracted book growing more convincingly with the business. Acquisition-adjusted recurring growth above 15% would reinforce that evidence; sustained growth below 13% would weaken it.

Neither metric would independently prove agent monetization. Renewal timing, hardware mix and conventional customer expansion still matter.

The next test is the opening 2027 outlook. Guidance above 13.5% revenue growth would be encouraging, while below 11% would warrant greater scrutiny. More important is whether actual 2027 growth ultimately exceeds 14% while non-GAAP operating margins recover above 45%.

The decisive evidence should emerge through 2027: larger software commitments, quantified add-on growth, improved renewal economics, or rising revenue per comparable customer. Better cash conversion and DSO moving toward historical levels would establish that the additional engineering value is reaching shareholders.

There are also risks outside the thesis. Synopsys may capture more of the agent interface through its OpenAI partnership, frontier models could erode the advantage of domain-specialized models, and Chinese demand remains exposed to export-control uncertainty. None is resolved by a strong quarter.

The Next Turn

July ended with a distinction between creating engineering value and capturing it. Three months later, Cadence has clarified how it intends to monetize that value, while full-flow wins and agent deployments provide stronger evidence of adoption.

The underlying advantage is clearer. As AI makes it easier to propose designs, the engineering infrastructure required to validate, optimize and manufacture them remains essential. Cadence has an opportunity to become more valuable as that infrastructure serves increasingly automated workflows.

What remains unproven is the capture rate. More design activity must eventually translate into larger contracts, additional capacity and stronger cash generation, rather than simply allowing customers to accomplish more under existing agreements.

The engine is becoming more important. Whether Cadence gets paid proportionately for running it harder is still an open question.

Until that evidence arrives, I would own Cadence as a high-quality compounder at a demanding but defensible valuation, not underwrite it as a newly discovered throughput monopoly whose pricing power the market has missed.

Subscribe now

• The operating thesis is working: revenue rose 138%, product margins improved, and customers are clearly paying for something scarce electricity delivered months or years before the grid can provide it.

• But the balance sheet is carrying more of the delay: receivables, contract assets, inventory and supplier advances have all risen sharply, while Bloom has committed $5.8 billion to suppliers. The company is increasingly absorbing scheduling risk, so customers do not have to wait.

• That changes the investment question: Bloom may create enormous value by delivering power early, but shareholders only benefit if the company can retain enough of that value after funding inventory, supplier commitments, installation and customer timing.

In May, we borrowed Matthew Boulton’s famous description of his business: “I sell here, Sir, what all the world desires to have, POWER.” Boulton and Watt did not merely sell steam engines. They charged customers for a portion of the coal their machines saved, capturing some of the economic value of removing a constraint. Their engines also freed factories from the rivers that had determined where industrial production could take place.

Our argument was that AI had made the electricity grid the new river. A data center cannot earn anything until power arrives at its chosen site. Bloom’s fuel cells offered a way to move electricity generation to where computing demand already existed, rather than waiting years for the grid to catch up.

We ended with a more demanding test:

“Bloom will deserve its premium only when it can charge for the time it saves without giving too much of that value back through warrants, financing structures, margin concessions, or accounting complexity.”

The third quarter answers the first half emphatically. Revenue rose 138% to $1.24 billion, beating our $1.09–1.15 billion forecast, while adjusted earnings reached $0.94 per share against consensus near $0.69. Management raised full-year revenue guidance to $4.4–4.6 billion. The factory is delivering, and customers are buying.

The second half is becoming more complicated. Bloom generated $303 million of adjusted net income but reported negative $181 million of free cash flow. Contract assets and supplier advances are rising, and the 10-Q disclosed $5.8 billion of supplier commitments, including $2.6 billion payable in 2027.

Bloom has become a better business. Whether it has become a better stock depends on what the company must spend, commit and finance to sustain that growth.

The Factory Is Proving Itself

The electricity shortage is no longer a debate about future demand. Large AI campuses require power at a scale and timetable that traditional utility infrastructure often cannot accommodate. For a customer with billions of dollars of computing equipment awaiting deployment, electricity available in six months may be worth substantially more than cheaper electricity available in three years.

Bloom’s solid-oxide fuel cells address that constraint. They convert natural gas into electricity without combustion, can be installed in modular increments, and may face fewer local-emissions and water-use obstacles than conventional generation. They still need fuel, permits and construction, but they can shorten part of the deployment process.

CEO KR Sridhar offered two examples:

“Just 15 months ago, we announced our first project with Oracle and delivered power to their Utah site in less than two months. We are achieving the same thing with Nebius in New Jersey.”

At Nebius, planning approvals took approximately three months after contract signing, followed by another two months before the first power block was ready.

That is what customers are buying: time.

