• The core thesis is being validated: Constellation signed roughly 920 MW of long-term nuclear agreements averaging 18.5 years, raised 2026 EPS guidance, and progressed Crane and the Calpine integration. The “boring” contracting thesis is working.

• Regulation looks more like friction than a veto: the new PPAs were signed before PJM and FERC rules were fully resolved, suggesting scarce, already-connected generation can be monetized even without perfect regulatory clarity.

• The rerating case is increasingly about earnings quality, not estimate upside: the opportunity is for more of Constellation’s already-expected earnings to become long-duration, contracted and defensible not necessarily for 2029 EPS to dramatically beat consensus.

“I recognize that the last time we spoke, I indicated that we expected to be done with an important transaction by this call, but we’re not ready to announce anything today.”

That was Joe Dominguez in March. Constellation had just laid out 20% base-EPS growth through 2029, authorized $5 billion of repurchases, and presented a financial framework designed to make the company look less like a merchant generator and more like a business with increasingly visible earnings. The stock fell 9.3% because none of that was what investors had shown up to hear. They wanted a hyperscaler deal.

Four months later, Constellation delivered one. Or rather, several. The company signed approximately 920 MW of long-term nuclear agreements averaging 18.5 years, raised 2026 adjusted EPS guidance to $11.50–$12.50, cleared two important regulatory hurdles for Crane, and reached agreement on the final Calpine disposal required by the DOJ. Dominguez’s summary of the fears from March was direct: “Those concerns have not materialized.”

And yet I don’t think Q2 changed the thesis. What it did was more useful: it reduced the probability that the thesis we already had would fail.

The Big Fundamental Question remains the same:

Can Constellation convert the growing scarcity of reliable, already-connected power into durable contracted earnings faster than new supply and regulation erode that scarcity?

Q2 moved the answer toward yes.

What We Said, What Happened

Our view of Constellation has evolved less than the language around it suggests. In August 2025, we argued that Constellation sold pressure rather than power. Nuclear was the reservoir: dependable, carbon-free generation available whenever customers needed it. The PTC was the levee protecting economics when power prices weakened; capacity markets rewarded readiness; long-term contracts were the pipes carrying that dependable output to customers.

We then corrected ourselves. We were right about the reservoir and early on the pipes. Co-location proved slower and more regulated than expected, while Calpine widened the ways Constellation could reach customers through gas, storage, retail relationships and powered sites. After Q1, we sharpened the customer problem again: increasingly, the buyer was not simply purchasing clean electricity but a date, power delivered when a data centre or industrial project needed it.

Q2 does not invalidate any of this. It does make one distinction clearer. The core thesis, that scarce operating generation can be converted into increasingly durable earnings, is working. The more ambitious extension, that Calpine lets Constellation repeatedly sell Freestone-like packages combining site, connection, generation and delivery schedule, remains unproved.

That matters because the 920 MW signed this quarter were not spectacular transactions. They were long-term nuclear PPAs. The boring version of the thesis worked.

The Pipe Moved Before the Permit

This is the most important fact from Q2.

In March, management blamed regulatory uncertainty for slowing contract execution. Customers did not know what PJM would require of large loads or how future rules would affect procurement. The obvious bear chain was straightforward: no clear rules meant no customer certainty; no certainty meant no contracts; no contracts meant scarcity remained stranded.

Q2 broke that chain.

When Jefferies’ Julien Dumoulin-Smith pressed management on whether the signed deals depended on the Reliability Backstop Procurement or future FERC decisions, management said they did not. These are executed contracts, not placeholders waiting for regulation to catch up.

PJM still matters. Management expects the RBP process to produce results by year-end and broader co-location clarity around the first or second quarter of 2027. But regulation increasingly looks like a constraint on pace and structure rather than a veto on monetizing the fleet.

The deeper reason is the time asymmetry underneath the entire power story. Large customers can commit capital and add load much faster than new dependable generation, and transmission can be permitted, interconnected and built. Existing, grid-connected generation is therefore the scarce asset during that gap, and Constellation owns 55 GW of it; management estimates that rebuilding the fleet today would cost more than three times the company’s current enterprise value.

Dominguez explained the same point in grid language:

“We have plenty of unused capacity in generation and in the wires grid over 99% of the hours of the year. We have a peak capacity concern, not an energy concern.”

His argument is that batteries, demand response, backup generation and curtailment can handle the relatively small number of difficult peak hours while existing generation serves incremental load during the rest.

That mechanism does not require a landmark co-location agreement. Capacity markets already pay more for scarcity. Bilateral PPAs already monetize long-term nuclear value. Commercial margins capture volatility. Buybacks turn cash into per-share earnings. The routing layer could make the flywheel turn faster, but Q2 showed that the flywheel does not depend on it.

Optionality Gets Attention; Base Gets a Multiple

Our Q1 article contained the line that still best explains the financial mechanism: “Optionality gets attention; base earnings get a multiple.”

An uncontracted nuclear MWh is an option on tighter power markets, rising capacity values and customer urgency. Sign an 18-year agreement and Constellation gives away some of that open-ended upside. In return, it gets duration.

The new contracts average 18.5 years and take roughly 30% of expected clean baseload generation under long-term agreement by 2032. Yet Shane Smith made clear that they start late enough that they do not materially affect 2029 earnings.

That sounds disappointing until you look at what is already in the model. Bloomberg expects adjusted EPS of roughly $11.75 in 2026, $13.06 in 2027, $16.61 in 2028 and $19.10 in 2029, despite very little revenue growth after Calpine enters the base. The Street is already assuming earnings conversion.

What the new contracts do is improve the durability of earnings beyond the period that dominates today’s forecast.

The same idea appears in the Illinois transition. Constellation models 53 million MWh of Illinois CMC nuclear output in 2026 at $34.09/MWh, falling to 23 million MWh in 2027, while nuclear volumes supported in the PTC bucket rise from 101 million to 127 million MWh and the assumed floor increases from $44.75 to $45.75/MWh. A casual reading sees a subsidy ending. The economic reading sees a large block of output migrating from one support mechanism into another.

That is what management’s base-earnings architecture is trying to show.

The obvious bull case is no longer differentiated. Nuclear scarcity is well understood. AI electricity demand is well understood. Twenty-one analysts rate CEG a Buy and none rate it a Sell.

I do not think the interesting bet is that $19.10 of 2029 EPS is wildly too low. The interesting bet is that the character of those earnings changes faster than the multiple anticipates. If more of the $19 becomes long-term contracted, inflation-linked and backed by assets that cannot be quickly replicated, then CEG does not need a dramatic estimate beat to work. It may simply deserve a better multiple on the earnings already expected.

The risk is the mirror image. Constellation could deliver the EPS and still fail to re-rate if enhanced earnings remain large, PPA pricing sits near the low end of management’s $20–$50/MWh premium range, and the broader customer-solutions business never becomes repeatable.

Steve Fleishman pushed Shane Smith directly on the pricing range. Smith refused to narrow it. That leaves an important gap in the evidence. Volume is disclosed. Duration is disclosed. Counterparty quality is disclosed. The economic premium is not.

What Management Is Really Selling

The prepared remarks and Q&A suggest management is trying to make three arguments at once.

The first is that the existing fleet, not speculative new build, is the asset. When KeyBanc’s Sophie Karp asked about new nuclear, Dominguez said conversations in New York were interesting but “not that imminent.” Management wants investors focused on the value of megawatts that already exist.

The second is that regulatory clarity will increase deal velocity rather than create demand from scratch. When JPMorgan’s Jeremy Tonet asked how customer conversations had changed, Dominguez said ambiguity had been the enemy of deal execution and predicted that contracting could “kick off with a bit of a bang” once rules settle. That is a strong claim, and it now has a timetable.

The third is that Calpine should be judged as more than $2 of EPS accretion. Calpine contributed $1.028 billion of Q2 RNF, $2.147 billion of operating revenue less $1.119 billion of purchased power and fuel, while Brazos Valley is being sold for about $1,420/kW. Required disposals should generate roughly $5.9 billion at close to $1,200/kW against approximately $960/kW implied by the acquisition.

That supports the asset-value argument. It does not yet prove the full time-to-power thesis. We said after Q1 that if Freestone remained a one-off, Constellation could still own an excellent clean-firm contracting business while the broader opportunity turned out to be less valuable than we imagined.

That test remains open.

A Good Quarter With Messy Accounting

Adjusted EPS of $2.55 beat Bloomberg’s $2.32 estimate, and management raised full-year guidance by $0.50 to $11.50–$12.50. But adjusted EPS was also $1.13 above GAAP EPS of $1.42.

The bridge is substantial. Unrealized fair-value adjustments added $0.94 per share to the reconciliation, acquired commodity-contract amortization added $0.41, Calpine merger and integration costs added $0.23, pension and OPEB added $0.06 and legal/environmental items added $0.10, partly offset by $0.61 of decommissioning-related activity.

I think most of those adjustments are understandable for a power business absorbing a large acquisition. Still, an improving earnings-quality thesis eventually needs to appear in cash. H1 operating cash flow was $1.55 billion versus $1.58 billion last year, while capex rose to $2.52 billion as Calpine, Crane and other projects absorbed capital.

Bloomberg expects free cash flow to reach about $6.1 billion by 2029 and net debt to fall below $14 billion. If that does not happen, the multiple argument weakens considerably.

Three Years From Now

For an August 2029 value, I care more about earnings composition than revenue growth.

The base case is not heroic. Bloomberg already has $19.10 of EPS in 2029. It requires continued contracting, Crane broadly on schedule, nuclear reliability near historical levels, higher free cash flow and a growing base share of earnings.

The bull case needs more: several gigawatts of premium contracting, stronger gas utilization, at least one repeatable Freestone-like model and continued accretive capital allocation. The bear is not “AI disappears.” It is that supply and regulation catch up before Constellation converts enough scarcity into contracts, while enhanced earnings normalize and the market returns the stock toward conventional power-company valuation.

The Scoreboard

The next year should tell us whether Q2 was the beginning of a cadence or merely a strong quarter. Cumulative new long-term nuclear contracting above 2 GW by the Q4 2026 earnings call would signal acceleration; below 1.5 GW would suggest that Q2 did not change velocity. Contract pricing remains harder to observe, but evidence that deals are clearing comfortably inside the $20–$50/MWh premium range would strengthen the rerating case; repeated deals near the low end would weaken it.

PJM needs to deliver the expected RBP outcome by year-end and broader co-location clarity by 1H27. Meaningful slippage beyond that would restore some of the regulatory discount. Nuclear capacity factor should remain around 93–94% or better outside planned-outage noise; sustained readings below 92% would raise a different and more serious question about the asset itself.

Finally, the accounting has to catch up with the story. The absolute GAAP-to-adjusted gap should shrink as Calpine purchase-accounting and integration effects fade, while free cash flow should begin moving toward the $6 billion level Bloomberg expects late in the decade. And if Constellation signs another material Freestone-like agreement by the end of 2027, the broader time-to-power thesis would finally begin to move from possibility to repetition.

The Same Bet, Better Evidence

For a year, we kept asking when the routing layer would activate. Q2 suggests that was too narrow a question.

The structural advantage is the time gap between demand and new supply. Capacity markets monetize it one way. Long-term PPAs monetize it another. Commercial execution captures volatility. The PTC raises the floor. Buybacks turn cash into per-share growth. None of those mechanisms requires the spectacular transaction investors have spent eighteen months waiting for.

The routing layer could make the machine better. It could give Constellation more ways to solve a customer’s entire power problem rather than simply sell electricity. But Q2 showed that the core engine can already compound without it.

The thesis did not change.

The evidence did.

Subscribe now

• The quarter was exceptional, but the thesis has moved: Sandisk reported an 84.6% gross margin and $39.25 of adjusted EPS, yet the more important development was $93.9 billion of minimum contracted revenue at floor prices.

• Sandisk is trading upside for duration: more than half of FY2027 bits and roughly two-thirds of FY2028 bits are already committed, creating a contracted core with multi-year agreements, price protection and financial guarantees.

• The market may be using the wrong trough: current margins are not sustainable, but the old boom-and-bust framework may no longer apply. The investment case now depends less on further estimate upgrades and more on how much normalized earnings persist through the next downturn.

“We want to increase the visibility and the durability of the franchise. We want to get this kind of boom and bust out of it.”

David Goeckeler gave that answer when Cantor’s CJ Muse asked why Sandisk’s next-quarter gross-margin guide was not higher even though NAND pricing was still rising. The wording matters. A chief executive whose company had just reported an 84.6% gross margin described the mid-80s as a fair return, then immediately changed the subject from price to duration. Management is no longer trying only to extract the highest possible margin from the shortage. It is trying to make an extraordinary margin last.

The quarter was certainly extraordinary. Revenue reached $8.97 billion, gross margin was 84.6%, adjusted EPS was $39.25, and adjusted free cash flow was $5.04 billion. Datacenter revenue doubled sequentially, while Datacenter’s share of bits rose from 12% a year ago to 38%. Sandisk guided the next quarter to $10.3–10.8 billion of revenue, an 83–85% gross margin, and $44–46 of adjusted EPS. Yet the most important disclosure was not in the income statement. It was that Sandisk now has $93.9 billion of minimum contracted revenue at floor prices.

The question is no longer how high earnings can go. It is how much of them stay.

The Thesis Has Moved

I have written three pieces on Sandisk. The first argued that AI created a missing storage layer between HBM and hard drives. The second connected that layer to context memory and KV cache inside new AI systems. The third accepted that the demand thesis had been proved and shifted the debate to contracts: could Sandisk turn temporary scarcity into a higher earnings floor?

I ended that article by asking whether the gross-margin floor was 55%, 45%, or 30%. At the time, New Business Models covered just over one-third of FY2027 bits. I said coverage needed to cross 50%, Datacenter needed to become a larger part of the business, and the contracts needed enough financial protection to matter when pricing eventually weakened.

FY4Q crossed those thresholds. More than half of FY2027 bits are committed, roughly two-thirds of FY2028 bits are spoken for, QLC Stargate is shipping for revenue, and customers that signed only one quarter ago are already asking for more supply over the next three to five years. The contract mechanism is no longer a thesis. It is becoming the business.

That changes my view. I no longer think Sandisk should be analysed mainly as a memory producer enjoying the strongest cycle in its history. It is building a contracted core inside that cyclical business, and the size of that core is now large enough to change normalised earnings.

A Fair Return

Sandisk has agreements with eight Datacenter and Edge customers. Their weighted-average duration is more than four years. Pro forma remaining obligations are $91.1 billion, backed by $16.5 billion of financial guarantees. Pricing combines fixed and variable components, with floors and ceilings, while supply and demand are specified by year and quarter.

This is a different commercial relationship from the one memory companies have traditionally had. In the old model, Sandisk invested against an industry forecast and renegotiated with procurement teams every quarter. The company carried the capacity risk; the customer retained the right to wait for lower prices. In the new model, customers secure future supply by committing volume, product mix, and financial protection years ahead.

Management says NBM gross margins should remain around 80%, with some upside when market prices rise. That means Sandisk is giving up part of the final price spike in exchange for a floor. The choice is deliberate. Management is choosing duration over maximising every last dollar of the boom.

The customer relationship may matter as much as the price floor. Sandisk says the agreements include detailed forecasts of product mix and deployment timing. That pulls the company closer to the architecture decisions that determine which capacities, controllers, and performance characteristics future AI systems require. A supplier that learns about demand after the product is designed competes on price. A supplier involved while the product is being designed competes on execution and trust.

The buyback reinforces the same choice. Sandisk repurchased $4.5 billion of shares during FY4Q and has $15.5 billion of authorisation remaining. A contracted earnings base makes future cash flow easier to plan; a low memory multiple lets the company turn that cash into a larger claim on future earnings for each remaining share.

From Acceleration to Duration

Two things moved in opposite directions this quarter. Contractual duration became much stronger, while cyclical acceleration weakened.

Revenue grew 51% sequentially, but one-third of the increase came from bits and two-thirds from pricing. The next guide assumes both bit growth and only modest further price increases. Gross-margin guidance is roughly flat. The business remains at an exceptional level, but the slope is changing.

That is not just a Sandisk issue. Memory spent five quarters as the cleanest expression of AI scarcity: every report brought larger price increases, larger beats, and another estimate reset. The TMT discussion described the current phase well: the business has reached cruising speed. Across the sector, beat magnitudes are shrinking, some Asian suppliers are missing, and leadership is rotating toward other parts of the AI buildout. Part of the selling is fundamental, not merely a crowded-position unwind.

CJ Muse framed the new debate directly:

“The challenge for semi investors is moving less of a focus on margin and EPS revisions, but rather buying into the durability of this cycle.”

The rest of the call followed that question. Bernstein’s Mark Newman pressed on the lighter revenue guide. Analysts asked why margins were not rising further, whether customers could walk away, and when supply would catch demand. They were not challenging whether AI demand exists or whether the contracts are real. They were challenging management’s larger claim that those contracts have changed the cycle.

I think both sides have part of the answer. The rate of estimate improvement is peaking. The old source of stock upside—another ASP surprise followed by another guide raise—is fading. But slower acceleration is not deteriorating durability. Sandisk is building more duration just as memory supplies less momentum.

Two Books, One Income Statement

Sandisk now contains two economic books. The contracted book has multi-year purchase commitments, price floors, financial guarantees, and detailed product visibility. The uncontracted book still floats with the NAND market.

The valuation debate comes down to how those books blend when spot pricing falls. If 60% of output earns an 80% gross margin and the residual 40% falls to 25%, the blended margin is still 58%. At two-thirds coverage, it is about 62%. Revenue per bit will differ across products, so this is not an earnings forecast. It is the reason the old trough framework no longer fits.

The market is right that an 84.6% company-wide gross margin is not a sensible terminal assumption. It is wrong to assume that Sandisk must therefore return to old-cycle margins. That would require contract coverage to reverse, contract economics to fail, or both.