The manufacturing results reinforce the point. Product revenue rose 179% year over year to $1.07 billion. Non-GAAP product gross margin improved from 35.3% in Q1 to 38.4% in Q3 despite capacity expansion. Service margin rose from 18.0% to 21.4%, while installation margin recovered from negative 29.4% to negative 3.5%.

These were the operational improvements we expected to see in May. The first large installations establish reliability and operating records; subsequent projects should become easier for customers, financiers and regulators to accept. We called this acceptability as a moat.

In physical industries, trust can compound. A successful installation provides evidence for the next, while greater manufacturing volume supports automation, lower costs and better execution.

The question is whether that process eventually makes Bloom easier to finance as well as easier to manufacture.

So far, the evidence is stronger on manufacturing.

The Customer Base Is Broadening, but the Proof Is Incomplete

Our May thesis distinguished three kinds of progress: Oracle provided technical proof, Brookfield provided financing proof, and utilities could eventually provide distribution proof.

Q3 brought encouraging evidence of broader demand. Bloom’s largest contractual counterparty accounted for 43% of revenue, compared with 73% for the largest counterparty in Q2. Yet the top two still represented 69%, and revenue from Brookfield-related joint ventures returned to $322 million, approximately 26% of the quarter.

That matters because the contractual buyer is not necessarily the ultimate electricity consumer. A project-financing entity purchasing Bloom’s equipment is evidence of demand and financing capacity but not equivalent to a diversified collection of end customers independently underwriting projects.

Oracle remains another source of uncertainty. Its Jupiter campus has faced pipeline-related delays, and Bloom’s Oracle-related warrants create accounting adjustments as contractual conditions are met. The Q3 net warrant adjustment was only about $0.5 million, but that figure cannot reliably tell us how many Oracle systems shipped. It would be a mistake to treat an accounting adjustment as a delivery schedule.

The broader improvement in revenue concentration is encouraging. It does not establish complete independence from any particular project.

The more interesting customer development may be American Electric Power’s planned 702 MW Bloom installation with carbon capture to serve Google.

A utility purchasing fuel cells to meet demand is different from a data center buying equipment because the utility cannot supply it. The incumbent is adopting the alternative technology. If other utilities follow, Bloom could become part of routine power procurement rather than merely an emergency solution to grid delays.

That would be stronger evidence of institutional acceptance than another large announcement from a project-financing affiliate.

But utilities have alternatives. AEP also works with Mainspring, while Caterpillar and conventional generator suppliers are pursuing the same data-center opportunity. Bloom’s faster delivery, modular design and permitting characteristics may remain valuable, but scarcity alone cannot establish a permanent pricing advantage.

For now, Bloom is demonstrating that it can supply a fast-growing market. Whether successful installations translate into preferential financing and recurring orders remains open.

The Wait Moves Onto Bloom’s Balance Sheet

This is the quarter’s most important finding.

Bloom recognizes revenue when equipment meets the relevant delivery and acceptance requirements. Cash collection can occur later because contractual billing milestones do not necessarily coincide with revenue recognition.

That difference is manageable when projects proceed on schedule. It becomes more consequential when the factory accelerates faster than customers can commission, finance or permit their sites.

In Q3, operating cash flow was negative $27 million. After $153 million of capex, free cash flow was negative $181 million.

The balance sheet shows the accumulation:

Receivables plus contract assets increased from approximately 72 days of quarterly revenue at the end of 2025 to 111 days in September. That indicates more revenue is tied up in collection and billing milestones, although the ratio is affected by timing and revenue growth.

Management’s explanation deserves consideration. Bloom generated $395 million of free cash flow in Q4 2025 and $473 million in Q4 2024. Large installations and year-end collections can create substantial quarterly swings. Rising contract assets are not automatically bad if they convert predictably into invoices and cash.

CFO Simon Edwards made the distinction explicit:

“While contract asset balance might grow, what’s really important for us is that they’re turning because we’re reaching those billing milestones and we’re collecting cash.”

That is the claim we need to test.

The more consequential disclosure is the $5.8 billion of supplier commitments extending beyond twelve months, including $2.6 billion scheduled for 2027. Some arrangements contain minimum-purchase or take-or-pay provisions.

Those commitments are not debt, nor can they be compared directly with the $661 million of reported remaining performance obligations. The latter is an accounting disclosure that excludes certain customer contracts, not Bloom’s complete order book.

But the underlying asymmetry is real.

To deliver power in months, Bloom must order components, reserve capacity and build supplier relationships before every customer project is fully ready. If customers encounter permitting or financing delays, some commitments remain even while payments and deliveries move later.

Customer deposits provide an important offset, and strong demand may eventually absorb the inventory. Bloom is not necessarily lending customers money. It is carrying some of their scheduling uncertainty.

The customer pays Bloom to stop waiting for power. Bloom increasingly commits its own capital so the customer does not have to wait.

That is the financial cost of its competitive advantage.

Who Keeps the Value of Time?

Boulton’s achievement was not simply producing more efficient engines. He developed a commercial arrangement that allowed his company to capture a share of the savings customers received.

Bloom’s customers may generate extraordinary returns by powering GPUs months or years earlier. Bloom captures only the portion of that value it can negotiate through equipment pricing, service contracts and commercial terms.

Today, scarcity strengthens that negotiating position. Turbine manufacturers have long backlogs, AI electricity demand is accelerating, and the ability to deliver quickly is exceptionally valuable.

But supply is responding. Caterpillar, GE Vernova, Wärtsilä and others are expanding their offerings. Bloom’s long-term advantage cannot depend solely on rivals being unavailable.

The strongest bull case is that manufacturing experience and successful installations make Bloom increasingly trusted by utilities, regulators and infrastructure investors. If that reduces customer-acquisition costs, financing friction and deployment uncertainty, the company can retain attractive margins even after alternative capacity becomes more available.

The opposite outcome is also possible. Bloom could become an excellent industrial supplier that must continuously out-execute competitors, fund inventory and absorb scheduling risk to sustain growth.

Its installed-base economics have not yet resolved the distinction. Service revenue grew 15% year over year, against 179% product growth, while performance guarantees continue to consume cash. The recurring service business is improving, but it has not yet become a powerful independent source of economic value.

Our May framework described three possible futures: Newcomen, where Bloom supplies temporary bridge power; Toll Road, where it becomes a strong industrial supplier; and Boulton, where its technology becomes sufficiently trusted and economical to sustain premium returns.

Q3 moves our confidence toward Toll Road, with a credible path to Boulton. It does not yet prove the final outcome.

A Better Business at an Expensive Price

We have rebuilt the three-year valuation around revenue growth, normalized profitability, stock compensation, dilution and cash conversion. The figures below are illustrative end-2029 values, not present values or company guidance.

Our base-case revenue is close to the Street’s approximately $13.2 billion FY2029 estimate. The disagreement is principally about what that revenue becomes.

We model $1.68 billion of FCF in 2029, well below the roughly $2.7–2.8 billion consensus. We are not disputing the electricity demand opportunity. We are questioning how much cash Bloom will retain after funding growth.

At the October 22 pre-results close of $217.77, the base-case terminal value of $215 implies essentially no nominal appreciation over three years. At a 12% required return, that future value is worth only $153 today.

Even the base-case 34-times terminal earnings multiple assumes investors continue assigning Bloom a substantial premium after years of rapid expansion. At 25 times, the same earnings would support a share price of approximately $158 in 2029.

The probability-weighted terminal value is around $252, worth approximately $179 today using the same discount rate. The upside is real, but the expected return is insufficient at the reference price.

The Next Test Is Cash, Not Another Contract

October 22’s broader AI-related selloff began before Bloom reported. The stock fell alongside other companies exposed to AI infrastructure financing. That price action cannot be attributed entirely to Bloom’s quarterly results, but the concern has a direct counterpart on its balance sheet.

The next two quarters should provide a clearer answer.

At September’s end, receivables plus contract assets represented approximately 111 days of quarterly revenue. Revenue growth alone would lower that ratio mechanically, so we need more than a smaller headline number.

Below 80 days at year-end, accompanied by meaningful invoice conversion and cash collections, would support management’s timing explanation. Remaining above 110 days would indicate that the burden persists despite higher sales. Between those levels, the underlying movement in contract assets and actual collections matters more than a binary verdict.

We also want 2027 guidance showing that customer-backed deployments can support the $2.6 billion of scheduled supplier commitments, sustained product margins near or above 37%, and more evidence of utility adoption and independently financed customers.

Bloom has already demonstrated something important: it can manufacture and deliver power at a speed increasingly valuable to AI customers. The original technology thesis is stronger than it was in May.

But the investment thesis requires one more step. As the installed base grows and institutional acceptance improves, Bloom should need to carry less scheduling and financing risk relative to the revenue it generates. If that happens, acceptability really can become a moat. If it does not, the business may continue growing impressively without delivering the returns shareholders expect.

Boulton got paid for the coal his customers saved. Bloom must now demonstrate how much of the time it saves customers it can keep, rather than pay for itself.

Subscribe now

• Spotify’s customer relationship looks increasingly durable: Premium subscribers beat expectations, personalization remains useful, and external agents may become another distribution layer rather than a substitute for Spotify itself.