I would not build the thesis around management’s forecast that the NAND market approaches $500 billion in 2027. Sandisk did not show how much of that forecast comes from bits, price, or mix, and the number requires extraordinary pricing to persist. The contracts are better evidence. They show what named customers have agreed to buy, the minimum economics Sandisk expects to earn, and how much future production is already committed.

This is the updated variant perception: the next phase of the Sandisk thesis is a transfer from estimate upside to earnings duration. The company can earn less than the most aggressive FY2027 forecast and still be worth more if the remaining earnings persist for longer and the share count keeps falling.

The Test That Comes in a Downturn

There is an awkward feature to this thesis: the floor can only be proved when the cycle turns. At today’s margins, it is invisible. A downturn will initially look like the thesis is failing because spot pricing and the uncontracted book will weaken first. Only then will we see whether the contracted economics hold.

The $16.5 billion of guarantees matter, but I would not treat them as the whole enforcement mechanism. The stronger commercial deterrent may be allocation. A customer that breaks a commitment during a shortage risks losing preferred access to Sandisk’s future supply when qualified capacity is hardest to replace. That is an inference, not a disclosed contract term, but it explains why customers are willing to sign and then expand multi-year commitments. Management expects bits to remain on allocation beyond calendar 2027, while the guarantees rise in coverage relative to the remaining obligations as the agreements mature.

I do not need to wait for a downturn to decide whether the earnings floor has moved. The contract coverage, margin framework, purchase orders, and customer expansions are enough to conclude that it has. The downturn will determine the exact height of the floor, not whether one exists.

Three Futures

A three-year stock price should be based on the earnings available to investors at that point, so I use FY2030 EPS as the forward number for a 2029 valuation. The scenarios below are my estimates, not company guidance.

The bear case assumes the commercial model fails the test: customers renegotiate, variable pricing resets lower, coverage retreats, competitor capacity returns, and the uncontracted book collapses. This is a real bear case, not a softened base case.

The base case does not require Sandisk to escape the cycle. Revenue reaches $44 billion, margins normalise well below current levels, and EPS falls far below peak-cycle expectations. The return comes from applying 12 times earnings to a more dependable $151 rather than 7–8 times an apparently temporary $200.

The bull case requires contract coverage above 75%, continued customer expansions, disciplined industry supply, and aggressive buybacks. Under those conditions, Sandisk would have created a new earnings regime rather than merely a better cycle.

At roughly $1,350, the probability-weighted value is about $1,790, or close to a 10% annual return over three years. More important, the base case works without requiring the bull case to carry the argument. I would own Sandisk here.

What I Track Now

The first three items matter most. Coverage tells us how much of the business has been moved out of the quarterly market. Gross margin tells us whether the promised floor is real. The mix of bits and price tells us whether Sandisk can grow after the pricing cycle stops doing the work.

The Earnings That Stay

Management is no longer choosing between a slightly higher or lower quarterly margin. It is choosing what kind of company Sandisk becomes when the shortage ends.

The old Sandisk maximised the upcycle and accepted the bust. The new Sandisk is contracting supply, sharing risk with customers, planning products several years ahead, and using the resulting cash to shrink the share count. The peak rate of earnings improvement may be behind it. That is not the end of the thesis.

Sandisk spent five quarters proving how high earnings could go. FY4Q showed why far more of them may stay.

Subscribe now

• The margin arrived from an unexpected direction: adjusted gross margin rose 200 basis points to an all-time high even though net revenue margin was unchanged. DoorDash did not charge more; the fulfillment system became more efficient.

• The Digital Shelf remains under-monetized: DashPass is accelerating grocery adoption, advertising is already contributing to earnings, and merchant relationships contain commercial value that DoorDash has yet to quantify or fully exercise.

• Per-share conversion remains the constraint: the widening GAAP-to-adjusted gap, stock compensation, recurring legal costs and continued reinvestment mean stronger network economics may not flow cleanly to shareholders.

Two numbers from DoorDash’s second quarter define everything that follows.

Adjusted Gross Margin: 54.2%. All-time high. Up two hundred basis points year-over-year. Net Revenue Margin: 13.5%. Identical to Q2 2025. To the decimal. DoorDash expanded margins by two hundred basis points without moving its take rate a single basis point.

Adjusted EBITDA of $914 million beat consensus by 8.4% and cleared the top of management’s own guidance range. GAAP operating income was $156 million, down from $163 million a year ago on 36% revenue growth, missing consensus by 19%. The stock moved four-tenths of a percent.

In The Two Businesses Inside after Q1, I wrote that margin expansion had to come from the intent layer, because: “Marginal costs are real and irreducible. Labor scales linearly with volume.” In The Meituan Referendum after Q4, I identified the Digital Shelf, grocery-enabled CPG advertising, as the single variable determining whether DoorDash was a $150 stock or a $300 stock. Both pieces assumed the fulfillment layer was the substrate that needed to reach zero. The intent layer was where the margin lived.

That framework needs correcting. The fulfillment layer didn’t get to zero. It got cheaper. On its own. While the intent layer had only been partly exercised.

Management knows it. They broke their shareholder letter format, said so explicitly, and spent five pages arguing DoorDash is a compounding membership flywheel, organized around strategy rather than financial results. They published two cohort charts they have never shown before: Adjusted Gross Profit per MAU and DashPass penetration compounding together over five years. Ravi Inukonda volunteered that the EBITDA beat came “later in the quarter” and they didn’t have time to reinvest, pre-framing Q3 before anyone asked. Twelve analysts asked questions. Not one pushed back.

This is a management team constructing the communication infrastructure for reclassification. The cohort charts are not evidence offered in passing. They are strategy. And the sell side is buying the narrative without interrogation.

The Cost Curve

Here is the mechanism behind the 200 basis points, and the caveat that matters.

Cost of revenue, as a percentage of Marketplace GOV, fell from 6.7% in Q2 2025 to 6.4% in Q2 2026. Contribution margin went from 34.9% to 36.8%. When Doug Anmuth of JP Morgan asked about the take rate improvement, Ravi gave an answer more useful for what it ruled out:

We’re not operating the business towards take rate or net revenue margin percentage. Our goal has been always to optimize for overall profit dollars… I would think about it as flattish from Q2 to Q3, and then Q4, which is normally a quarter with higher Dasher costs, it’ll be slightly lower.

Management is guiding you not to model take rate expansion. The margin improvement came from below the revenue line.

I should be precise about what this does and does not prove. The year-over-year comparison strips out seasonality, Q2 2025 had the same seasonal Dasher cost profile as Q2 2026, and adjusted gross margin was 52.2% then versus 54.2% now. That is not a seasonal pattern. It is also not fully explained by Deliveroo consolidation, since Deliveroo was already contributing in Q4 2025 and Q1 2026 without producing a comparable gross margin expansion.

The evidence is consistent with a density-driven cost curve: better batching, routing, Dasher utilization per zone, and international unit economics converging toward the US standard. But the materials do not provide a full bridge showing how much came from structural delivery efficiency versus Deliveroo mix, advertising contribution, insurance timing, or basket composition. Management cited ads and subtotal as specific drivers and attributed take rate movement primarily to seasonal Dasher costs.

The honest read: something structural improved in the economics of fulfillment. The precise decomposition is not observable from outside. The direction of the change matters more than the magnitude of any single quarter, because if fulfillment itself has an improving cost curve at DoorDash’s density, it changes the architecture of the entire thesis.

The prior articles framed it as: fulfillment gets to zero, then advertising takes over. The actual structure may be different. If fulfillment gets cheaper over time, and advertising layers on top, the two businesses don’t merely stack. They compound. That shifts the terminal margin ceiling upward, not by the 200 basis points per year I asserted in an earlier draft (one quarter does not establish an annual rate, and cost curves flatten as efficiency improves), but by a cumulative 100-200 basis points through 2029 in a base case, with more available in a bull case where the global technology platform and autonomous delivery contribute additional gains.

Directionally, DoorDash’s delivery economics appear to be converging toward Uber’s levels on a steeper improvement trajectory, though segment definitions differ enough that a precise numerical comparison is unreliable. The conceptual point stands: DoorDash has order-volume share leadership and the relationship between density and delivery cost is non-linear. A logistics business whose unit costs decline non-linearly with scale is exhibiting economics that sit somewhere between pure logistics and platform.

DashPass Before the Digital Shelf

The most important update to the earlier thesis is one of sequencing.

The prior articles assumed the path ran: grocery reaches breakeven → CPG advertising unlocks → margins expand. Q2 suggests the mechanism starts one step earlier.

DoorDash disclosed that it added more paid U.S. DashPass members in the twelve months through Q2 than in the previous twenty-four months combined. As cohorts age, DashPass penetration rises, order rates increase, and adjusted gross profit per MAU increases. Most tellingly: DashPass members placed approximately 75% of U.S. grocery and retail orders during Q2.

DoorDash is not acquiring a restaurant consumer and then separately acquiring a grocery consumer. It is making restaurant delivery reliable enough to become habitual, converting the habitual consumer into a member, and using membership to lower the cost of trying grocery and retail. More categories make DashPass more valuable. More DashPass members make new categories easier to scale. The Digital Shelf may be the margin engine. DashPass is what keeps filling it.

There is a cost to this loop. DashPass orders carry lower gross-margin percentages because members pay lower consumer fees. Management accepts this trade when frequency, retention, and lifetime profit compensate, and Q2’s cohort charts argue it does. But the company still does not disclose the cohort-level cash economics needed to settle the question quantitatively. The direction is clear. The magnitude is taken on trust.

The Lever Nobody Asked About

If the cost curve is doing margin work on its own, the monetization layer becomes additive rather than compensatory. That changes the math.

Michael Morton of MoffettNathanson opened Q&A with the question that matters most. Some grocery partners reportedly pay near-zero take rates. What’s the repricing opportunity? Tony Xu declined to comment on any specific partner, then said this:

When you are the fastest grower in the market for them, and you are their source of growth, we might be 100% of the growth that they see, you certainly have opportunities to grow your business with them.

I don’t want to overstate what that means. “Opportunities to grow your business with them” could mean higher take rates, but it could also mean advertising, fulfillment services, data products, or expanded assortment. It is not a confirmation of unused pricing power. It is a statement that commercial bargaining position exists and management sees ways to monetize it. But it is still a CEO describing latent economic value in a relationship the market cannot observe.

The grocery thesis is advancing alongside it. DashMart warehouses run near 24/7 with what Tony described as “10x better error rates” because DoorDash controls the inventory. The blind storefront problem I identified in The Ice King, the information layer promising something the physical layer cannot reliably deliver, is being solved through vertical integration, exactly as the Tudor framework predicted. Management reaffirmed new verticals reaching gross profit positive in H2 2026. The Digital Shelf thesis isn’t wrong. It is understated, because it sits on top of a cost structure that is itself improving.

Then the silence. When Youssef Squali of Truist asked what drove the EBITDA outperformance, Ravi answered:

The unit economic improvement came in ahead of our expectations, specifically in a couple of areas. Ads was one of them, subtotal was the other one.

Twelve analysts asked questions. Not one, across MoffettNathanson, Evercore, Bernstein, JP Morgan, Morgan Stanley, Bank of America, Citi, Barclays, asked how large the advertising business is.

The advertising lever is not untouched. It is already contributing to results, Ravi said so. But it is under-disclosed. Instacart breaks out advertising separately and is valued on it. Uber discloses an advertising run-rate. DoorDash does neither, which means sell-side models assign marketplace economics to a revenue stream that likely carries much higher incremental margins. Companies segment-report when they want a business valued separately. SevenRooms venues growing 100%+ YoY, digital ordering at 150,000+ merchants with revenue up 40%+, reservations bookings up 150% sequentially, all described in prose, none tabulated.

The variant perception, stated carefully: the market appears to capitalize DoorDash’s current margin trajectory while assigning limited incremental value to a separately disclosed advertising or merchant-services segment. That segment is growing and already affecting EBITDA, but cannot be modelled from outside. The day DoorDash tabulates it, the sum-of-the-parts changes. That day hasn’t come.

The Ice Houses, Rebuilt

Two threads from the prior articles that Q2 resolved or materially advanced.

The Ice King assigned meaningful probability to a $1.5-2 billion Deliveroo impairment by 2027-2028. European food delivery had collectively destroyed $20 billion in shareholder value. Q2 showed the opposite. Deliveroo accelerated across MAUs, Total Orders, and subscription growth, all at two-year highs, while exceeding internal profit expectations and turning contribution profit positive. The global technology platform completes H1 2027; the efficiency benefits haven’t arrived yet. Management backed the thesis with capital: $1.05 billion of buybacks at roughly $154 average, the same buyback whose twelve-month absence I flagged in Q4 as a bear signal. That signal reversed.

Deliveroo has moved from probable problem to promising proof point. It has not yet been proven as a successful acquisition. DoorDash still does not disclose Deliveroo’s standalone EBITDA, cash contribution, integration spending, or return on purchase price. Contribution-profit positivity is evidence. It is not a return calculation.

The Ice King also posed the refrigeration question: would DoorDash’s infrastructure compound or be obsoleted? Justin Post of Bank of America asked about agentic AI traffic. Tony:

I think there’d be two big wars. One is the battle for attention, and you see that playing out with chat assistants. The other is the battle for atoms… Our focus is squarely on making sure that we master the physical world so we can be the most useful to all these digital assistants.

DoorDash isn’t waiting for the refrigerator. It’s building it, the Autonomous Delivery Platform orchestrating human dashers, Dot robots, DoorDash Air drones, and third-party AVs in a single routing system. Dot targeting a high single-digit percentage of Phoenix orders by year-end is the first falsifiable commitment. And Meituan’s 2025 experience, where JD.com and Alibaba destroyed sixty billion yuan of profit at full maturity, is a reminder that the cost curve, not pricing power, is the moat that survives competitive attack. DoorDash appears to be building the right one.

The Counterweight

The bears are not wrong about everything.

GAAP operating income declined 4.3% year-over-year on 36% revenue growth. The bridge from $199 million of net income to $914 million of EBITDA is $715 million, up 93% year-over-year, now 78% of the adjusted number.

Three items that need stating plainly. SBC was $349 million, 7.8% of revenue. Management “lowered” the FY guide to $1.2-1.3 billion, but H1 was $580 million, implying H2 of $620-720 million. Q2 is the run-rate, not the anomaly. The legal and regulatory add-back ran $29M, $48M, $29M, $45M, then $98M across five quarters. A charge excluded from EBITDA every quarter for years is an operating cost, and at $250-300M annualized it represents 7-8% of adjusted EBITDA. R&D grew 52% to $535 million, funding the technology platform and autonomous delivery.

Nobody on the call raised any of it.

The honest way to see this company’s earnings is through three lenses, not one. GAAP operating income was $156 million, the floor, burdened by acquisition amortization, SBC, and everything else. Adjusted operating income excluding the roughly $112 million per quarter of finite Deliveroo intangible amortization, which is non-cash and runs off on a defined schedule, but whose associated acquisition consideration was very real, was approximately $268 million, up roughly 64%. Owner earnings, after adding back the amortization but deducting normalized SBC and recurring legal costs, sits somewhere between these two. None of these views is more correct than the others. Each answers a different question.

On cash: management disclosed that FY26 reported free cash flow will be reduced $700-800 million by year-end merchant-payment timing. Clean FY26 FCF is therefore approximately $3.7-3.8 billion. The stock trades at roughly 24 times clean free cash flow, not the 30 times the headline implies. That is a useful anchor.

The GAAP-to-adjusted gap is real, and it widened this quarter. The cost curve argues it is transitional, the underlying business economics are improving even as the P&L carries acquisition and investment noise. But the gap is widening faster than the business is growing, and if the cost curve flattens while the add-backs persist, the adjusted story was always the generous one.

Three Worlds, 2029

The Q4 article’s base case was $175. The Q1 article’s was $210. FY2026 EBITDA is tracking approximately $3.73 billion, roughly matching the Q4 article’s 2027 base case. We are ahead of schedule on profitability. But the right response to being ahead of schedule is to revise the estimates, not to assume the rate continues.

A margin bridge makes the assumptions auditable:

The reinvestment deduction is where intellectual honesty lives. DoorDash’s reward for proving one investment works is permission to begin three more. Management keeps finding attractive areas to fund, grocery infrastructure, autonomous delivery, merchant software, international expansion. The company may grow for longer than expected while converting less of that growth into per-share free cash flow than expected. That tension must be in the model, not outside it.

Share counts assume continued buyback at roughly the current pace (~$1.5B annually) partially offsetting SBC dilution of 8-10 million shares per year. Net cash of $8-12 billion by 2029 (varying by scenario) is added to the EV-to-equity bridge. These are approximations, not a full equity model, but the arithmetic is visible.

At approximately $208, the stock prices the lower half of the base case against a probability-weighted value of roughly $260. That gap is real but not extreme, roughly 25%, and the catalysts are defined: Q3 EBITDA in November, advertising disclosure at any point, and Deliveroo’s continued trajectory.

The updated BFQ. The prior articles asked whether DoorDash was a logistics company or a media company. That question has partially resolved. The new question: How long does the fulfillment cost curve run before it flattens, how much under-disclosed monetization remains above it, and how much of both reaches each share rather than funding the next layer of expansion?

The first two parts are bullish inputs. The third is the constraint that keeps this from being a simple call.

The variant stated carefully. Consensus sees a delivery company in an investment cycle that will eventually produce margin expansion. The variant: the margin expansion has already begun in the cost curve, and the monetization layer, advertising, grocery commercial terms, merchant services, is under-disclosed and only partly exercised. The market is pricing the margin DoorDash has earned. The less visible question is how much additional margin exists that DoorDash has not yet chosen to show.

What Resolves This

Six signposts, each with a threshold.