• But monetization is still the unresolved part of the thesis: Premium gross profit per subscriber-month is up meaningfully year over year, yet it was flat sequentially, while add-on products such as Audiobooks+ still lack enough disclosure to prove they materially lift retained economics.

• That distinction matters at the current valuation: the stock rerated sharply while 2027 earnings expectations barely moved. The franchise may be stronger than before, but at roughly $584 the upside depends on proving Spotify can keep more of each incremental euro after royalties, investment and dilution.

Spotify’s third quarter was close to expectations. Revenue of €4.97 billion was slightly below consensus, operating income of €672 million essentially matched it, and free cash flow declined year over year. Yet the shares rose 11% to $583.66. The market seemed relieved that a year of elevated investment was ending and reassured that AI agents might extend Spotify’s reach rather than replace it.

Both conclusions have merit. Neither settles the argument we made in August’s The Moat Learns to Monetise. Spotify has spent years learning what each listener wants, turning that knowledge into habit, and improving the economics of a subscription constrained by music royalties. The next step is to charge its most engaged customers for more than basic access: additional audiobook hours, differentiated experiences, and products whose price rises with use. We called that the Meter.

Q3 strengthened the case that Spotify can keep the listener. It offered little new proof that the Meter can materially increase what Spotify earns from that listener. And whatever new value it creates will still be shared with rights holders. The customer relationship looks increasingly durable. The incremental economics are still negotiated.

The Ceiling Was Broken, Not Abolished

Spotify’s gross margin reached 33.0%, up from 31.6% a year earlier. For much of its history, the company was judged against an apparent mid-20s ceiling: as subscription revenue rose, music royalties rose with it. That ceiling has broken in the consolidated numbers. But at its May Investor Day, management said only about a third of margin expansion since 2022 came from music. Podcasts and audiobooks supplied most of the rest. The music business improved; Spotify also changed the mix of what it sells.

The core subscription remains healthy. Spotify added six million Premium subscribers, beating its target, while adding 11 million monthly active users. Fewer free-user additions alongside solid paid growth are consistent with August’s expectation that Spotify would prioritize more valuable users. They do not, on their own, prove better conversion: Spotify does not disclose the cohorts or churn required to establish that.

Operating margin was 13.5%, essentially flat with a year earlier, because marketing and investment in AI and cloud services absorbed the gross-margin improvement. Share-price-linked payroll taxes depressed the comparison: excluding a €22 million charge this quarter and a €16 million credit a year ago, operating income grew roughly 23%, rather than the reported 15%.

The Q4 guide is encouraging. Revenue of €5.2 billion, gross margin of 33.8% and operating income of €820 million imply operating expenses of roughly €938 million, down from €967 million in Q3. CFO Christian Luiga said expense growth was now “coming down meaningfully” and expected that moderation to continue into 2027. The year of accelerating investment may be ending, but that does not mean the elevated cost base disappears. Spotify must grow into it.

One accounting wrinkle matters when the next quarter arrives: the guide assumes a September share price of $487.38 for payroll-related charges, well below the post-results price. A sustained rally could push reported expense above the guide without changing the underlying business. We should judge expense moderation after adjusting for that mechanical effect, not on a headline miss alone.

The Agent Is a Door, Not the Destination

The strategic highlight came from co-CEO Gustav Söderström. As people move from opening an app to asking an assistant for “something to run to,” the assistant could become the interface that chooses the music service. That is an obvious threat to Spotify, particularly when Apple, Google and Amazon own both assistants and competing music subscriptions.

Söderström described something different:

“Listening is actually not shifting. All the listening is happening through Spotify ... [The agent] remote-controls Spotify in the background.”

That is Spotify Connect for the agent era. Connected speakers expanded Spotify’s reach without becoming its replacement; an external assistant can work the same way. Spotify was early to ChatGPT and, according to management, was Muse’s only music service at launch. Meta and OpenAI do not have competing music subscriptions, making an independent partner attractive.

This is a distribution advantage, not an exclusive right. The assistant may still decide which service to call; Apple Music or YouTube Music could become equally accessible. And while a general model knows which songs exist, Spotify knows years of individual listening choices, the reason Söderström has argued that “taste is not a fact.” Those preferences are a source of convenience and switching friction, though rival services have listening data too.

Nearly 40 million Premium users now listen to mixed playlists, with consumption rising 90% sequentially. That supports the usefulness of Spotify’s personalization. It does not tell us whether outside agents generate incremental listening, paid conversion or profit. Spotify disclosed none of those numbers. The agent may open a new door to customers, but we have not yet seen what is coming through it.

The Meter Has to Survive the Labels

A new door matters only if Spotify earns more from the people who enter. That is where the quarter was less reassuring. Constant-currency Premium revenue per user grew 7.5% in Q3, but management expects only 4.5%–5% in Q4. Part of the slowdown reflects lapping price increases across more than 150 markets, not necessarily weakening pricing power. Still, broad subscription increases arrive in rounds. Between rounds, Spotify needs more subscribers, richer plan mix or additional purchases to sustain growth.