Q3 Adjusted EBITDA (November). Above $1,060M strengthens the case that Q2’s improvement was structural. Below $990M suggests it was timing and reinvestment deferral.

Q3 Adjusted Gross Margin. At or above 54% means the year-over-year expansion holds on a different seasonal base. Below 52.5% means the curve may have flattened or Q2 was mix-driven.

Advertising disclosure. Any quarter, any format. The single cleanest re-rating catalyst, the moment the second business becomes defensible in numbers rather than prose.

Organic GOV growth. Above 24% reverses the Q1-to-Q2 deceleration. Below 21% compresses the multiple regardless of margin.

Dot in Phoenix by year-end. Hit means the autonomy R&D is a funded option with observable progress. Miss or silence means discount the autonomous delivery narrative.

Legal/regulatory add-back. Below $55M in Q3 means the $98M was discrete. Above $80M again means treating it as a permanent operating cost and haircutting EBITDA accordingly.

The ice houses are still useful. Q2 showed us something the prior articles didn’t expect: the ice is getting cheaper to cut even before Tudor raises the price. The DashPass flywheel is filling the shelves. The advertising layer is contributing before being disclosed. Deliveroo is improving before the platform rebuild is complete.

The compounding mechanism appears real. The remaining question, the one that keeps this from being a simple buy, is how much of that compounding reaches each share, and how much funds the next ambition management decides to pursue. DoorDash has begun proving the network generates more value as it grows. It has not yet proved how much of that value it intends to keep

Subscribe now

• The flywheel is working: Brazil items sold rose 56%, ecosystemic users grew 37%, credit quality improved, and operating expenses gained 260 basis points of sequential leverage.

• Shareholders still cannot see the payoff: management reinvested all the operating leverage into lower prices and incentives, while declining to disclose the ecosystemic-user penetration and economics needed to value the strategy.

• The debate is now about time: MercadoLibre appears to be building a more valuable integrated system, but the stock needs evidence that customer value will eventually become reported margins and per-share earnings.

From MercadoLibre’s Q2 2026 shareholder letter:

Contribution profit per ecosystemic user is multiples of the sum of the profit of a marketplace-only user and a fintech-only user. Profitability doesn’t simply add when a user becomes ecosystemic, it multiplies.

A year ago, we called the free shipping threshold cut a strategic gambit, a victory lap disguised as a skirmish. Since then: four articles, one consistent thesis, and a stock that went from roughly $2,700 to $1,888. Down 30%.

The commerce call was right. Every signpost cleared. Brazil items sold grew 56% year-on-year. Unit shipping costs kept falling. The managed network hit 96% penetration. Ecosystemic user growth in Brazil accelerated from 35% before the threshold change to almost 50%. The logistics density thesis played out exactly as we described in the first article, then the second, then the third.

The credit call was wrong in Q4 and Q1. We underestimated how fast a $12.5 billion lending operation would change how the market categorized the entire business. We wrote about the logistics war while the credit transformation happened in peripheral vision. Q2 materially weakened the immediate bear case, NIMAL improved 290 basis points, NPLs dropped to historic lows, but the damage to the stock was already done.

The margin call was half right and half wrong. We said operating leverage would emerge. It did, opex improved 260 basis points sequentially this quarter. We assumed management would show some of it. They showed none. They reinvested every dollar.

The hard lesson, five quarters in: being right about the flywheel was necessary but not sufficient. The market doesn’t pay for flywheels it can’t see in the P&L. What we underestimated, across every article, was that management would reinvest 100% of the operating leverage, every quarter, with no signal of when it ends. The gap between the business and the stock is not analytical error. It is a time horizon mismatch between a management team playing a decade-long game and a market that prices in quarters.

Q2 doesn’t close that gap. But it gives us, for the first time, the economic unit that might explain what’s on the other side of it.

The Surplus Question Gets a Candidate Answer

Our Q1 article asked the question that matters: who captures the surplus MercadoLibre is creating? The company was producing enormous value, cheaper shipping, faster delivery, broader selection, easier credit, better financial tools, but the stock case required showing that value accrued to shareholders after subsidies, provisions, capex, and funding costs.

Q2 gives that question a candidate mechanism. Not yet a complete answer. A mechanism.

The concept is the ecosystemic user, someone who engages with both the marketplace and Mercado Pago. Management says these users generate 70% more GMV than marketplace-only users. They buy 55% more items. Their payment volume is 90% higher. Their assets under management are more than double. And their contribution profit is not additive but multiplicative, multiples of the combined profit of a marketplace-only and fintech-only user.

These users are growing at 37% year-on-year, the fastest-growing segment since Q4 2023, and their share of the total user base has expanded by eight percentage points over that period. Credit card holders are two to three times more likely to become ecosystemic than users without the card. MELI+ subscribers grew 72%.

Here is what connects every investment we’ve written about across four articles. The free shipping threshold converts marketplace-only buyers into more frequent buyers who adopt Mercado Pago. The credit card converts Pago users into marketplace shoppers. PIX discounts drive payment adoption. 1P and cross-border trade expand selection that pulls users deeper into the ecosystem. MELI+ bundles it explicitly. These are not separate initiatives with separate ROI calculations. They are different doors into the same room.

The argument is credible. The disclosure is incomplete. MercadoLibre did not tell us what percentage of users are ecosystemic. It did not disclose absolute contribution profit, conversion cost, or the time required to recover that cost. Management gave us the uplift because it supports the story. It withheld the denominator that would let us value it. When BTIG’s Marvin Fong asked directly for the penetration number, Martin de los Santos declined:

We don’t disclose the actual share of users. We just want to make sure that those are important users to us, growing very fast.

Management wants the concept in the narrative without the number that constrains it. That tells you the ecosystemic user may be MercadoLibre’s real economic unit, but investors still cannot build an ecosystemic-user model.

Three things need to be separated. Q2 provides strong evidence that the system creates more user value. It provides meaningful evidence that the system creates more company-level economic value. It provides incomplete evidence that this value will accrue to shareholders within a reasonable period. The article, and the investment case, rests on preserving that hierarchy.

Management Is Selling the Ecosystem. The Market Is Pricing the Balance Sheet.

Management spent the entire shareholder letter and earnings call selling this reframing. CEO Ariel Szarfsztejn closed the call with the thesis stated as plainly as a management team will ever state it:

We are building something quite unique globally, an ecosystem of commerce and financial services that compounds on itself. These are not two businesses running in parallel. This is one flywheel, each side making the other one more valuable.

No timeline. No margin target. No guidance the market can model against.

The questions on the call revealed where the disagreement sits. Goldman’s Irma Sgarz led with “feel obliged to ask about margin” and pushed on gross margin compression sources. Cantor asked whether current margin levels were sustainable through the second half. Citi pressed on credit card NIMAL inflection. UBS challenged the 90+ day NPL metric. BTIG fished for the ecosystemic penetration number.

The questions split into three camps: analysts exploring the flywheel and its upside, analysts testing the timing of margin realisation, and analysts still focused on credit risk. The second group dominated the call. Nobody disputed that engagement improved. Everyone wanted the bridge from engagement to earnings.

Management’s answer was consistent: credit profitability recovered, opex scaled, and management then reinvested those benefits into commerce. That is exactly the pattern we’ve tracked for five quarters. The analysts accept the evidence that the investments work. They are not willing to pay today for profits management has neither quantified nor scheduled.

“Elected”

Here is the section that matters for the stock.

Product development expense scaled from 8.4% of revenue to 7.2% year-on-year. AI productivity drove it: code submissions up 110%, engineering headcount flat for the first time in years, more than half of Help Portal requests resolved by the AI assistant without human intervention. Provisions flat quarter-on-quarter despite credit portfolio growing 75%. Total opex improved 260 basis points sequentially. The operating leverage we spent four articles waiting for showed up in the numbers.

Then management launched new PIX buyer discounts and seller take rate cuts in Brazil. In April. During the quarter. Gross margin compressed 280 basis points sequentially, consuming every dollar of opex improvement. Martin de los Santos on the call:

We elected to reinvest that margin into other areas of the business.

Elected. He used the word three times.

The business gets stronger. The margin doesn’t move. The stock sits.

Our Q1 article asked whether the dial has a minimum setting that competition enforces. Q2 doesn’t fully answer that. Whether the new Brazil initiatives were proactive or reactive doesn’t determine whether they’re discretionary in an economic sense. Management may invest proactively because it sees competitive dynamics, Shopee in low-ASP, Amazon in logistics, Nubank in financial services, that make it necessary even without an immediate threat. The earlier framework remains the honest one: management controls the intensity, but competition and customer expectations influence the minimum.

What Q2 does clarify is that the leverage mechanism works. The opex lines proved it. If management ever chose to let two or three quarters of opex improvement flow through without launching a new initiative, operating margin would move above 8% rapidly. They have not chosen to. There is no external catalyst that forces them to. It comes when they decide, and they have given no signal of when that will be.

Behind the reinvestment decision, though, something structural is building that deserves attention. The AI-driven productivity gains, Product Development at 7.2% and falling, customer service running on 30% fewer humans while the business tripled, 2026 the first year engineering isn’t growing, represent 100-200 basis points of annual operating leverage that compounds regardless of management’s commerce and fintech investment decisions. This happens in the background. Over three years, it could contribute meaningfully to margin recovery even if management continues to invest aggressively elsewhere. The market is not modelling this because it is distributed across multiple line items. It may be the most underappreciated structural feature of the entire P&L.

Credit Recovered. The Clock Did Not.

The Q4 article’s $12.5 billion identity crisis, the market re-categorizing MELI from tech platform to Latin American lender, has a materially better data set now. Not a resolution. A better data set.

NIMAL recovered from 17.8% to 20.7%. The 15-90 day NPL dropped from 8.0% to 7.0%, near historic lows. Credit card NPL at 4.6%. Provisions flat quarter-on-quarter. Credit revenue net of provisions as a share of revenue higher than a year ago. The portfolio is scaling profitably at the current level of growth.

Our Q4 signpost said NIMAL must hold above 22%. It didn’t. The miss is from credit card mix, now 47% of the portfolio at -2.5% NIMAL because issuance accelerated to 2.6 million cards per quarter. Each cohort breaks even in 12-18 months, but every wave of new issuance resets the average age of the book. The portfolio is majority young. Recalibrate: watch NIMAL trending toward 22% as issuance growth moderates and cohorts season.

The immediate bear case, credit growth outrunning underwriting, weakened materially this quarter. The structural credit debate, whether the portfolio can survive a full Brazilian downturn at this scale, did not resolve. UBS correctly flagged the 90+ day NPL uptick. Management attributed it to technical factors and product mix. Q3 will confirm whether it was noise or signal.

Our View, Updated

We were right that logistics density could be used offensively. We were right that integrated commerce, payments, credit, and data create a system competitors cannot easily replicate. We were right that Engine A, advertising growing 62% FX-neutral and now above 10% of Latin American digital ad share, AI productivity compounding in every operating line, would eventually become visible.

We underestimated how quickly credit would change the company’s financial identity. We underestimated how long the investment cycle could persist. Most of all, we treated value creation as if it would automatically become shareholder value. The Q1 article corrected that by asking who captures the surplus. Q2 gives the surplus question a named mechanism, the ecosystemic user, without yet proving how much of the value reaches the bottom line.

Our variant view is narrower than it was a year ago: MercadoLibre is becoming a more valuable integrated system even while reported margins remain low, and Q2 suggests the cost of building that system may be temporary rather than permanent. The missing variable is the meaning of “temporary.”

The bear case does not require MercadoLibre to lose. It requires the surplus to keep flowing to customers, sellers, and credit losses while margins stay near current levels. The base case assumes the ecosystem keeps compounding, credit seasons normally, and the cost of defending the system is higher than we once assumed but manageable. The bull case is management’s story: ecosystemic users become a large share of the base, credit cards turn profitable, advertising and AI change the margin structure, and the company gets re-categorized as integrated digital infrastructure.

Signposts graded and updated: Brazil FXN GMV above 35%, cleared at 39%, maintain. NIMAL trending toward 22%, below at 20.7%, recalibrate to “watch as issuance moderates.” 15-90 NPL below 8%, cleared at 7.0%. Operating margin above 7.5% as inflection signal, below 6.5% as stress. Advertising FXN growth above 50%, cleared at 62%, watch for quarterly revenue approaching $500 million. New: ecosystemic user penetration disclosure, when management starts reporting the number, it means they’re confident in the curve. New: 90+ day NPL below 18% confirms noise, above 20% reengages concern.

Five quarters. Five articles. The commerce thesis was right. The credit thesis scared us and recovered. The surplus question now has a named mechanism, if not yet a proven one. The moat is wider than it has ever been.

And the stock is down 30% since we started writing.

MercadoLibre is building something that does not exist anywhere else, an integrated commerce and financial services flywheel in a market where both digital shifts are happening simultaneously to 650 million people. The flywheel works. Engagement compounds. Q2 suggests the economics may follow. Management still has to let them appear.

The price of the flywheel is time. And time, for now, is what management is buying, whether the market likes it or not.

Subscribe now

• The operating thesis strengthened: bookings grew 22% in constant currency, profit grew faster, AV utilization is improving, and Uber’s market position remains resilient in mature autonomous-vehicle cities.

• The valuation thesis weakened: Uber still withheld AV take rates and comparable Uber One metrics, while committing more than $10 billion to help build the autonomous supplier ecosystem.

• The key question has shifted: it is no longer whether Uber matters in an autonomous world, but how much capital it must contribute and how much cash ultimately remains for shareholders.

In May I wrote that Uber had moved from being priced as an autonomous-vehicle casualty in the high $60s to being priced as a business with a plausible coordination thesis at $80. I thought the remaining distance to my $120–$130 base case depended on Uber One becoming more deeply embedded, Delivery margins reaching 4%, and insurance savings accelerating U.S. Mobility.

Three months later, Uber delivered the strongest evidence for that thesis yet, and the stock returned to roughly $68. The repricing did not merely stall; it reversed.

That does not mean the original thesis was wrong. It means my explanation for what would move the stock was incomplete.

I had assumed Uber would rerate as management proved that autonomous fleets need its demand. Q2 suggests that proof is no longer sufficient. Uber is now providing increasingly precise evidence of what preserving that position will cost, while still withholding the numbers that show what it will earn.

The Scorecard

The honest place to begin is with the tests I set before the quarter.

After Q1, I wrote:

“If Uber reports Q2 without disclosing AV take rates, I will interpret that as a signal the economics are not yet where management wants them.”

Uber did not disclose the take rate. When Bank of America’s Justin Post asked how the more-than-$10-billion autonomous commitment would affect earnings and margins, Balaji Krishnamurthy said Uber would provide more visibility as deployment approached scale.

I do not think that answer means the economics are necessarily poor. The commercial arrangements are early, likely vary by partner, and may change as vehicle volumes grow. But I committed in advance to what continued silence would mean. I do not get to rewrite the test after seeing encouraging evidence elsewhere.

Uber One created a second disclosure problem. Q1’s defining number was fifty million members generating half of total Gross Bookings across Mobility and Delivery. In Q2, management said membership reached another record and that members generated more than 70% of Delivery Gross Bookings, but it did not update the member count or the share of total bookings.

The new statistic may be useful. It is not an update to the old one. It changes both the metric and the denominator. One quarter may be harmless; a second omission would look deliberate.

The operational evidence was almost the mirror image. Gross Bookings reached $58.0 billion and grew 22% in constant currency. Adjusted EBITDA increased 33%, non-GAAP operating income rose 40%, Delivery bookings grew 25%, and Delivery operating income grew 38%.

Insurance savings passed through lower prices produced faster trip growth in Los Angeles and San Francisco. Autonomous partners were completing trips in the high twenties to low thirties per vehicle per day. Uber’s category position in Los Angeles, San Francisco, and Phoenix; the mature U.S. AV markets, was higher than a year earlier.

The pattern is hard to dismiss: the operating tests strengthened, while the two disclosures needed to value that strength failed.

The Proof

The original Uber thesis began with utilization.

Human drivers provide their own vehicles and decide when to work. When demand falls, they turn off the app. An autonomous vehicle is a fixed asset that depreciates whether it is carrying a passenger or sitting idle. Its owner must cover charging, cleaning, maintenance, insurance, remote assistance, and financing before earning a return.

That makes dense demand more valuable after autonomy arrives, not less.

Q2 supplied the best evidence yet that Uber can provide it. AVs are now live on Uber in seven cities, with as many as fifteen planned by year-end. Partners have committed approximately 120,000 vehicles over the coming years, and management says Uber’s position has strengthened rather than weakened in the cities where autonomous competition is most advanced.

AV Labs adds another layer. Uber is operating sensor-equipped vehicles that collect rideshare-specific driving data for its partners. Combined with demand, dispatch, fleet operations, insurance, financing, charging, and local regulatory work, Uber is becoming an input into autonomous commercialization rather than merely another app through which a completed product can be booked.

The strategic argument is much stronger than it was eighteen months ago. Uber will almost certainly matter in an autonomous world.

And yet relevance is not ownership.

The Invoice

In 2005 Apple prepaid $1.25 billion to five memory suppliers to secure NAND flash for the iPod. Apple was not trying to manufacture semiconductors; it was using the strength of a customer relationship it already controlled to make sure scarce supply expanded around it.

Uber is attempting something similar, with one dangerous difference. Apple’s memory suppliers did not aspire to own the iPod customer. Uber’s autonomous partners may eventually want the rider, the app, and the entire trip economics.

Dara Khosrowshahi’s desired market structure is clear: several AV developers compete across technologies, vehicles, and regions; each benefits from Uber’s demand and local operating capabilities; and no single provider gains enough power to dictate commercial terms.

I had treated that fragmentation as the likely outcome of the industry. Q2 suggests Uber may have to help finance it into existence.

Uber expects to commit more than $10 billion across equity investments, infrastructure, and vehicle offtake arrangements. It is helping software partners raise capital, anchoring vehicle demand, collecting shared data, and building the capabilities needed to launch fleets city by city. Management intends to bring in outside financiers as deployments scale.

The intended loop is compelling. Uber’s recurring demand raises vehicle utilization; better utilization attracts suppliers and outside capital; more suppliers reduce Uber’s dependence on Waymo or any other developer; better coverage and price produce more demand.

But the loop begins with Uber’s balance sheet.

That creates two risks. The first is that Uber supports weaker suppliers that never become competitive with the leading autonomous systems. Lucid can have PIF backing, Nuro software, and a large Uber order without proving that it can manufacture the right vehicle at the right cost.

The second is that successful suppliers use Uber’s capital, data, and demand to reach scale before seeking better economics or distributing directly. Uber may help create the competition it needs and still fail to own the value that competition produces.

The non-obvious insight from Q2 is that Uber is using cash from its capital-light business to create the supplier market its capital-light future requires.

Where the Cost Appears

Uber does not disclose a clean bridge from AV investment to current operating expenses, so it would be wrong to treat every increase in corporate costs as autonomous spending. The consolidated results still show where the broader ambition first meets the income statement.

Source: Uber Q2 2026 earnings release.

Mobility and Delivery generated $775 million of incremental operating income, while corporate costs absorbed $168 million. That line contains shared finance, legal, technology, mapping, payments, and other functions, so it cannot be assigned solely to AV Labs, partnerships, or M&A. It is still the visible line through which some of the expanding corporate agenda reaches shareholders.

The Q3 guide made the tension more immediate. Uber guided to non-GAAP EPS of $0.84–$0.88, versus an analyst expectation of approximately $0.89, while its bookings midpoint was broadly in line with the Street. The quarter proved the strategic case but did not create near-term estimate upside.

Capital allocation told the same story. Uber generated $10.1 billion of trailing free cash flow, but deployed approximately $4 billion to establish a 37% economic interest in Delivery Hero and repurchased only $510 million of its own stock.

Reported free cash flow measures what the operating business generates. It does not measure what remains available to owners after acquisitions, strategic equity investments, and AV commitments.

Management is winning the strategic argument and losing the multiple because the costs can be placed into a model today, while the returns remain a distant assumption.

What the Next Year Can and Cannot Tell Us

The next few quarters cannot determine whether Uber’s AV strategy creates value. The vehicles, financing structures, supplier relationships, and commercial agreements will take years to mature. Uber itself has said meaningful commercialization will take much longer than the underlying technical progress.

Corporate costs moderating in Q3 would not prove that the $10 billion earns an attractive return. A restored buyback would not prove that outside financiers will eventually absorb the vehicle risk. Even a disclosed take rate would tell us little without knowing what Uber spent or guaranteed to obtain it.

What the next year can show is whether Uber can carry the build without weakening the business that funds it.

Through early 2027, the test is whether Mobility and Delivery continue producing bookings growth near 20%, profit grows faster than bookings, central costs become more controlled, leverage remains bounded, and buybacks recover. That would show that management can incubate a long-duration project without sacrificing current per-share compounding.

During 2027, the test shifts toward risk transfer. Uber needs to demonstrate scaled third-party vehicle financing, clarify which of the 120,000 commitments are firm or conditional, and show how much capital or guarantee exposure remains with Uber.

Only from 2028 onward can the final economic question begin to resolve: AV gross profit, take rate after all support, return on Uber capital, utilization across several suppliers, and whether those suppliers remain on Uber after reaching scale.

The next few quarters do not resolve the long-term thesis. They determine whether Uber earns the right to keep pursuing it.

Three Futures

These scenarios use my assumptions from the FY2026 earnings base:

The bear case is not Uber disappearing. The company remains relevant, but central costs stay elevated, Delivery Hero consumes capital through its integration, suppliers gain bargaining power, and buybacks become residual. Earnings rise, but the multiple remains compressed because strategic importance never becomes distributable cash.

The base case assumes that Mobility and Delivery continue compounding, Uber contains the incubation cost, and third-party financing gradually takes over much of the vehicle funding. AVs remain too small to drive the consolidated model, but they no longer justify a large displacement discount.

The bull case requires several competitive AV suppliers, visible economics above a 20% take rate, limited retained fleet risk, and a core business strong enough to fund investment while reducing the share count.

The probability-weighted value is approximately $112. The wide range does not mainly reflect uncertainty about how many robotaxi trips Uber completes. It reflects uncertainty over how much capital Uber must leave behind and how much cash per share remains for owners.

I still believe the coordination thesis. Q2 made it stronger. What changed is my understanding of why the stock has not responded.

I expected Uber to rerate when it proved that autonomous vehicles would need its demand. Instead, each new piece of proof arrived with a more visible invoice and no disclosed toll.

Uber is building a long-duration asset whose value will not be known for at least another year or two. The next few quarters will not tell us whether that asset works. They will tell us whether Uber can afford to wait.

Uber has proved that it can make autonomous vehicles busy. It has not yet proved that it can stop paying once they are.

Subscribe now

• The core continues to compound: Toast added a record 9,500 net locations, recurring gross profit reached $595 million, and monetisation rose to 98 basis points without relying on higher pricing.

• AI now has revenue attached: Toast IQ Grow is on track to become the company’s fastest product to $10 million in ARR, using transaction-level data to identify opportunities, launch campaigns, and measure the orders they generate.

• The economics remain unproved: Early demand and customer outcomes are encouraging, but Toast has not disclosed paid accounts, retention, service intensity, or gross margins, the evidence needed to show that Grow can scale like software rather than a managed service.

One Idea, Several Names

For eighteen months our argument has been the same. Financial technology accounts for most of reported revenue, so the market files Toast alongside Square and Clover and applies a payment multiple. That is the wrong lens, because no restaurant selects Toast for its processing rates. They select it because Toast runs the business, and payments flow through because of owning the transaction layer. Toast is a vertically integrated restaurant operating system with payments as the monetization rail.

We gave that one idea three names across four articles, the integration layer, the profit sanctuary, the invisible manager, which was a writing habit rather than a change of mind. Two things in the record deserve correcting.

The first is timing. For six quarters we pointed at Toast IQ adoption figures and treated usage as evidence the AI layer was arriving. Usage is not revenue, and a market that declines to underwrite what it cannot measure was behaving sensibly. We were describing a mechanism and asking to be paid for optionality.

The second matters more. We consistently defended Toast’s decision to reinvest rather than harvest, distinguishing chosen lower margins from structurally lower margins. That distinction holds. But we treated the act of reinvestment as though it were evidence of value creation, and it is not. Investment is not the conclusion. Return on investment is. Toast still discloses no acquisition cost, payback period, or contribution margin for any business it is funding.

Q2 does not require a new identity for Toast. It gives better evidence about the company we have been studying since the start.

The System Still Compounds

Toast added a record 9,500 net locations, taking the base to roughly 180,000, with the majority from core SMB and mid-market rather than a few large enterprise deployments. GPV reached $60.7 billion and recurring gross profit $595 million.

Location count measures distribution, ARR measures monetization, and recurring gross profit measures how much Toast actually captures from that distribution. All three rose, and gross profit rose fastest. GPV per location was flat after falling 1% in the fourth quarter, SaaS gross margin expanded roughly 240 basis points, and total take rate, recurring gross profit over GPV, reached 98 basis points, up five year over year. Elena Gomez attributed the take rate gain to product adoption and cost optimization rather than pricing, which makes it durable. Toast did not sustain growth by adding visibly weaker customers.

One detail sits awkwardly. Net adds beat expectations by more than a third while ARR landed at consensus, which raises the question of why such a large location surprise produced no monetization surprise. New locations ramp over time, many may have gone live late in the quarter, and enterprise or international deployments may carry lower initial ARPU, but Toast disclosed no cohort data to settle it. Worth tracking rather than a thesis break.

The Closed Loop, and What It Hasn’t Proven

Toast IQ Grow is a marketing product built on restaurant data. Management said it is on track to be the fastest product in company history to $10 million in ARR, that it already runs at positive gross margins, and that one restaurant cut agency spending 70% while generating over $100,000 in marketing-attributed sales in under two months.

What matters is not that Toast built an AI marketing tool. What matters is that it can close the loop. A standalone vendor can produce a competent email. Toast can observe that Tuesday nights are underused, identify guests who stopped returning, select an offer with attractive margin, deliver the campaign, process the resulting orders, and verify whether incremental sales occurred. Narang put it directly:

“Because we power the point-of-sale, we can close the loop with each campaign tied directly to the orders it drives.”

The model is not the scarce asset, since foundation models are available to everyone. The harder combination is context, distribution, workflow access, and transaction-level measurement inside one system. This is why AI most likely increases the value of Toast’s existing position rather than creating a different company.

The prize is adjacent spending. Restaurants outsource marketing, payroll, bookkeeping, tax, and scheduling, and as Narang noted, they are often “spending a multiple of what they spend on software” on those functions. Toast’s total ARR works out to roughly $13,400 per location. It does not need to capture all of the surrounding budget for the economics to change; a modest share would materially raise revenue per location. How much is technologically addressable and economically capturable, this quarter did not answer.

Which is the discipline required here. Three standards of proof apply. Demand proof asks whether restaurants will buy, and Grow’s trajectory suggests some will. Outcome proof asks whether the product creates measurable value, and the early figures suggest it can. Economic proof asks whether Toast can deliver at scalable margins, and that remains entirely open.

Bernstein’s Harshita Rawat pressed on exactly this, given the marketing success managers who review AI-generated work. Narang answered:

“As we’ve gotten some early scale, we’ve already seen the gross margins improve. Frankly, I have no concerns long term about what the gross margins of that business could be.”

Confident, possibly right, not a disclosed unit-economic model. Toast provided no paid customer count, no conversion rate, no ARPU, no retention, no current gross margin, and no ratio of accounts to marketing managers. A human-in-the-loop service can show improving margins at small scale and still settle far below software economics, and that gap is most of the upside case.

One further qualification to something we have argued before: more locations do not automatically make the agents smarter. They expand distribution and may enrich the benchmark set, but only if Toast can standardize patterns across restaurants that differ enormously in cuisine, format, and check size. A new taqueria improves the dataset for taquerias. Whether it improves the agent serving a steakhouse is an engineering question, not an arithmetic one.

The Same Disagreement, Third Quarter Running

Toast beat on revenue, recurring gross profit, EBITDA, GPV, and location adds, and the stock fell 1.3%. Adjusted EBITDA of $221 million included roughly $10 million from a tariff refund, leaving underlying EBITDA near $211 million against consensus of $195.7 million. Full-year EBITDA guidance rose to $805–825 million, a $15 million midpoint raise against a $25 million beat. Gomez explained the gap without hedging:

“We strategically chose to reinvest the tariff refund into key growth initiatives and to seed long-term bets. As a result, we’re increasing our full-year adjusted EBITDA guidance by less than the Q2 beat.”

Close to close, Q4 2025 took the stock down about 7%, Q1 2026 down 14.7%, and this quarter down 1.3%. Three consecutive quarters of strong results, guidance withheld, negative reaction. This is not a communications failure that better investor relations can repair. Management is optimizing for the number of durable businesses it can build; the marginal shareholder wants evidence core profits will reach consolidated earnings. The disagreement ends only when the newer investments produce enough gross profit to make their returns visible.

Q2 moved that forward without settling it. New market ARR is near $200 million and doubling, retail SaaS ARPU is approaching core levels within two years of launch, and Grow has revenue attached rather than usage statistics. Against that, free cash flow fell to $130 million from $208 million as inventories rose to $217 million from $114 million at year-end, with Toast securing components ahead of memory-market pressure. That may be the right trade, but the cash has left the building and the explanation now carries a deadline.

A note on something we over-weighted. Last quarter Toast Local had its own section as the answer to DoorDash’s point-of-sale ambitions. This quarter it appeared once, briefly, at the end of the final answer. The demand-side thesis is not dead, but the evidence is much weaker than on the merchant side and we should weight it accordingly. DoorDash remains the real counterpoint: Toast knows the restaurant better, aggregators own more consumer intent. Toast needs enough connection to the guest to avoid becoming a back-office utility while someone else controls demand.

Three Outcomes for the Same Company

None of these requires Toast to become a different business. In each it remains a restaurant operating system, and the only question is how deeply it monetizes that position. From roughly $32, with about 580 million diluted shares and $1.7 billion of net cash, looking to FY2028.

The margin trap (20%). Restaurant activity weakens, GPV per location turns negative, Toast Capital losses rise, and Grow finds adoption but settles into managed-service economics. The core stays valuable while reinvestment fails to earn its cost. Revenue compounds at 12–14%, margins near 11–12%, and at 10–12 times EBITDA that supports roughly $22–28.

The quiet compounder (55%). Toast keeps taking share, attachment rises, enterprise and international scale steadily, retail becomes a credible adjacency, and AI improves support costs and retention without becoming a separate profit pool. Revenue compounds at 17–19%, margins reach 15–16%, and at 16–18 times EBITDA that supports roughly $45–57.

The expanding wallet (25%). Grow proves repeatable, a second agent succeeds, human review per account falls, and restaurants shift part of their outsourced spending onto the platform. Total monetization moves from 98 basis points toward 120. Revenue compounds at 21–23%, margins reach 18–20%, and at 20–22 times EBITDA that supports roughly $72–88.

The base case has sat in the $45–57 range across three articles while the stock traded $37, then $26, then $24, and now $32. The destination has not moved. What changed is that the upside now has a mechanism attached rather than an aspiration.

What Would Change Our Mind

Grow above roughly $50 million in ARR by mid-2027, without service headcount rising at the same rate. A second agent shipping inside four quarters. Disclosure of paid accounts, attach rate, or gross margin, usage anecdotes have stopped being sufficient.

Recurring gross profit growth staying several points above GPV growth, with the spread coming from subscription adoption and paid services rather than pricing or credit.

GPV per location positive by the fourth quarter, once the World Cup benefit washes out. Cash conversion back toward 70% of EBITDA in the second half. Toast Capital credit losses below roughly $30 million a quarter.

And the core holding above 20% growth at 40% margins, because that is what funds everything else.

Our view has not changed. Toast is a vertically integrated restaurant operating system with payments as the monetization rail, and workflow ownership is what drives distribution, attachment, retention, and data context. Enterprise, international, retail, Toast Local, and agentic products are extensions of that architecture, not quarterly reinventions of the company. Q2 strengthened the core thesis and produced the first measurable evidence that the same position can be used to perform work restaurants currently outsource, which is not a new definition of Toast, but a larger answer to the question we have been asking from the start: how much can be built on top of a workflow Toast already owns?

The register was the wedge, and the operating system is the moat. The opportunity now is making that operating system worth more to every restaurant that joins it.

Subscribe now

• The transaction is getting richer: Revenue grew 8% against room-night growth of 5.3%, extending an eight-quarter pattern in which monetisation has outpaced volume as merchant mix has risen to 73%.

• The acquisition treadmill remains: Marketing expense grew faster than revenue, marketing intensity increased to 4.7% of gross bookings, and direct mix remained stuck in the mid-60%s despite greater Genius and app engagement.

• The decisive test is the second booking: Booking has shown it can extract more value from the trip it executes today. The moat will be proven only when that trip lowers the cost of acquiring the customer tomorrow.

From Booking Holdings’ Q2 earnings call:

Room nights, gross bookings, revenue, and adjusted EBITDA all exceeded the high end of our guidance for the second quarter... Overall, if I also may point you to our full-year guidance, actually, if you look at our full-year guidance, we are at a high single-digit level for gross bookings and revenues and at mid-teens level for EPS at the high end. That is still in line with our long-term algorithm and is also in line at a constant currency basis with our original guidance for the year, even though in our assumptions we have seven out of twelve-month impact of the Middle East.

That last sentence from CFO Ewout Steenbergen is the quarter. Booking absorbed seven months of a regional war, a war that closed the Strait of Hormuz, grounded flights across the Gulf, and pushed airfares up across every long-haul corridor, and held its original full-year guidance at constant currency. Not a trimmed guide with an asterisk. The original algorithm.

The numbers underneath: room nights grew 5.3% to 325 million, gross bookings rose 9% to $51 billion, revenue grew 8% to $7.35 billion, adjusted EBITDA grew 9% with margins expanding 40 basis points to 36%. Every metric above the high end of guidance.

And yet the composition is what caught my attention. Room nights missed the street, consensus was around 329 million, while revenue beat it by $167 million. Fewer nights, more money. That gap, roughly three percentage points between revenue growth and volume growth, is where this quarter’s real argument lives.

In my Q1 piece I set a threshold I refused to explain away: if room nights missed even the lowered 2-4% guide, the thesis needed revision rather than another external excuse. They came in at 5.3%. The threshold held. I’ll also note, because the series deserves the honesty, that I’ve overreached before, the “fintech disguised as travel” framing, the Amazon analogy, the suggestion that Booking’s payment rails could escape travel entirely. I’m retiring that. What survived the pruning is narrower and better: Booking is moving from discovery, which AI is commoditising, toward execution, payments, settlement, refunds, service, coordination, which it is not. The question was always whether the execution layer captures its own economics. Q2 gave evidence on both sides.

The Spread

Start with the side that worked.

The three-point spread between revenue growth and room night growth needs careful handling, because I’ve learned the cost of assigning too much causality to one quarter. Walk the bridge: constant-currency ADRs rose about 2%, driven by genuine pricing in Europe and the US, management was explicit this wasn’t a mix artifact. FX added roughly a point. Flights and other verticals contributed. Revenue timing from March’s elevated cancellations added noise. Those factors get you close to the full spread without invoking merchant economics at all.