That is the purpose of the Meter. Audiobooks+ had already passed $100 million of annual recurring revenue in August. Spotify now plans to expand audiobooks from 22 to more than 180 markets and has launched Creator Memberships. Usage-linked products could let heavy listeners pay for what they consume, while matching extra revenue to the added cost of delivering it. But Q3 provided no new Audiobooks+ revenue, attach rate or product-level margin. The ability to charge is growing faster than the evidence that customers are paying materially more.

Our most useful measure is therefore not the number of new features, but Premium gross profit per subscriber per month, average Premium revenue per user multiplied by Premium gross margin.

Spotify now keeps about 14% more gross profit per paid listener than it did a year ago. That is substantial progress. But the number did not improve sequentially: higher ARPU was offset by a lower Premium margin. This is the economic gap between better personalization and better monetization.

The labels determine part of what fills it. Universal, Sony, Warner and Merlin accounted for about 72% of label-delivered streams in 2025. Royalties are negotiated, often with revenue-sharing features, so not every extra euro reaches Spotify’s gross profit. Universal and Warner have also supported differentiated tiers and bundles in recent agreements; richer products could expand the pool for everyone rather than merely redistribute it. The continuing dispute over how audiobook bundles affect publishing royalties shows why the definition of a new product matters as much as its popularity.

By Q3 2027, we want to see gross profit per Premium subscriber approaching €1.80 a month with stable retention. A fall toward €1.65, particularly alongside adverse royalty terms, would be far more concerning than a disappointing feature-launch statistic. Advertising provides no offset yet: revenue per quarter-end free user was roughly €0.97, against €0.98 a year earlier, although that is an imperfect yield measure because free users are also a recruitment funnel.

A Stronger Franchise Is Not Necessarily a Better Stock

The share-price reaction highlights the difference. Analysts raised their targets after results while leaving 2027 earnings expectations almost unchanged. They became more willing to pay for essentially the same profit. Consensus also expects roughly 14.8% revenue growth in 2027, which is demanding against the near-term ARPU guide without another broad pricing round. We are closer to 13.5% unless those changes.

Our own August base-case range of $640–750 also needs revisiting: the company will carry a higher investment base than we assumed, and our earlier cash-flow approach treated stock compensation too generously.

Rather than compound that error, we value after-tax earnings, explicitly model diluted shares, and use free cash flow as a separate reality check. The figures below are end-2029 values, not present values. They assume €1 = $1.12, different tax rates and modest net finance income across the cases; cash is not added to the values a second time.

The base assumes continued subscriber growth of about 7%–8%, core revenue per user up roughly 4%–5%, and a few hundred million euros of additional annual metered revenue by 2029. It is not a heroic success case. Yet it offers only about 6% cumulative upside from $583.66. The bear implies a 42% loss; the bull, a 55% gain.

At these probabilities, the weighted end-2029 value is $620. That is around 6% total appreciation, or 2% a year across roughly 3.2 years. To earn 10% annually on the same assumptions, the entry price would need to be approximately $460; for 15%, about $400. These are mathematical thresholds, not promises that the scenarios or probabilities are right.

The base case also requires scrutiny beyond earnings. Its roughly €4.9 billion of indicative 2029 reported free cash flow assumes stronger cash conversion than after-tax operating profit; that is not automatically owner earnings, particularly with employee equity awards and buybacks. If the Meter remains small or royalty economics deteriorate, a 26-times exit multiple could prove generous.

We would hold rather than add near $584 and require either a meaningfully lower price or new evidence of higher retained profit before increasing exposure. A stock that rises faster than the underwriting improves is not made safer by a better quarter.

What Changes the Decision

February’s Q4 results should show whether expense growth really moderates, adjusting for stock-linked payroll taxes, and whether Premium revenue per user holds near the guided 4.5%–5% growth. During 2027, the decisive evidence will be gross profit per Premium subscriber, the size and margins of add-on products, and the outcome of label negotiations. An initial number for agent-originated listening and conversion would help distinguish a genuine new distribution channel from an attractive integration announcement.

Spotify has already answered the old question of whether music streaming can be a durable, profitable subscription business. Q3 made the answer more convincing. The open question is whether deeper knowledge of listeners and more places to reach them allow Spotify to keep substantially more money per customer.

The listener is loyal. The next leg of profit is still being negotiated.