So, the single quarter proves less than I’d like. The pattern is what matters:

Eight consecutive quarters of positive spread, widening as merchant mix climbs from 66% to 73%. Merchant revenue grew 15% this quarter while agency revenue declined 7%. Management confirmed again that incremental payment revenues exceeded incremental payment costs over the trailing twelve months, a statement that has appeared before, but one that survived another quarter at higher mix while sales and other expenses still showed cost efficiency despite the heavier merchant load. The payment economics are confirming, not arriving.

If the pattern holds, room night growth systematically understates Booking’s earnings trajectory, and a market that values this stock on volume is using the wrong gauge. That’s the bull case, and the table is the best evidence for it that exists. But a pattern is not yet proof of permanence. Two quarters of the spread collapsing toward zero and this argument weakens considerably.

The Treadmill

Now the side that didn’t work.

If executing more of the trip, payments, service, loyalty, bundling, makes each transaction richer, the next test is whether it makes the next customer cheaper to acquire. That’s the difference between a better margin and a genuine moat. And Q2 did not show it.

Marketing expense grew 11% against revenue of 8%. Marketing as a percentage of gross bookings rose to 4.7% from 4.6%. B2C direct mix held in the mid-60%s, stable, not improving. SEO declined, and when Mizuho’s Lloyd Walmsley pushed on whether that reflected AI search or something else, Fogel answered more directly than I expected:

I do believe that some of the changes that were made in the display at Google definitely put some pressure on SEO... Our direct, mid-60s%. Hasn’t gone down at all. In fact, total number, absolute number, obviously, had to go up to maintain the same percentage share in the direct, because obviously we grew.

The deflection is fair as far as it goes. Both direct and paid traffic grew in absolute terms; the ratio stayed flat because both sides rose. But a ratio can stay flat on a treadmill. Booking is spending more on paid channels to hold direct mix steady while organic discovery erodes underneath it. That is not what a flywheel looks like.

I want to be careful here too, because one quarter of marketing deleverage doesn’t invalidate customer loyalty. Booking is still investing in the US and Asia. Connected Trip is a low-double-digit share of transactions, large enough to matter directionally, too small to move the consolidated marketing line. Paid spend can rationally grow during geographic expansion. But the strategic metrics have reached enough scale that some evidence should be showing up in acquisition economics. Genius Level 2 and 3 members generate a high-50% share of room nights, up from mid-50%s. Mobile app mix is in the high-50%s, up from mid-50%s. These loyalty layers are supposed to convert engagement into cheaper demand. Q2 showed the engagement. It did not show the conversion.

The Second Booking

Steenbergen gave the behavioural disclosure the whole thesis rests on:

Travelers who book with us across multiple travel verticals return more frequently.

That’s the mechanism, one trip creates context, trust and service history, and the next booking should come back more directly and more cheaply. The mechanism is visible. Its economics are not. In my Q4 piece I called the real AI risk toll extraction rather than replacement: Booking stays essential while paying rising rent to whichever interface controls customer intent. Google collects that rent today through $8 billion-plus of annual marketing spend. Q2’s evidence on the interface itself was actually encouraging, LLM referral traffic remains below 1% of room nights and flat, Google launched agentic booking with Booking as a first partner, OpenAI went CPC, which Fogel called “right in our wheelhouse.” The AI layer is fragmenting, not concentrating. But the rent did not fall. Both things were true in the same quarter, and the tension between them is now the entire investment debate.

AI’s clearest contribution so far is on the other side of the P&L anyway. Customer service cost per booking is falling at a double-digit rate. Voice AI handles most eligible inbound calls. Steenbergen disclosed internal metrics, AI cost per merge request, declining, that tell you this management team treats AI spending the way it treats marketing spending: measured, ROI-gated, unsentimental. AI is making the execution layer cheaper to run before it makes demand cheaper to acquire. That’s worth something. It isn’t the moat.

So, the question I’ve been circling for four articles has narrowed to one sentence: can Booking turn the trip it executes today into the customer it owns tomorrow? The merchant model makes each trip richer; the table shows it. Genius, the app and Connected Trip give the traveller reasons to return, Steenbergen’s disclosure confirms it. But until the return shows up as lower acquisition cost, the moat is behavioural rather than financial, and the market won’t pay a premium multiple for behaviour.

Three Years Out

From roughly $195:

The bull case requires no new category label; I’ve spent enough of this series asking the market to rename the company. It requires direct mix rising above 70%, Connected Trip reaching 20-25% penetration, and marketing falling toward 4.0-4.3% of gross bookings. Operational outcomes, not a reclassification event. The bear case, the toll road, marketing intensity rising, margins flat, is roughly what’s already priced in.

Management is making its own bet on which scenario plays out. Booking returned a record $4.1 billion to shareholders in Q2 and repurchased $7.4 billion of stock in the first half at an average price of $173, a pace that retires roughly 6% of the float per year. When a company with a $10 billion free cash flow engine buys back its own equity that aggressively at these levels, it is telling you where it thinks intrinsic value sits.

I own it here, and I’d add meaningfully toward $175, near the company’s own buyback average, absent new evidence of thesis deterioration. Sized for patience, not certainty: two of three scenarios pay you, and the one that doesn’t cost almost nothing from here.

What Settles the Argument

The Q1 threshold was met, so here is the new one, and I will not attribute a breach to conflict, channel mix, or geographic investment: if marketing exceeds 5.0% of gross bookings in any of the next four quarters, the switching-cost thesis at the core of my bull case needs revision. Below 4.5% in any quarter, and the conflict-distortion explanation is confirmed and the flywheel is turning. Between those bounds, watch B2C direct mix, above the mid-60%s means repeat behaviour is converting into cheaper demand, and watch for the number nobody has disclosed: what it costs Booking to acquire a customer through an AI agent versus Google Search. Every quarter of silence on that number extends the uncertainty discount.

Booking has shown it can earn more from the trip in front of it. The moat will be proven only when that trip lowers the cost of winning the next one.

That proof will not come from AI demos or partnership announcements. It will come from the income statement.

It always does.

Subscribe now

• The proof: Premium subscribers reached 300 million, ARPU rose 7.4% in constant currency, and gross margin hit a record 33.4%, challenging the view that royalty costs permanently cap Spotify’s economics.

• The bet: Management is deliberately trading some low-quality user growth for higher conversion and monetisation, introducing more friction into the Free tier and prioritising revenue over raw MAU expansion.

• The next layer: Audiobooks+ has surpassed $100 million in annual recurring revenue, offering an early glimpse of a metered model in which Spotify charges its most engaged users for additional hours, credits, and AI-powered experiences.

Three months ago, I argued that the market wanted financial proof the Taste Graph was an engine, not a thesis. Q2 supplied some of that proof: 300 million Premium subscribers, constant-currency ARPU growth of 7.4%, a record 33.4% consolidated gross margin, and €797 million in free cash flow. Yet the most important moment in the quarter was not a number. It was a decision, one that changes how I think about the next phase of Spotify’s business.

The Old Ceiling and How It Broke

For most of its existence, Spotify had a problem that scale could not fix. Music royalties are generally calculated as the greater of a percentage of revenue and a per-user amount. As Spotify grew, content costs grew in rough proportion. Gross margin sat around 25-27% for years. The company could add subscribers and revenue, but it could not add margin, because the content cost structure moved in step with the top line. That was the bear case from the IPO through 2023: a business that could get bigger but never meaningfully richer.

The margin trajectory tells you when something changed: 25.3% in 2022, 26.7% in 2023, 31.1% in 2024, and now 33.4%. Q2’s Premium segment was even sharper, gross margin of 34.9%, up 174 basis points, as Premium revenue grew 16% on a constant-currency basis while associated content costs grew only 12%. The filing attributes the improvement to revenue growth outpacing music royalty costs net of marketplace programmes, audiobook licensing costs, and Spotify Partner Program costs. That is all Spotify disclosed about the mechanics, and it is enough: the old claim that scale could never improve Spotify’s economics is broken in the data.

The Taste Graph matters here because it is the asset that makes price increases absorbable. Spotify raised prices across dozens of markets and subscriber growth remained healthy, 7 million net additions against guidance of 6 million, with management reporting no visible damage to intake. If Spotify were a commodity distributor, a 7% ARPU increase should have produced meaningful churn. It did not. The Large Taste Model, AI DJ at roughly 100 million subscribers, Prompted Playlist, years of accumulated listening history, these create a switching cost that is invisible in any single quarter and powerful across many. The subscription proved it can escape the old economic ceiling. The question is what Spotify builds on top of it.

The Decision

Alex Norström, on the Q2 earnings call:

“We’ve been making changes like tweaking the sign-up to get a higher quality MAU throughput. We’ve deprecated a lower-end Android device support which builds the business and makes it more efficient for us. We’ve carefully introduced some friction in both ad load and some limitations in our Free tier. Gustav and I have this saying where we say sometimes we pull the growth lever, and sometimes we pull the monetization lever. Here we’re starting to pull the monetization lever.”

Spotify guided Q3 MAUs to 788 million, up 11 million from 777 million, a meaningful step down from the 16 million added in Q2 and the 15-20 million quarterly pace of recent years. The company is not shrinking. It is choosing fewer marginal users in selected emerging markets in the belief that the remaining funnel will generate more value. Norström pointed to Latin America as the precedent, where Spotify followed the same progression from high MAU growth and low conversion to a more monetisable base over several years.

I think this framing is credible. At 777 million users, raw scale is less scarce than monetisable intent, and the conversion gap in Rest of World, which represents a large share of MAUs but a much smaller share of subscribers, is precisely the kind of opportunity that rewards tighter funnels rather than wider ones.

But I also think it is important to name what this is. It is a wager, not a proof. The proof comes later: does conversion improve, does churn hold, does advertising revenue per Free user rise, does the emerging-market base generate more revenue despite slower intake? Management has changed the scoreboard before showing the new score. That does not make the decision wrong. It means the evidence that validates the decision has not yet arrived.

The Meter

The smaller number in Q2 that I think matters more than most analysts will credit: Audiobooks+ passed $100 million in annual recurring revenue.

Norström described Spotify’s usage curve in a way that connected this directly to the Taste Graph:

“At the head of that curve are many millions of people who simply want more, and they’re willing to pay for it.”

Premium includes an audiobook allowance. Heavier listeners pay for more hours. Personal Podcasts follows the same logic, with monthly credits included in Premium and the option to buy more. Gustav Söderström made the connection to AI inference explicitly: generated experiences carry a real marginal cost, so unlimited use at a flat price would separate cost from revenue at the exact moment Spotify needs them connected.

This is the business model that may sit on top of the Taste Graph. The Taste Graph identifies what each user values. The usage-based layer, the meter, lets Spotify charge its most intensive users more, rather than raising the flat subscription price for everyone equally. That creates a different kind of revenue growth than conventional price increases: it is driven by engagement intensity rather than across-the-board repricing, and it scales with the depth of the personalisation engine rather than with management’s appetite for pricing risk.

The loop this creates is stronger than better recommendations alone. More users generate more data. Better data improves discovery and generated products. Better products reveal which users sit at the head of the usage curve. Those users buy credits, hours, or add-ons, and the revenue funds the next product. At $100 million of ARR against nearly €4.8 billion of quarterly revenue, Audiobooks+ proves the architecture works. It does not yet prove materiality. The meter needs to become a business rather than an elegant explanation.

What Remains Unproved

Advertising has stabilised but not inflected. Ad-supported revenue grew 3% in constant currency, reversing the decline from Q1. Automated channels reached nearly 40% of ad revenue, up from roughly 30% last quarter, and active advertisers grew 60% to 33,000. CFO Christian Luiga, whose guidance credibility on gross margin has been near-perfect for two years, committed to double-digit ad growth in the second half. The system is built. The excuse period should end. If double-digit growth arrives in Q3 or Q4, the market has to reprice a segment it had written off. If it does not, Luiga’s credibility on this specific promise takes a hit.

One risk I have not previously discussed sits on page 25 of the 6-K. The Mechanical Licensing Collective has filed an amended complaint and has sought permission for an interlocutory appeal of the January 2025 court ruling that classified Spotify’s Premium Service as a qualifying bundle for royalty purposes. If the MLC were to prevail in full, Spotify estimates a potential liability of approximately €473 million for March 2024 through June 2026, plus possible penalties. More important than the one-time charge, an adverse outcome could raise the forward royalty rate on the music component of Premium, which would directly threaten the improved content economics that Q2 demonstrated. Management did not discuss this on the call. I am not treating it as a base case, but it is the structural tail risk to the margin story and it deserves attention.

The €200 million of incremental marketing and AI spending in 2026 has not yet been tied to quantified returns. Management calls it temporary, expects Q4 moderation, and has kept headcount flat at 7,302. Söderström’s claim that “our margin is a managed outcome” is believable given the evidence. It is not a permanent pass. Q4 needs to show that “controlled” eventually becomes “productive.”

Three Years From Now

The bear case assumes add-on tiers remain small, advertising stays weak, and AI costs absorb pricing gains, or the MLC ruling resets royalty economics. The base case assumes consensus delivery, several hundred million euros of add-on revenue, and margins approaching management’s long-term targets. The bull case requires usage-based products to become a material revenue category while Spotify retains enough economics from rights holders to exceed a 20% operating margin.

The tracking items: Premium net additions holding near 5 million per quarter despite emerging-market friction. Ad-supported revenue reaching double-digit constant-currency growth by Q4. Gross margin returning above 33% after Q3’s regulatory timing charge. Q4 operating expense growth moderating below 10% excluding currency and social charges. Paid add-on revenue moving well beyond today’s $100 million ARR. And the MLC appeal, any adverse procedural development would change the margin calculus directly.

Q2 proved that the core subscription has escaped its old economic ceiling. The monetisation decision and the meter are management’s wager that the same understanding of taste that improved retention can now support a richer, more layered revenue model. The next phase of this series is no longer about whether the moat exists. It is about whether the moat can bear weight.

Subscribe now

• The mechanism is now visible in the accounts. In twelve quarters, revenue growth rose from 17% to 93%, the sequential addition from $50 million to $303 million, and net dollar retention from 107% to 157%. Revenue is outrunning customer count while RPO, billings, and margins rise with it.

• The Ontology is an institutional compiler. Foundation models supply generic intelligence; the Ontology translates a firm’s objects, permissions, workflows, and judgment into executable instructions. Better models make more workflows possible. The compiler makes them cumulative.

• The race is depth against distribution. Palantir has the deepest deployed model of high-consequence operations; incumbents have far greater distribution. My end-2029 outcomes are approximately $72, $190, and $375.

A compiler is invisible when it works. A programmer writes in human-readable language; the machine requires exact instructions. The compiler resolves definitions, links libraries, enforces rules, and turns intention into execution. The tenth program reuses what the first nine compiled.

Large organisations face the same problem. Their operating knowledge is scattered across software, employees, policies, permissions, exceptions, and accumulated judgment. A foundation model may understand supply chains in general; it does not know which substitution this company permits, who may approve it, what SAP must record, or whether the decision worked.

The model understands the world. It does not understand the firm.

Palantir calls the translation layer between the two the Ontology. It maps an institution’s objects and relationships, connects them to live systems, encodes who may act, and records what follows when an action is taken. I think a better description is an institutional compiler: Palantir turns the messy logic of an organisation into an executable model that employees, applications, and AI agents can use.

The big fundamental question is: can each compiled workflow increase the speed, scope, and value of the next before Microsoft, ServiceNow, AWS, Google, and SAP compile enough of the firm through systems they already distribute?

Q2 matters because the accounts now look like the compiler has begun to compound.

From Coordination to Compounding

Our view of Palantir has changed in stages. In October, I described the Ontology as the modern equivalent of MS-DOS: a common language across incompatible enterprise systems. After Q4, the argument became stronger. Once an institution had mapped its objects, permissions, and operating relationships into Palantir, replacement was no longer a normal software switch; it required reconstructing the institution’s operating model.

Q1 moved the thesis from coordination to absorption. Workflows that once lived inside individual applications could increasingly move into the Ontology, while the applications became interfaces. Our Q1 article also applied the Jevons idea to enterprise AI: as inference prices fall, firms use more intelligence, which creates more agents, tool calls, and governed actions. The force commoditising models increases demand for the layer controlling what those models may see and do.

What I missed was the effect on Palantir’s own growth curve. I understood why cheap models should create more demand. I had not fully understood why every successful workflow might reduce the difficulty of the next one, allowing Palantir to accelerate even as its base became larger.

The Ontology does not merely govern more work. It preserves what was learned while that work was being compiled.

The Acceleration That Should Not Be Possible

Palantir’s year-over-year growth has accelerated for eleven consecutive quarters. The percentage progression is striking; the absolute additions are harder to explain away.

Source: Palantir quarterly earnings materials; calculations.

Palantir added roughly $326 million to its quarterly revenue base in the year ending Q2 2025. It added about $932 million during the following year. Q2’s sequential addition alone was larger than the entire US commercial business one year earlier.

A temporary AI spending surge could produce a few strong quarters. It would be less likely to produce accelerating percentage growth, larger dollar additions, higher retention, expanding contracted demand, and wider margins at the same time. Q2 revenue grew 93% to $1.935 billion; US commercial grew 149% to $764 million; US government grew 90% to $809 million. Adjusted operating margin reached 62%, and adjusted free cash flow was $1.22 billion at a 63% margin.

The most revealing comparison is revenue against customer count. Total customers rose 24% over the year while revenue rose 93%. US commercial customers increased 35%, from 485 to 653, while US commercial revenue rose 149%. Sequentially, customer count grew 6% and revenue grew 28%. The system is going deeper much faster than the customer base is becoming broader.

Net dollar retention confirms that interpretation. It was 107% in Q3 2023 and reached 157% in Q2 2026. Existing cohorts are not merely renewing; they are becoming much larger relationships.

The forward indicators trace the same progression:

Source: Palantir quarterly presentations. Figures may not sum perfectly because of rounding.

Total RPO has increased almost fivefold; long-term RPO more than sixfold. Late 2025 brought a surge in multi-year commitments. Since then, short-term RPO has accelerated toward recognition while the long-term bucket keeps refilling. Billings have nearly quadrupled. Palantir is converting backlog without emptying it.

RPO also understates part of the government business because contracts with initial terms of twelve months or less, and obligations outside termination-for-convenience clauses, are excluded. Management says the reported balance is therefore primarily commercial.

The cost structure completes the argument. From Q2 2025 to Q2 2026, Palantir added about $932 million of revenue and $730 million of adjusted operating income, an incremental adjusted margin near 78%. Revenue grew 93%; adjusted expenses grew 37%. Those are not the economics of a business that repeats the same amount of labour for every new workflow. They imply reuse.

How the Compiler Compounds

Every enterprise AI system contains two models. The first is a foundation model—GPT, Claude, Gemini, Llama, Nemotron—trained on broad information and supplying generic intelligence. The second is a model of the institution: its objects, relationships, permissions, policies, actions, decision history, benchmarks, and outcomes.

The two assets age differently. A foundation model begins losing relative value when a competitor releases a better or cheaper one. The model of the firm becomes richer each time it is used. The first workflow requires Palantir’s Forward Deployed Engineers to connect systems, define objects, encode policy, and establish safe actions. The second workflow inherits much of that work. The tenth inherits the first nine.

This is where the collapse in inference costs matters. Our Q1 article noted that inference prices had fallen roughly a thousand-fold in three years. That did not reduce AI consumption; it expanded the number of tasks that could be run economically. Each new task still needs institutional context, permissioning, auditability, and a way to act.

Three mechanisms reinforce one another. Cheaper models make more workflows viable. Each workflow enriches the customer’s Ontology, making adjacent work faster to compile. Each engagement adds connectors, industry patterns, and field knowledge that Palantir can reuse elsewhere.

Shyam Sankar described the result as an automated model factory inside the customer’s security boundary. Operational telemetry feeds customer-specific benchmarks and post-training, while the customer controls the resulting weights. He also said that a standard open model beat frontier models on five production tasks once Palantir tested it against the customer’s actual work. The important asset was not the model ranking; it was the institution’s own definition of success.

That is why “institutional compiler” is more precise than “control plane.” Palantir is not merely routing agents or governing tool calls. It is translating tacit institutional knowledge into a living system through which agents can reason, act, and learn. Forward Deployed Engineers are performing specification work. The Ontology turns that work into a reusable asset.

The old consulting critique is therefore dead, but a new risk replaces it. If every new institution still requires proportional FDE effort, Palantir has a human-capital ceiling. If templates, tools, partners, and prior deployments allow each engineer to compile more operating surface area, the ceiling rises. Q2’s incremental margins suggest that it is rising; the company does not disclose workflow-level economics, so this remains an inference rather than a settled fact.

The foundation model makes a workflow possible. The institutional compiler makes it cumulative.

The Other Compilers

Palantir is no longer alone in understanding where value is moving. Microsoft, ServiceNow, AWS, and Google now describe products that govern agents, connect enterprise context, and translate intelligence into action. The strategic question is not whether these products exist. It is whose representation becomes authoritative when an agent must decide what an object means, which rule applies, and what action is allowed.

Microsoft starts with the graph of work: identity in Entra, documents and communication in Microsoft 365, code in GitHub, data in Fabric, and processes in Dynamics and Power Platform. Fabric IQ Ontology, still in preview, defines entities, relationships, and rules for agents and workflows. Microsoft has less proven operational depth, but the broadest ingredients and cheapest route into existing customers.

ServiceNow is the closest conceptual rival. It begins with the history of services, incidents, approvals, assets, cases, and business rules. Context Engine maps people, assets, policies, and operational history; ServiceNow says that intelligence “compounds with every workflow.” AI Control Tower governs agents across vendors, while Action Fabric exposes workflows to third-party agents. Its thesis now sounds like Palantir’s: “Intelligence will keep getting cheaper. Trusted execution will keep getting more valuable.”

AWS and Google are strongest one layer below. AgentCore lets customers use any model or framework while handling runtime, tool access, identity, tracing, evaluation, and security. Google’s Agent Platform offers runtime, registry, identity, gateways, policies, and observability for agent interactions. These products complement Palantir by making agents easier to run, but they also let technically capable customers assemble a good-enough compiler themselves.

SAP remains different. It already contains executable transactional logic for finance, procurement, supply chains, and manufacturing. Any cross-system compiler must respect that authority rather than wish it away; Palantir’s own prior examples treated SAP as a source system to connect and, at times, a workflow to simplify.

The contest can be reduced to one line: Palantir’s depth is racing the incumbents’ distribution.

The concept of an ontology is not the moat; everyone now understands it. Palantir’s possible moat is the installed system: a live model already connected to source systems, trusted to execute consequential actions, and accumulating knowledge faster than a rival can reconstruct it. Microsoft and ServiceNow do not need to beat that depth everywhere. They need to make a separate Palantir purchase unnecessary for enough workflows.

Sovereignty Cuts Both Ways

Management’s “AI sovereignty” argument is a claim about ownership of the second model. If a model provider captures a customer’s prompts, reasoning traces, benchmarks, and operating outcomes, the provider’s asset compounds while the customer rents the result. Palantir says the institution should own that learning loop. Its Q2 products extend this idea into model switching, post-training, agent orchestration, and customer-controlled weights.

There is an obvious contradiction: a company may reduce dependence on model laboratories by becoming deeply dependent on Palantir. The more operating logic compiled into the Ontology, the harder it becomes to leave. Palantir’s best answer is architectural rather than rhetorical—make the compiler open above and difficult to reproduce below. Let Microsoft, ServiceNow, AWS, Google, and customer-built agents use the Ontology; keep the advantage in constructing and maintaining the live institutional model.

Sovereignty also cuts geographically. US revenue grew 115% and now represents more than 81% of Palantir’s revenue. International commercial revenue grew 26% and only 2% sequentially. American companies may see a US vendor with defence-grade controls as an advantage; foreign institutions may see dependence on that same vendor as a sovereignty risk.

I would therefore treat international as upside, not as a required base-case input. The US opportunity is large enough to carry the next several years. A company valued as a universal institutional layer will eventually have to prove that the compiler can travel.

What Q2 Changed

Before Q2, I believed model commoditisation increased the value of governed action, the Ontology was absorbing workflows, and delivery capacity was the main constraint. I still believe all three. What changed is that increasing returns inside customers are no longer merely an architectural possibility. Revenue density, retention, contracted demand, billings, and margins now move together in the pattern the thesis predicted.

The new variant view is therefore about duration. Conventional software models assume each sale is another unit added to a larger base, so growth must eventually slow. Palantir’s relevant unit may be different: the amount of institutional operating surface compiled into the Ontology. Each new workflow can increase the value of the workflows already there.

The underappreciated risk also changed. It is no longer that Palantir is consulting in disguise. It is that Microsoft or ServiceNow becomes good enough for ordinary workflows before Palantir can industrialise its depth. The market may be underestimating both the duration of Palantir’s growth and the power of incumbent distribution.

Three Years From Now

I value the outcomes at year-end 2029 using FY2030 earnings, because that is the period public markets would then be pricing. The starting point is Palantir’s new FY2026 revenue guide of roughly $8.15 billion and a current reference price of $125.65.

In the bear case, incumbents capture ordinary workflows while Palantir keeps the hardest government and industrial settings. In the base case, delivery scales and Palantir remains authoritative in high-value cross-system work. In the bull case, third-party agents consume Palantir’s context rather than replace it, making the compiler a neutral layer across clouds and models.

The probability-weighted value is about $207, roughly 65% above the current price, but it is not a precise target. The bull case contributes heavily, and small changes in growth duration or terminal multiple move the answer sharply. The $9.2 billion cash balance supports the downside, but I do not add it separately to a P/E-derived value because part of its benefit already appears through interest income and future capital allocation.

Palantir has proved the margin structure. The harder question is how much of the firm it can compile, and for how long.

What Would Prove Us Wrong

The thesis should remain falsifiable. Revenue must keep growing far faster than customer count; net dollar retention should stay above 145%, with a fall below 135% signalling weaker customer compounding. Short-term RPO must convert while long-term RPO continues to refill, and billings should not trail revenue growth for several quarters. Adjusted operating margin should remain above 55%, with incremental margins above 60%, or the human-capital ceiling deserves another look.

US commercial growth below 60% before late 2027 would weaken the duration case. International commercial growth above 30–40% would expand the runway; sustained growth below 20% would confirm a geographically narrower opportunity. The competitive test is more qualitative but just as clear: are Microsoft and ServiceNow agents using Palantir as the authoritative source of institutional context, or is Palantir becoming one tool inside their systems?

Compilers usually disappear into the systems built on top of them. Users do not think about the definitions, libraries, and dependencies being resolved beneath the surface; they notice that new work becomes easier because the hard translation was done before.

That is what Q2 may be showing. Model intelligence is getting cheaper. More work becomes automatable. Each Palantir workflow adds context that can be used again, while each deployment adds knowledge that can shorten the next one. Revenue accelerates, contracted demand refills, and marginal costs fall.

The model laboratories are teaching machines to understand the world. Palantir’s bet is that the greater prize lies in compiling each institution into something those machines can safely understand and act upon.

Q2 is the first quarter where that compiler looks capable of accelerating itself.

Subscribe now

• Lam cleared nearly every test we set far earlier than expected: gross margin reached 52%, NAND doubled to 23% of systems revenue, and CSBG rose to $2.47 billion.

• The structural thesis is now stronger: more three-dimensional and technically demanding chips require more etch, deposition, and service intensity per wafer—but the spending that expresses that complexity remains cyclical.

• The old bear case was China and margin compression. The new one is that customers are compressing several years of investment into a short window, creating a future capacity digestion cycle that Lam has not yet proved it can compound through.

In April, at $275, I argued that AI was making semiconductor manufacturing more Lam-shaped, increasing the etch and deposition work required per wafer, independent of how many wafers the industry processes. I set four signposts. Gross margin above 50%: proves the operational flywheel. NAND above 15% of systems revenue: proves the conversion pull-forward is real. CSBG above $2 billion: proves the installed-base compounding is structural. CY27 WFE above $155 billion: proves the multi-year cycle extends.

I expected those tests to take time.

Lam cleared almost all of them in one quarter. Gross margin hit 52%, a 20-year high. NAND hit 23% of systems. Doubled. CSBG jumped to $2.47 billion. WFE was raised to “the low $150 billion range” for 2026, with 2027 described as an “extraordinary setup.” Then Lam guided September revenue to $8.1 billion, $970 million above consensus, with a 39.5% operating margin.

Our thesis was right. Our sense of speed was wrong.

One correction. In January I modeled a 48–49% gross margin floor after China declined. The actual number is 52%, and management is targeting mid-50s. I was wrong by 300 basis points. The R&D maturity flywheel I described in April was real, but I underestimated both its magnitude and the pricing contribution that comes from solving the hardest manufacturing problems at the most advanced nodes.

And one prior debate that can be closed. China fell from 34% to 26% of revenue. Lam posted record margins and a massive guide. The old thesis, that declining China revenue would compress margins, has not merely weakened. It has been falsified.

The question in April was whether AI was making semiconductor manufacturing more Lam-shaped. June answered it. The better question now: can rising semiconductor complexity increase Lam’s content per wafer fast enough to outrun the capacity cycle that its own success is helping create?

The same problem, spreading

Lam’s served available market reached roughly 36–36.5% of WFE, ahead of the timeline set at the 2025 Investor Day. Doug Bettinger, when Mizuho’s Vijay Rakesh asked:

“We at that point were talking about our SAM expanding from the low 30% of WFE range into the high 30% range. As we sit here today, we’re probably trending already to that high level. We’re, I don’t know, I’d guess 36%, 36.5% this year.”

Low-30s to 36.5% in 18 months. That doesn’t mean Lam invented a new market. It means more semiconductor spending is flowing toward etch, deposition, and surface treatment where Lam is already strong. The same technical problem, building three-dimensional structures with tighter geometries, is spreading across device categories.

In NAND, the problem is depth. In January I called it the coiled spring. This quarter it fired. NAND went from 12% to 23% of systems revenue, more than doubling sequentially. Tim Archer framed it as structural AI demand:

“Expanding context windows and persistent memory requirements are driving significantly higher demand for flash storage.”

That completes a narrative shift SK Hynix started in January. If NAND spending is genuinely AI-driven, the historical cyclical discount investors apply is wrong.

But I want to note how Archer handled the pushback. When Citi’s Atif Malik raised TrendForce forecasts of NAND oversupply in 2027, Archer didn’t defend NAND directly. He widened the lens, “the Lam story is a lot bigger than just NAND.” A fully confident defense would have met the oversupply question head-on. He redirected instead. The structural reframing is still the right emphasis, but when management deflects rather than defends, I want to register that.

The air pocket risk from April, what fills the revenue stream when the $40 billion conversion completes?, has not been answered, only deferred. Whether conversion hands off to greenfield capacity or enterprise SSD demand sustains the build is what the next articles need to track.

The same physical constraint is appearing in DRAM, where advanced nodes now require technologies first proven in foundry logic. Lam’s Akara platform, adopted for GAA logic at 2nm, is winning conductor etch positions in DRAM, its installed base doubling annually. Advanced packaging is growing more than 70% year-over-year, with panel-level tools shipping across multiple geographies.

Different device. Different format. Same physical problem.

The 52% gross margin confirms the mechanism is financial, not just technical. The toll rate rises, more process steps per wafer, and the collection cost falls: more mature tools, closer-to-customer manufacturing, scale leverage. Pricing contributed, but management didn’t quantify the component, so I can’t cleanly separate it from mix and efficiency gains. The prior articles correctly identified customer concentration as a margin constraint, and TSMC, Samsung, and SK Hynix remain formidable procurement organizations.

Management raised long-term targets to mid-50s gross margins and mid-40s operating margins, a material upgrade from the 2025 Investor Day framework. Several quarters need to pass before declaring 52% the permanent floor rather than a favorable mix quarter.

The installed base earns its keep, and the balance sheet absorbs the strain

CSBG at $2.47 billion, up 43% year-over-year, is the compounding layer from April made visible at scale. Bettinger confirmed it’s “growing faster than the installed base.” That’s ARPU expansion. Equipment Intelligence and Dextro are in “the very early stages of a multi-year rollout” across 100,000+ chambers.

Here’s the tension. The Q4 upside came entirely from CSBG. Systems revenue of $4.25 billion missed consensus by $177 million. If complexity means more tools per wafer, systems should accelerate too.

And the ramp’s financial strain shows. DSO jumped from 64 to 72 days. Receivables grew 29% on 15% revenue growth. Free cash flow was 56% of adjusted net income. Buybacks slowed to $246 million from $796 million. Headcount rose roughly 1,800 in one quarter.

The systems miss is probably timing, the $8.1 billion September guide implies ~$5.6 billion in systems, a 32% sequential jump. Deferred revenue grew $213 million. Japan acceptances rose to $490 million. The tools are in the pipeline.

One quarter of this mismatch is timing. If systems miss again in September and DSO stays above 70, the quality of the growth deserves harder scrutiny.

Success creates the next bear case

The old bear cases have weakened. China-margin compression: falsified. NAND as management talk: $978 million says otherwise. The 50% margin needing China: disproved at 52% with China at 26%.

The new bear case is that a real structural advantage is being compressed into an unusually short spending window. Customers have announced fabs through 2027 and beyond. DRAM, NAND, foundry logic, and packaging are all rising together. That is excellent for orders. It also creates the conditions for excess capacity when those fabs produce simultaneously.

The right distinction, and the one I collapsed too easily in prior versions, is between complexity as a persistent technical direction and spending on complexity as a cyclical financial expression. Complexity may never peak. The rate at which customers spend to address it absolutely can.

What the market prices and what it doesn’t

At what point does our variant perception become consensus?

At $298, with analysts raising targets to $350–$450 and the structural narrative spreading, the qualitative thesis is no longer ours alone.

It’s also worth recognizing that management is actively selling the re-rating. Every disclosure choice, SAM quantified at 36.5%, advanced packaging upgraded to >70%, new profitability targets at peak performance, “compelling” escalated to “extraordinary”, serves one purpose: convince long-only capital that Lam deserves a compounder multiple. That doesn’t make it wrong. During an upcycle, authenticated structural shifts and peak-cycle management confidence produce identical signals. They separate when the cycle turns.

The narrative has shifted. The numbers have not. FY27 consensus at $31.1 billion and $8.20 EPS is difficult to sustain after a $8.1 billion quarterly guide. Estimates need to move materially higher. The revision cycle should provide a mechanical tailwind over the next 60–90 days.

The bear case is not catastrophe. It’s the possibility that the physical thesis proves correct while the financial expression proves cyclical, customers compress spending into FY27, capacity overshoots, and rising content per wafer does not fully offset the decline. Lam exits as a better business. Just not one that has escaped cyclical valuation.

Six signposts:

1. December revenue guidance. Above $8.3B = September was a step. Below $7.8B = timing concern.

2. Gross margin. Below 50% for two consecutive quarters challenges the thesis.

3. CSBG. Above $2.5B confirms the franchise. Below $2.1B during high utilization weakens it.

4. Cash conversion. DSO below 65 and FCF conversion above 80% by December = June was timing.

5. The NAND handoff. Conversion must bridge to greenfield or sustained enterprise SSD demand.

6. WFE after 2027. Spending must stay high enough that content and CSBG offset systems digestion.

The toll booth metaphor has served its purpose across four articles. What this quarter showed is that it compounds: the toll rate rises with complexity, the collection cost falls with scale, and every tool enlarges an installed base that generates rising revenue for years.