Subscribe now

• The structural thesis is working: semiconductor manufacturing is getting more complex, Lam’s served market is expanding, Aether dry resist is gaining traction, and backside-power wins are adding new process positions. The physics is creating more places for Lam to participate. Pasted text

• But the pricing-power thesis needs to be revised: ASML is showing long-dated customer commitments and negotiated price increases, while Lam’s evidence points more toward process wins, higher content and switching costs than true scarcity pricing. Pasted text

• The investment case is therefore more cyclical than we previously assumed: ex-China growth is strong, but China supplied most of the latest sequential revenue increase, support revenue was nearly flat, and the current valuation still requires Lam to prove that more manufacturing steps translate into durable shareholder economics. Pasted text

In August 2013, Samsung began mass production of a flash memory chip constructed from 24 layers stacked vertically. For decades, the industry had made memory denser by shrinking it, requiring increasingly precise and expensive lithography. Stacking changed the manufacturing problem. Each layer needed to be deposited, and holes had to be etched through the resulting structure, deeper and narrower with every generation.

The difficult work was no longer only drawing the pattern. It was building the structure.

Lam Research was particularly well positioned for that transition. Its expertise in etching and deposition became increasingly important to the construction of three-dimensional memory, while every successful equipment installation established a position that could generate years of upgrades, spares and services.

Thirteen years later, the same architectural shift is spreading across DRAM, logic and advanced packaging. Lam has just reported the strongest quarter in its history: revenue of $8.22 billion, up 54% year over year, non-GAAP operating margin above 40% for the first time, and free cash flow of $3.24 billion. Management now expects calendar 2027 wafer fabrication equipment spending to exceed $200 billion.

The shares nevertheless fell after the announcement, having already declined sharply from their June peak of $438.50.

On July 30, The Physics Arrived Early ended with a distinction:

This quarter proved Lam captures more when customers spend. The next cycle will show how much persists when they stop.

I think the market has started pricing the second sentence. But this quarter reveals a more fundamental distinction that our earlier articles did not adequately recognize.

Lam is becoming more technologically important. That does not necessarily mean it is acquiring proportionately greater pricing power.

The Numbers Passed

Our July article established six signposts. September cleared most of the near-term tests.

The financial progress is undeniable. Margins reached records, receivables collection improved, and free cash flow rose from $1.27 billion in June to $3.24 billion in September.

But the historical signposts are not the same as the current earnings bar.

Our two independent October 19 preparations both expected more. Claude forecast $8.30 billion in September revenue and a $9.10 billion December guide, with an inferred buy-side guide threshold of $9.0–9.3 billion. ChatGPT’s frozen ledger forecast September revenue of $8.25–8.40 billion and a December guide of $8.70–8.90 billion.

Actual revenue was $8.22 billion, and December guidance was $8.60 billion. Both forecasts were too optimistic, especially about the next quarter.

The July thesis passed several historical tests while our newer expectations failed. That is not a contradiction. It means the business is progressing, but the financial upside we expected was already harder to deliver.

Our earlier July estimate of approximately $8.20 in FY2027 EPS also needs care: it reflected a pre-June-results estimate that was already being revised upward around publication. Consensus near $9.54 by October is not entirely a post publication forecast victory.

More importantly, two of our earlier conclusions about what kind of business Lam is have become less defensible. Those conclusions determine the valuation multiple.

The Hard Part Moved. The Pricing Power Did Not Move Equally.

The technological argument has become stronger.

On the October call, UBS’s Tim Arcuri asked Tim Archer about Micron’s decision to prioritize three-dimensional DRAM over a proposed 4F² architecture. Lam’s shares had reacted negatively to the roadmap change, yet Archer argued that a more direct transition into 3D DRAM could accelerate the need for etching and deposition.

That is the original thesis arriving in another category.

Lam says its served available market has expanded by approximately four percentage points as a share of WFE. Foundry/logic and DRAM contributed about 80% of systems-revenue growth during calendar 2026 to date, reflecting product investments begun several years earlier.

The company also increased its expected cumulative Aether dry-resist revenue to approximately $2.5 billion by the end of the decade, roughly $1 billion above its previous outlook. It reported additional backside-power process wins, including production tool-of-record positions for several technologies at a leading customer.

This is persuasive evidence that semiconductor complexity is creating new commercial opportunities for Lam.

The question is who captures the value.

In January, The Atomic Toll Booth argued that the industry’s manufacturing bottleneck was moving from lithography toward etching and deposition, and that Lam would increasingly capture the economics previously associated with ASML’s scarcity.

October provides a useful test of that claim because ASML and Lam reported within a week of each other.

ASML’s non-current contract liabilities rose by approximately €5.73 billion during its September quarter, reaching €9.30 billion. These balances include customer commitments and advances associated with future deliveries, although the increase should not be interpreted as €5.73 billion of cash newly received during the quarter.