Management is selling the conclusion, Lam compounds through cycles, before the test arrives. The test arrives when the fabs being built today begin producing simultaneously and equipment orders decelerate. Until then, the structural narrative and the cyclical data will be indistinguishable.

The mechanism is real. I believe that more than I did in January, more than I did in April, and more than the evidence strictly requires me to. What I do not yet have is evidence that it survives a spending pause. That evidence cannot exist until the pause arrives.

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

Subscribe now

• AI has moved toward Amazon’s architecture. Agents require CPUs, storage, databases, security, and enterprise data—not just accelerators—turning AI into a multiplier on the rest of AWS.

• The Foundry is strengthening the Engine Room. Amazon’s custom silicon and AI services increasingly support retail, advertising, logistics, and internal automation, creating a reciprocal flywheel across the company.

• The remaining question is cash. AWS growth, margins, and backlog validate the investment, but each successful capacity expansion appears to justify an even larger one before shareholders see the free cash flow.

“We’ve long believed AWS could become a few hundred billion dollars revenue business and now believe it will be at least double that and very possibly be $1 trillion annual revenue business for us in time.”

— Andy Jassy

I have spent the past three quarters asking whether Amazon’s Engine Room could finance its Foundry.

In Q3, commerce, Prime, sellers, and advertising were generating enough cash to fund AWS, AI capacity, and custom silicon. In Q4, the Foundry stopped being one destination for Amazon’s surplus and became the organizing principle of capital allocation. Q1 proved that customers wanted what Amazon was building, but the balance sheet entered the story, and I ended with a deliberately hard conclusion: evidence was not yet cash.

Q2 makes the evidence overwhelming and the cash question harder.

AWS growth needed to remain above 27%; it reached 36.7%. Advertising needed to hold above 22%; it accelerated to 26%. Paid units needed to grow at least 12%; they rose 17%. Trainium needed customers beyond Amazon’s investment portfolio; OpenAI, Uber, Pinterest, and a growing group of AI companies appeared.

The two cash tests failed. Trailing free cash flow moved from positive $1.2 billion to negative $7.6 billion. Long-term debt reached $128.9 billion, close to the $150 billion level I had marked as a warning. Amazon then raised expected 2026 cash capital spending from $200 billion to $220 billion.

I was right about the demand. I was too conservative about how quickly it would become revenue and profit, and too optimistic about when those profits would become free cash flow.

That is enough reconciliation.

The deeper question is whether the nature of AI has changed in a way that makes Amazon’s old architecture once poorly suited to frontier-model training unusually well suited to inference and agents. If it has, AWS may have entered a cycle in which AI demand pulls the rest of the cloud behind it, cheaper internal silicon widens the profit pool, and better economics justify still more capacity.

The risk is no longer that the Foundry fails. The risk is that it works so well it never finishes being built.

The Workload Moved Toward Amazon

Three years ago, the case against AWS was persuasive.

The first generative-AI wave revolved around training giant models across thousands of tightly connected Nvidia accelerators. That job rewarded the fastest horizontal networking, full Nvidia systems, and clouds willing to reorganize their architecture around one synchronized cluster. AWS had spent years moving in a different direction: Nitro, its own networking, disaggregated resources, managed services, and custom chips built around the cost structure of general cloud computing.

Those choices made AWS weaker for frontier training. The criticism was right. Then AI changed shape.

The first wave generated answers to prompts. Reasoning models began generating many more tokens before producing an answer. Agents now trigger those reasoning loops repeatedly, retrieve context, call software, query databases, check permissions, save memory, and act again without waiting for another human instruction.

Each shift changes what the underlying system must do well.

Training wants one giant accelerator cluster. Production inference needs reliable serving near the model. Agents need CPUs for orchestration and tool calls, storage for memory, databases for context, identity and security for permissions, and access to the applications and proprietary data on which they are supposed to act.

That sounds much more like AWS.

Nitro was designed to separate and route different resources efficiently. Graviton was built to own the CPU cost curve. AWS’s broader advantage was that much of the enterprise data and software an agent needs already lived there.

AWS did not suddenly repair every weakness identified during the training era.

The workload moved toward the architecture Amazon had already spent a decade building.

Jassy described the commercial consequence:

“Growth in AI drives core because post-training, reinforcement learning, and agent tool use is mostly done on CPUs versus AI accelerators.”

This is a claim about the structure of demand, not a product boast. AI spending does not sit beside conventional cloud spending. Once a workload enters production, it pulls CPU cycles, storage, databases, networking, security, and observability behind it.

AI becomes a multiplier on the rest of AWS.

Microsoft independently described the same mechanism two days earlier. Satya Nadella said CPUs were “just as important as GPUs” for agents, while Microsoft highlighted rising demand for PostgreSQL, Fabric, Foundry, enterprise context, identity, and security. Amy Hood said additional CPU and GPU capacity was monetized almost immediately once engineering teams made it available.

Two competitors are describing the same system. That makes it harder to dismiss Jassy’s explanation as earnings-call framing.

Amazon may be better placed to collect on it because more applications and data already reside in AWS, and because Graviton gives it a credible, widely adopted CPU cost advantage.

The financial evidence is striking:

Q2’s $4.6 billion sequential increase was about 80% larger than AWS’s prior record. Growth accelerated for the fifth straight quarter. Backlog reached $496 billion, more than three times current annualized AWS revenue, although Amazon did not disclose the duration, customer concentration, or conversion schedule. Most 2027 capacity is reserved, and Jassy described demand already committed for 2028 as “striking.”

Then there was margin.

Reported AWS operating margin reached 39.4%, including a $600 million energy-contract gain. Strip that out and clean margin was still roughly 37.9%, slightly above Q1, while depreciation grew more than 30%.

Growth accelerated by nine points. Margin held.

At a $169 billion revenue run rate.

Part of that reflects high utilization, mature core services, software improvements, and cost discipline. Trainium cannot explain every basis point. It does, though, explain why margins like these may persist.

The lower layers of inference are becoming more commodity-like. Customers increasingly consume an outcome through Bedrock or another managed service rather than selecting a chip directly. When models can be swapped and the interface hides the hardware, cost structure matters more. A provider paying full third-party silicon margins sets one cost benchmark; Amazon can run a growing share of its services on Trainium and retain the difference.

That is the old Graviton playbook: sell the higher-value service and use cheaper internal silicon underneath.

Amazon’s AI and chips businesses have each passed $25 billion annual revenue run rates, both growing at triple-digit rates. Graviton is used by 98% of the largest EC2 customers. Trainium has multi-year, multi-gigawatt commitments from Anthropic and OpenAI. Jassy previously said Trainium could save tens of billions of dollars in annual capital spending and add several hundred basis points of margin at scale.

Q2 does not prove the whole claim. It makes it much easier to believe.

The Engine Room Learns from the Foundry

It would be a mistake to write this as an AWS-only quarter.

Online stores grew 15%, up from 9% in Q1. Third-party seller services grew 16%, up from 12%. Paid units increased 17%. North America and International revenue grew 16% and 15%, respectively. Advertising stepped from four quarters at 22% growth to 26%.

Retail’s mechanism is speed.

Amazon delivered more than 40% more items on the same day or overnight in the first half. Monthly active perishables customers rose more than 50% since January. Same-day orders containing perishables carried more than three times as many units. Pharmacy doubled new customers, and same-day prescription deliveries increased nearly fivefold.

A two-day service is useful for a planned purchase. Same-day delivery can become the default for milk, medicine, toiletries, and whatever the household forgot this morning.

Speed increases frequency. Frequency improves inventory and route density. Density lowers unit cost and supports wider selection and faster delivery.

Then it repeats.

There was one crack. Shipping costs grew 19%, two points faster than paid units. For three quarters I had treated units growing faster than shipping costs as evidence that the retail network was gaining operating strength. That spread reversed. Olsavsky blamed fuel inflation and higher linehaul rates, with an FBA surcharge offsetting part of the pressure.

I am willing to accept the explanation for one quarter. I will not ignore the metric.

North America’s reported 7.9% margin also included a $600 million tariff refund. Clean margin was closer to 7.3%, below Q1. Retail demand accelerated, but the cost advantage did not widen.

Advertising converts that demand into cash.

The business reached $19.8 billion in quarterly revenue, nearly an $80 billion annual run rate. Amazon does not disclose its profit, but the economics are attractive: the shopping traffic, purchase data, and customer relationship already exist. An additional sponsored recommendation does not require another warehouse or delivery route.

AI is creating more monetization surfaces. Shoppers who click a sponsored prompt convert 48% more often and spend 21% more. Advertisers using Ads Agent see lower costs per impression and acquisition. Prime Video adds attention outside the store; inventory across the NFL, NBA, WNBA, and NASCAR sold out.

There is a real risk. AI discovery may send more shoppers directly to brand sites and Shopify merchants, leaving Amazon with fulfillment while weakening the discovery moment where its most valuable ads sit. For now, agentic commerce looks more like a referral channel than a distributor that owns checkout. That gives Amazon time to build sponsored prompts into Rufus and Alexa without surrendering the customer relationship.

This is the update to the Engine Room and Foundry framework.

The Engine Room still supplies the cash. Retail creates traffic and transaction data. Advertising monetizes it. Prime deepens the relationship.

The Foundry is beginning to pay some of that investment back. Trainium runs Bedrock; Bedrock supports Rufus, Alexa, advertising tools, and internal automation; better shopping experiences drive conversion and purchase frequency; more commerce creates more advertising revenue; that revenue helps finance the next data center.

The relationship is now reciprocal.

Amazon Supply Chain Services shows the same corporate formula in the physical world. Amazon built logistics for itself, reached extraordinary scale as its first customer, then began selling the capability to P&G, 3M, Lands’ End, and American Eagle.

AWS was born this way. Logistics may be next.

The Market Paid for Proof

Amazon raised expected 2026 cash capital spending from $200 billion to $220 billion, largely because memory costs increased, and the shares rose.

The market does not hate capital spending. It hates spending it cannot underwrite.

Alphabet delivered extraordinary Google Cloud growth, margin, and backlog, yet a larger capex shock and negative quarterly free cash flow overwhelmed that evidence. Microsoft paired accelerating Azure with positive free cash flow and a clearer path to containing consolidated margin pressure. Amazon supplied the most explicit asset-level return math of the three.

Data centers begin consuming cash roughly two years before monetization, but the buildings last more than 30 years and can host several generations of servers. Servers are bought only months before deployment, when Amazon has much better demand visibility. They break even in slightly under three years, last five to six years, and most current AI capacity is contracted for at least five.

The math is attractive. It is not complete.

The sub-three-year break-even applies to servers and networking, not the fully burdened cost of land, buildings, power, financing, and unused capacity during construction. Amazon proved that equipment placed into a busy facility can earn strong returns. It has not shown when the total capital program generates cash for shareholders.

The peer comparison adds another wrinkle. Estimated annualized net-new cloud revenue in the quarter was roughly $19.0 billion for GCP, $18.6 billion for AWS, and $15.9 billion for Azure. The figures are estimates and contract timing differs, so I would not build an entire thesis around them. They do invert the lazy percentage-growth narrative: AWS is adding immense absolute dollars despite its larger base.

Microsoft owns the strongest enterprise distribution through Office, GitHub, identity, security, and Agent 365. Those products create Azure demand, though they also claim scarce compute for Microsoft’s own use. Google has the deepest integration among models, TPUs, Search, YouTube, advertising, and Cloud.

Amazon’s consumer businesses are more physical. AI tends to improve shopping, logistics, and advertising rather than directly replace them. That may give Amazon less defensive urgency when allocating capacity between internal products and external customers.

Less. Not none.

Amazon is building Rufus, Alexa, Quick, Kiro, its own models, and a deep Anthropic partnership. The neutrality advantage remains a hypothesis.

Management’s new disclosures were designed to make the capital easier to accept: $496 billion of backlog, $25 billion-plus AI and chips businesses, capacity reserved through 2028, five-year contracts, and explicit break-even math.

What Amazon left out is where the uncertainty now lives.

It did not disclose how much backlog belongs to Anthropic and OpenAI, how much the AI and chips numbers overlap, when the backlog converts, the split between short- and long-cycle capital, fully burdened returns, or the expected free-cash-flow inflection. Asked about funding, Olsavsky said Amazon had “a lot of options.”

A non-answer.

Anthropic is part asset, part tenant, and part concentration risk. The $53.4 billion pre-tax gain made Q2 net income meaningless for operating analysis. It also confirmed that Amazon owns a valuable financial asset whose value and AWS consumption are tied to the same relationship.

Useful. Complicated.

A Foundry That Never Finishes

The old bear case was that Amazon would build capacity before demand appeared. Q2 weakened it. The new risk is rational perpetual reinvestment.

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.

The Q3 headline guide is distorted because Prime Day moved from Q3 last year into Q2 this year, with nearly four points of calendar impact and another 80 basis points of currency pressure. AWS has no Prime Day.

I would underwrite a wide $2.5–4.0 billion sequential AWS addition, implying roughly 35–40% growth. The range is wide because Amazon gave strong demand commentary but no quarterly capacity schedule. Below 30% would reopen the question of whether Q2 was a concentrated capacity release. Clean margin below 35% would suggest that depreciation is arriving faster than the benefits from utilization and custom silicon.

For valuation, consolidated operating income is the cleaner primary anchor. A sum-of-the-parts view is useful because AWS, advertising, commerce, and Anthropic have very different economics, but it can count the same traffic, data, and AI investment twice.

At a post-print price around $258, my three-year ranges are:

The numbers are not the thesis. They show where the thesis matters.

AWS must keep adding at least $3 billion of quarterly revenue or explain the capacity timing. Clean AWS margin should remain above 35–36%. Advertising needs to hold above 20%. Shipping-cost growth must fall back below unit growth. Debt should remain near or below $150 billion at year-end. Above all, operating-cash-flow growth needs to begin outrunning capital-spending growth during 2027.

Q4 was the declaration. Q1 proved the demand. Q2 proved the operating economics.

The remaining proof is cash.

The Foundry works. The question is whether Amazon’s success keeps extending the buildout faster than shareholders can ever see it.

Subscribe now

• The advertising thesis is working. Meta grew ad revenue 27% despite daily users rising only 3%, as better recommendation and advertising models increased both inventory and pricing.

• Product return is no longer the same as shareholder return. Capital spending, depreciation, research costs and widening AI ambitions absorbed nearly all of Meta’s operating cash flow during the quarter.

• The key question is whether the lab remains subordinate to the advertiser. Meta does not need to win every frontier-model contest, but it must show that unproven projects face a real capital hurdle and that free cash flow eventually returns to shareholders.

A year ago, in The Infinity Machine Boots Up, I wrote that “the ROI isn’t coming; it’s here.” Meta had shown that better recommendation models could increase engagement, create more ad inventory and improve conversion at the same time. Revenue was accelerating, margins were expanding, and the company could reinvest the resulting profit into still more compute. I called it an infinity machine.

I still believe the core idea.

What I no longer believe is that the return from Meta’s advertising systems and the return from Meta’s entire AI budget can be treated as the same thing.

That distinction did not matter much in 2Q25. It matters now. Meta grew advertising revenue 27% while daily users increased only 3%. Family revenue per person rose about 24%; ad impressions increased 14%, and the average price per ad rose 12%. Yet reported operating margin fell from 43% to 31%, clean margin was closer to 37% after removing legal and severance charges, and $31.1 billion of quarterly capital spending left only $784 million of free cash flow. The product return is real. The shareholder return is no longer automatic.

I now think there are three separate questions. Does AI improve the product? Does that improvement exceed the operating cost of producing it? Does the cash left over accrue per share? In 2Q25, all three answers appeared to be yes. In 2Q26, the first remained emphatic, the second became more conditional, and the third was no—at least for the quarter.

I was right about Meta’s return on intelligence. I had not separated that return from Zuckerberg’s appetite to reinvest it.

The Closed Loop

The fundamental question for Meta is no longer whether AI helps the business. It plainly does.

The question is whether Meta can keep the loop closed: can more intelligence create more relevant content, more attention, better advertising and more profit, which then funds still better intelligence, or will each proof of return simply widen the company’s ambitions until the advertising machine becomes the funding source for an open-ended AI race?

The productive loop is unusually powerful. Better models decide what someone wants to see before that person knows to ask. Better content keeps the person engaged for longer. More engagement creates more commercial moments. Better ad models match those moments with businesses willing to pay, and the resulting conversion data improves the next decision.

Scale strengthens every step. Meta has 3.6 billion daily users, a vast supply of content, millions of businesses and an auction learning from countless interactions. A rival can build a capable model. Recreating all four sides of that system is much harder.

The quarter confirmed the mechanism. Meta’s audience barely grew, but the economic value of that audience jumped. The mix was healthy too: impressions supplied more inventory while higher pricing showed that the new inventory retained value. Worldwide impression growth did slow from 19% in the first quarter to 14%, so I would not extrapolate 27% ad growth forever. Still, this was not a company squeezing a mature user base. It was a system becoming more productive.

The Advertising Company

Last year I wrote that Meta was no longer an advertising company. That was too clever.

Meta is an advertising company. That is the bull case.

What changed is the meaning of advertising. Twenty years ago, a business bought space beside content. Today it gives Meta a budget and an objective; Meta can create the ad, identify the customer, choose the moment, predict the conversion and adjust the campaign. The business is buying an outcome rather than a slot.

Google monetises intent after a user expresses it. Amazon monetises intent near the point of purchase. Meta is strongest one step earlier: it discovers desire. Someone opens Instagram to be entertained and encounters a product they did not know existed. AI improves the content, detects the latent interest, generates the commercial message and finds the seller.

This is why I do not think Meta needs to own the best general-purpose model in every category. Anthropic may lead coding, OpenAI may own more consumer productivity, and Google may have the broadest model-and-distribution bundle. Meta can still earn the highest return from intelligence because it owns a direct route from prediction to commercial demand.