ASML also completed negotiations that increased prices for its most critical products. The company expects some financial benefit in 2027, with the full effect beginning in 2028.

These two observations reinforce one another: customers are making substantial long-dated commitments to secure ASML equipment, and ASML has been able to negotiate higher prices.

Lam’s customer behavior looks different.

Its deferred revenue increased by $227 million to $2.66 billion, but CFO Doug Bettinger explained that stronger sales volumes were partly offset by declining customer down payments.

That does not mean Lam lacks pricing power. Its gross margin has risen from approximately 48% two years ago to more than 52%, and management has identified pricing initiatives alongside product differentiation and manufacturing efficiency.

Nor are ASML’s contract liabilities directly comparable with Lam’s customer deposits. Equipment lead times, contract structures, delivery schedules and payment terms differ.

Nevertheless, the contrast is informative.

ASML is the sole commercial supplier of EUV lithography systems. Lam holds valuable positions in qualified etching and deposition processes, but it competes with Tokyo Electron, Applied Materials and other suppliers across a wider range of applications.

Once Lam wins a process position, customers face meaningful switching costs. The difficulty is that the next technology generation can reopen competition for the next position.

There is another complication our previous articles missed.

Lam says its Akara conductor-etch technology can deliver up to 50% more wafer output per fab footprint.

That is valuable to customers. It also means that better technology can reduce the number of chambers needed for a given amount of production.

More process steps do not automatically translate into proportionately more equipment revenue. Lam must capture the benefit through higher prices, additional process wins, greater market share or increased lifetime service economics.

The same technological improvement can create substantial value for a chipmaker without allowing Lam to retain all of it.

This is the central correction to our original thesis.

We were right that semiconductor complexity would increase the importance of the physical construction processes Lam specializes in. We were too quick to assume that increasing technical importance would translate directly into ASML-like bargaining power.

Lam’s toll is primarily paid through a growing number of economically valuable manufacturing applications. ASML has demonstrated a stronger ability to charge for the scarcity of its machines.

From Tokens to Wafers

The most interesting new business argument in Lam’s October call concerned NAND.

Archer traced AI demand beyond GPUs and HBM into the storage hierarchy. As inference workloads become more complex, the demand for storage increasingly separates into different performance requirements.

QLC NAND is suited to high-density, lower-cost storage. TLC can serve higher-performance applications, including context memory and key-value caches. SLC-class storage offers lower latency and faster access for workloads that prioritize responsiveness over density.

The economic consequence is potentially significant.

Moving data from TLC to SLC-class devices reduces the number of bits stored per device. Lam estimates that each percentage-point shift of NAND bit demand toward SLC-class storage could require approximately 2% more installed wafer capacity to produce the same number of bits.

This is a different growth mechanism from the NAND layer-count conversion cycle that dominated our earlier articles.

Conversions require customers to replace or upgrade equipment to manufacture more advanced devices. A lasting shift toward less bit-dense, higher-performance storage could create a need for additional manufacturing capacity even without proportionate growth in total stored bits.

That potentially answers the question we have carried since April: what replaces the NAND conversion spending once the current upgrade cycle finishes in 2027?

But it remains a hypothesis.

When Mizuho asked how much NAND demand might migrate toward SLC-class devices, Archer declined to provide an adoption forecast. He also did not quantify the resulting equipment opportunity.

Additional wafer requirements need not produce an immediate wave of new factories. Customers may first use available capacity, reallocate production or expand existing facilities.

Until memory manufacturers commit meaningful investment specifically to this architecture, the SLC opportunity remains an emerging source of optionality rather than a reliable component of our base-case forecast.

It strengthens the long-term explanation for why NAND could become more important to AI. It does not yet settle the investment case for Lam.

What Compounds, and What Cycles

Our July article distinguished between increasing manufacturing complexity and the cyclical spending through which that complexity becomes revenue.

September provided evidence of both.

Ex-China revenue has risen substantially, reaching approximately $5.34 billion in September, roughly 76% above the prior year. This is encouraging evidence that Lam’s growth is not dependent entirely on Chinese spending.

Yet China supplied approximately three-quarters of September’s sequential revenue increase. Domestic Chinese customers, together with multinational customers operating there, contributed to the unusually strong quarter.

Management expects the Chinese contribution to decline meaningfully in December, particularly from domestic customers.

At the midpoint of Lam’s $8.60 billion guide, a China revenue share of 25–30% would require ex-China revenue of approximately $6.0–6.5 billion—an increase of roughly 13–21% in a single quarter.

Those percentages are illustrative assumptions, not management guidance. But they establish the scale of the replacement demand required.

They also force a correction to July.

When gross margin reached 52% with China down to 26% of revenue, I wrote that the thesis of China-related margin compression had been falsified.

That was too categorical.

We demonstrated that Lam could deliver attractive margins with one particular customer and geographic mix. We did not prove that future China declines could never affect margins.

September’s return to 35% China exposure reinforces the need to distinguish operating efficiency, product mix, pricing and geography rather than treating a favorable result as permanent evidence.

The installed-base business raises a similar question.

The support business has grown meaningfully in absolute dollars, but it has not become a steadily increasing share of revenue.

In September, systems revenue grew almost 35% sequentially, while support-related revenue grew less than 1%.

Management expects CSBG growth exceeding 15% in calendar 2027, following more than 30% in 2026. That is impressive, but the components have different economics.

Spares and maintenance benefit from a larger installed base and high utilization. Upgrades depend partly on technology transitions and customer spending. Reliant equipment remains exposed to mature-node investment, including China.

When JPMorgan asked management to rank the drivers, Bettinger emphasized the continuing strength of spares, the NAND upgrade cycle and Reliant. He did not disclose separate profitability or growth rates for those businesses.

That leaves an important question unanswered: how much of the reported support business is a durable annuity, and how much is another expression of the equipment cycle?

This is not an argument against the installed-base franchise. It is an argument for measuring its resilience rather than assuming it from the revenue category.

What the Physics Is Worth

My variant perception has changed.

I still believe Lam’s underlying engine is stronger than a conventional cyclical-equipment model suggests. The company is gaining exposure to increasingly complex semiconductor structures, its margins have improved materially, and the installed base creates valuable future opportunities.

But the January thesis went further. It argued that Lam’s position justified a structural re-rating toward the valuation of an indispensable infrastructure supplier.

The evidence for that conclusion is weaker.

ASML has demonstrated more pronounced supplier scarcity through future customer commitments and negotiated equipment pricing. Lam’s advantage is substantial, but it depends on repeatedly winning applications against credible competitors and turning those wins into profitable shipments.

Meanwhile, the industry outlook is no longer an undiscovered positive.

Before Lam’s results, Barclays had already estimated 2027 WFE spending at approximately $209.5 billion, and Bernstein had forecast $204 billion. Management’s preliminary outlook above $200 billion confirmed a large investment cycle, but it was not a new level of industry demand that investors had never considered.

At the October 21 regular-session closing price of $260.81, Lam traded at approximately 25 times contemporaneous next-twelve-month earnings expectations. This is a lower valuation than earlier in the year, but one that still assumes considerable earnings durability.

I would now use a more restrained terminal multiple, near 22 times FY2029 earnings in the central scenario, rather than the 28–32 times used in the January analysis.

The probability-weighted value is approximately $294, or 13% above the October 21 close before dividends.

That is not an especially compelling expected return for an equipment company exposed to capital-spending cyclicality and a potentially hostile valuation regime.

An interesting comparison is ASML. On FY2028 consensus, it trades near 20 times earnings versus approximately 22 times for Lam. But ASML’s fiscal year ends six months later, and a matched next-twelve-month comparison makes ASML the more expensive stock—approximately 28 times versus Lam’s 25 times.

ASML therefore is not obviously cheaper on a consistent earnings horizon. The case for prioritizing it is that its scarcer competitive position may deserve the premium.

For Lam, the conclusion is more modest: I would retain existing exposure, but would not prioritize additional semiconductor capital here simply because the industry is entering another period of exceptional spending.

The January Test

The next earnings report offers a direct test of our structural argument.

If December ex-China revenue comes in below $6.0 billion, Lam will have failed to sustain even approximately 12% sequential ex-China growth despite an expanding industry and extraordinary technology demand.

That would not by itself disprove the long-term process-intensity thesis; supply constraints and shipment timing could explain some of the shortfall. But it would weaken the argument that the underlying non-China engine is accelerating quickly enough to offset the current concentration in Chinese spending.

The same call should tell us whether December shipments delayed by supply constraints reached customers in March-quarter guidance, whether support revenue moves convincingly beyond $2.5 billion, and whether management’s more specific 2027 WFE outlook validates the implied equipment capacity requirements.

Beyond that, the decisive test comes when the existing NAND conversion cycle ends and the industry’s current wave of fab construction approaches completion.

If Lam continues growing its economic content per wafer, maintaining margins and generating installed-base cash flow through that period, the case for structural compounding becomes much stronger.

The physics is doing much of what our series anticipated. Semiconductor manufacturing is becoming more complex, and Lam’s products are increasingly relevant to the technologies making that complexity possible.

What the results have not established is that technical indispensability translates into the same economic leverage ASML possesses.

The distinction matters because it determines what we should pay for the business.

The toll is real, and the number of steps is rising. The question now is how much of the toll Lam gets to keep.

Subscribe now