That is my variant view. The AI winner will not necessarily be the lab with the best benchmark. It may be the company with the most profitable use for a model that is good enough.

When Success Widens the Bet

The old thesis assumed a closed financial loop. There is now a second possibility. A self-funding loop can become a self-expanding one.

That does not make the spending foolish. Meta’s aggressive GPU purchases in 2022 looked excessive before they helped repair recommendations after Apple’s privacy changes and TikTok’s rise. The willingness to commit before the payoff is obvious is part of the company’s strength.

But “AI spending” now covers several different claims on capital. Recommendation, ranking and ad automation have earned their spending because the return is visible. Enough model and compute ownership to avoid dependency may be necessary even when the direct return is harder to measure. Frontier research, enterprise agents, external compute, smart glasses and the continuing Reality Labs effort remain less proven.

Meta does not tell us how the budget is divided. That is the gap.

The reported quarter looked worse than the underlying business: $2.4 billion of legal charges and $1.18 billion of severance reduced operating income, and excluding them produces a clean margin near 37%. The legal cost cannot be dismissed—youth-related litigation may recur—but it tells us little about the return on AI. Even on the cleaner number, margin was down roughly six points from a year earlier. Research expense rose 67%, stock compensation rose 58%, depreciation rose 46%, and Reality Labs lost another $4.6 billion. Operating cash flow increased 25%, yet Meta made no share repurchases and issued almost $25 billion of debt as capital spending absorbed nearly all of the cash generated by operations. One quarter of construction payments should not be annualised, but the direction is clear.

The core machine did not fail.

More claimants arrived.

A Price for Ambition

Zuckerberg said Meta has received a “large number of offers” for its compute at a “meaningful premium” to what the company paid. Most people heard the outline of a new cloud business. I heard something more useful: an observable opportunity cost.

An outside rental price tells Meta what an internal workload must beat. If a recommendation model can generate more advertising profit than the same capacity would earn from a customer, Meta should take the compute back. If an internal project cannot clear that bar, it should explain the strategic value of doing so or let someone else use the machines.

Ben Thompson made this point before the quarter, and it is the part of his argument worth keeping: compute rental can be a hurdle rather than a destination. Meta does not need another reason to build. It needs a price for deciding which ambition deserves what it has built.

This is not an argument for replacing founder conviction with a committee. Zuckerberg’s willingness to invest early has created enormous value. The better answer is a rule that lets successful bets win more capital and forces unsuccessful ones to lose access to it.

Meta need not prove that every AI project will work.

It needs to show that failure has a cost.

What I Believe Now

I am more confident that Meta owns one of the best commercial applications of AI in the world. I am also more confident that the company does not need to win every frontier contest to benefit from the technology.

I am less willing to award the wider AI programme the returns already visible in advertising. Compute rental can protect the downside, but returns on rented hardware—however attractive during scarcity—should not receive the same valuation as a self-reinforcing ad system. Agents, glasses and enterprise products may become large businesses; for now, they are claims, not evidence.

This is why the stock’s decline was both understandable and incomplete. The quarter was slightly below the buy-side revenue bar and third-quarter guidance implied slower growth. Yet clean operating income was better than the headline, and the capex midpoint barely moved. Meta managed to spend too much for shareholders focused on free cash flow and too little for semiconductor bulls seeking another upward capex surprise.

The deeper conclusion is simpler:

The quarter raised my estimate of Meta’s underlying earning power and lowered my confidence that shareholders will capture all of it.

Three Years From Now

The valuation turns on whether product return, company return and shareholder return reconnect by 2029.

In the bear case, advertising remains a good business but slows before spending does. New products do not become material, annual capital spending remains near $200 billion, and Meta is valued as a capital-heavy advertiser.

In the base case, advertising compounds in the mid-teens, clean margins recover toward 40%, spending intensity peaks and free cash flow returns. Meta does not need another business of comparable size; it needs the lab to remain subordinate to the economics of the core.

The bull case requires a larger expansion of content and commercial inventory, plus one new profit pool—agents, messaging or glasses—becoming material. Internal uses of compute must earn far more than external rental, and capital spending must stop growing faster than revenue.

At roughly $538 after the print, the base case implies an annualised return of about 18–26% over three years. The bear case is roughly flat to down; the bull case requires evidence I am not yet willing to capitalise.

What Would Change My Mind

I will track five things. Advertising growth above 20% through early 2027 would confirm that the productive loop remains strong; below 15% while spending stays elevated would weaken it. Clean operating margin needs to stabilise above 37–38%, rather than settling below 35%. Quarterly free cash flow should recover above $10 billion before 2028. Meta needs to disclose enough about compute sales or internal returns to show that capacity has a real hurdle. And buybacks should eventually resume at a level that offsets stock compensation.

A year ago, I said the ROI had arrived.

It had.

What I missed was that the arrival of a return does not settle who gets to use it. Meta has proved that it knows how to turn intelligence into revenue. It now has to prove that it knows which ambitions deserve that revenue.

That is the cost of being right.

Subscribe now

• Microsoft’s AI distribution is working, but product leadership remains unresolved. Copilot seats, Agent 365 registrations and GitHub revenue are rising rapidly, even as the most important agentic products continue to emerge outside Microsoft.

• The investment thesis has shifted from owning the AI operating system to controlling the right-of-way. Azure provides the track; model companies provide the cargo; Agent 365, Entra and Purview could become the governance layer every enterprise agent must pass through.

• At 19× forward earnings, the stock appears priced for the utility outcome. The upside depends on whether Microsoft can govern third-party agents at scale and turn enterprise distribution into an exclusive ecosystem chokepoint.

Satya Nadella, on the Q4 earnings call, responding to a question about competitive dynamics:

All of us are reading this, “1873” is the book to be read.

Amy Hood, moments later:

You’ve got to run an efficient railroad.

Three months ago, I ended the Q3 essay with a line I meant as a commitment: “The next four quarters will tell us which one it is”, whether Microsoft owns the AI operating system or operates as middleware beneath it. This is quarter one. I owe the reader a score.

Agent 365 ubiquity. Forty million agents registered across tens of thousands of companies in roughly two months. Fast. But “registered” is undefined, a registered agent could be active and governed, or auto enrolled through an E7 bundle and sitting dormant. We don’t know how many enterprise agents run outside the control plane. This is adoption. Ubiquity is a different test. Open.

Copilot to 50 million seats. Went from 15 million to 20 million to 30 million-plus, with net adds more than doubling quarter over quarter. NHS England bought 505,000 seats. EY bought 400,000 on the new E7 SKU. Ahead of pace. But a seat procured by a CIO is not a seat chosen by a user when Claude, ChatGPT and Copilot are all available. And a seat chosen by a user is not a company that competes better because it deployed the product. We have evidence at the procurement level. We lack evidence at the competitive-performance level.

Ex-OpenAI RPO above 25%. Printed at exactly 25%. Down from 28%. Decelerating toward the threshold I set in January. One more quarter and this condition fails.

GitHub product leadership. Revenue accelerated 60%, on a billing-model change to usage-based pricing, not demonstrated product superiority over Claude Code. The franchise monetises better while the product contest remains unresolved. Half-met.

Two conditions ahead of pace. One at the line. One half-met. The score isn’t pass or fail. It’s ambiguous. And the ambiguity is the honest finding.

What Microsoft Saw and What It Missed

This is the hardest paragraph I’ve had to write in this series.

Microsoft had exclusive access to OpenAI from 2019. Seven years of the closest possible seat to the frontier. Hundreds of billions committed. And the defining product forms of the AI era appeared somewhere else. ChatGPT established the conversational interface. Cursor and Claude Code pushed coding from autocomplete to autonomous work. Anthropic built Cowork, the agentic tool for knowledge workers. Microsoft’s answer was to import Anthropic’s technology into Microsoft 365 and call it Copilot Cowork.

This quarter added a second data point. The flagship autopilots, autonomous, long-running agents with full enterprise compliance, ship “powered by OpenClaw.” That’s consecutive quarters where the defining agentic capability arrives from outside. Nadella’s own language has migrated from “chat” to “Copilot” to “autopilots” to “super app,” each rename conceding that the previous assumption about how AI would meet users was wrong.

Access to innovation is not product imagination. Microsoft saw the model. It misread the adoption vector.

Meanwhile, the model layer I had expected to commoditise did the opposite. Anthropic’s ARR is tracking at roughly $74 billion. OpenAI at $41 billion. Combined: $115 billion, growing at triple-digit rates. Seven months ago the combined figure was around $32 billion. In my October essay I described the Engine layer, AI models, as “rapidly commoditizing” and positioned it as a shrinking input to the Cockpit’s governance premium. That was wrong. The Engine didn’t compress into irrelevance. It became a massive, independently funded freight train, consuming hyperscaler infrastructure at unprecedented rates and building its own enterprise distribution that bypasses Microsoft’s EA channel entirely. Developers adopted Claude Code and ChatGPT, enterprises followed, and procurement caught up afterward.

That is the user-cycle pattern the Q3 essay warned about. It’s running faster than I expected.

The Enterprise Anesthetic

Copilot at 30 million seats is extraordinary distribution. The engagement data, satisfaction scores doubling over three quarters, weekly usage on par with Outlook and Teams, deployment-to-high-usage falling from months to days, is genuinely strong. I am not dismissing it.

But enterprise distribution can protect revenue while concealing product decay. And there is a meaningful gap between deployment and competitive advantage.

There are five levels of AI adoption:

A seat procured by a CIO. A seat assigned to an employee. A user who opens the product regularly. A user who chooses it when Claude, ChatGPT and Copilot are all available. A company whose competitive performance improves because it deployed the product.

Microsoft has strong evidence at levels one through three. What we still lack, from Microsoft, from the sell side, from anyone, is evidence at level five. Did the customer actually compete better?

That matters because disruption doesn’t reach Microsoft through its sales channel. It reaches Microsoft through its customers.

AWS didn’t defeat Microsoft by persuading CIOs that EC2 was architecturally superior to Windows Server. Developers and startups adopted AWS, moved faster, and changed competitive expectations. Then enterprise CIOs adopted cloud because the companies built on cloud were beating the ones that weren’t. The sequence was: better developer tool, faster product cycles, stronger customer outcomes, competitive pressure on incumbents, CIO adoption.

AI could follow the same path. A bank doesn’t switch from Copilot to Claude because a developer survey says Claude is better. It switches when another bank approves loans faster, ships software with fewer engineers, resolves complaints more cheaply. Unilever doesn’t abandon the Microsoft bundle because a power user mocks Copilot on Twitter. It reconsiders when a rival using AI-native tools compounds faster across every function.

The AI-native firms don’t need to defeat Microsoft’s salesforce directly. They need to help Microsoft’s customers’ competitors win.

Distribution buys time. Customer competition sets the deadline.

1873

In 1873, the railroad boom collapsed. The Northern Pacific Railway’s failure triggered a six-year depression. The industry had overbuilt, too many parallel routes, too much leverage, too many operators competing for the same freight. The railroads that survived were not the ones with the most track. They were the ones with the best routes, the most efficient operations, and the least debt. Vanderbilt. Hill. Huntington. They bought the wreckage at pennies and earned toll-road returns for a century.

The CEO of the company running the largest infrastructure build in technology history told investors to read this book. The CFO told them to run an efficient railroad.

This is not casual commentary.

Management is privately modelling a shakeout scenario and positioning Microsoft as the survivor. The evidence is specific: self-funded capex with no equity issuance (Alphabet raised $49.6 billion in stock in June; Microsoft did not); margin stability with FY2027 operating margins guided down less than a point; capex flexibility, with Hood, for the first time, explicitly telling analysts that two-thirds of spend is short-lived GPUs and CPUs that can be slowed if demand softens. And right-of-way: if your agent needs to access enterprise data, it authenticates through Microsoft.

The 1873 reference also clarifies the product-layer question. Railroads didn’t own the cargo. They didn’t need to. They owned the track, the gauge standard, and the toll schedule. The efficient ones didn’t need the best locomotives, they needed the best routes and the lowest operating costs.

My October essay called the three layers Grid, Engine, Cockpit. By Q3, the Engine was supposed to be commoditising. Instead it hyperscaled to $115 billion of ARR from two companies. The Engine isn’t a commodity input. It’s a massive, independently funded freight train running on hyperscaler track.

The framework that fits now: Track, Cargo, Right-of-Way. Azure is the track. The model companies are the cargo. Agent 365, identity, compliance, governance, is the right-of-way. You don’t need to own the cargo. You need the cargo to need your track. And you need the right-of-way to be exclusive rather than an open road.

What Makes the Right-of-Way Exclusive

All three hyperscaler clouds are accelerating simultaneously. The question is the shape of the acceleration.

Source: Alphabet and Amazon earnings releases; Azure constant-currency growth from Microsoft filings. AWS Q2 2026 is consensus.

Azure’s growth rate is strong. It is also the flattest line on this table. GCP accelerated 50 percentage points in five quarters. AWS accelerated roughly 15. Azure accelerated eight. All three are riding the same demand wave, the $115 billion in frontier-model ARR flows through their infrastructure, but the growth curves suggest Azure’s share of incremental cloud AI demand is not expanding. The tide is real. Azure’s position within the tide is stable, not gaining.

That’s fine for the stock if the stock were an infrastructure story. It isn’t. At 19x, the market has already discounted the infrastructure layer. What the stock needs to re-rate is evidence that the other layers, the right-of-way, are exclusive. Agent 365. Entra. Purview. Work IQ. The governance plane.

And this is where the answer to the Q3 question lives.

If the right-of-way is exclusive, if every agent, regardless of model provider, must authenticate through Agent 365 before touching enterprise data, then governance is the operating system and the CIO cycle wins, because compliance centralises whether or not users prefer Claude.

If the right-of-way is replicable, if enterprises can govern agents through open protocols, competing identity systems, or the model providers’ own enterprise offerings, then governance is a feature, not a chokepoint, and user preference determines which agent wins.

The answer depends on a single variable: Agent 365 adoption depth. Not whether enterprises register agents, 40 million says yes. Whether enterprises govern non-Microsoft agents through Agent 365. If Claude Code agents inside a bank authenticate through Agent 365 before accessing enterprise data, the right-of-way is exclusive. If they run through their own API harness with a separate compliance stack, Agent 365 is a Microsoft 365 governance tool, not an ecosystem control point.

That distinction is the difference between a $700 stock and a $430 stock. Both above today’s price.

What I Own and What Would Change My Mind

At $386, the stock is priced for the worst outcome, utility, IBM-with-better-infrastructure, 18-19x forward earnings. The right-of-way outcome, which I weight at 50%, implies $700-750 in three years on 17-18% revenue growth, 44-46% operating margins, and a 24-26x terminal multiple. The operating-system outcome, now 20% in my framework, down from 30% after two consecutive quarters of product-layer imports, implies $1,000-plus on 22-23% growth, 48-50% margins, and 28-30x. Even the bear case at 20% probability generates $430-480 on 12-13% growth, 40-42% margins, and 18-20x. Probability-weighted: roughly $700. The stock here prices the worst outcome and pays you to wait for the resolution.

Three things I’m watching:

Agent 365 depth. Not registrations, governing non-Microsoft agents. Any disclosure that Claude, ChatGPT, or third-party agents authenticate through Agent 365 at enterprise scale is confirmation the right-of-way is exclusive. Silence on this metric, or the disclosure quietly dropped, is the opposite signal.

Ex-OpenAI RPO. At exactly 25%, my own threshold. Below 22% next quarter means the broad-based governance premium isn’t holding.

Customer competitiveness. The hardest to measure, the most important to track. Any evidence that a company using AI-native tools outperformed a competitor using the Microsoft suite is the leading indicator that disruption has arrived through customers, not through Microsoft’s sales channel.

The Q3 essay asked whether AI is a CIO cycle or a user cycle. After quarter one of four, the honest answer: it’s both, simultaneously, and the outcome depends on which cycle compounds faster. Microsoft’s CIO machine is converting at record pace. The user cycle, Anthropic at $74 billion, Claude Code leading developers, ChatGPT Work growing, is also converting, and it’s not flowing through Microsoft.

The efficient railroad survives 1873. The question this series will answer is whether it emerges owning the right-of-way or just operating the trains on someone else’s track. At 19x, you’re buying the track at cargo prices.

That’s the bet.

Subscribe now

The five hyperscalers at the centre of the AI buildout, Alphabet, Amazon, Meta, Microsoft and Oracle, reported approximately $405 billion in capital expenditure in 2025.

Goldman Sachs expects that figure to reach roughly $750 billion in 2026 and $1.14 trillion in 2027.

The funding mix is changing with it.

The companies issued $108 billion of investment-grade bonds in 2025, equivalent to approximately 26% of their capex. In the first half of 2026, issuance had already reached $194 billion. Goldman expects it to approach $250 billion for the full year, or roughly 33% of capex.

For 2027, the projection is $400 billion of bond issuance against $1.14 trillion of capex - a 35% share.

Alphabet has already shown what this pressure can look like. Its capital spending pushed quarterly free cash flow to negative $5.9 billion in Q2, the company’s first negative quarter since listing. Its trailing-12-month free cash flow remained positive, but the direction is clear: internal cash generation is covering a smaller share of the buildout.

And direct hyperscaler bond issuance is only one layer.

Goldman separately estimates that $489 billion of AI-related debt has been issued across the wider ecosystem in 2026, with hyperscalers accounting for approximately 40%. The remainder has come through data-centre finance, semiconductors, software and other parts of the infrastructure chain.

The structures are evolving too. Meta and BlackRock’s newly announced El Paso data-cententre venture has approximately $14 billion in development costs, with part of BlackRock’s investment funded through $12.5 billion of debt financing. Meta will lease the entire campus from the venture.

The point is not that a credit crisis is imminent. Many of the largest borrowers retain substantial financing capacity.

But capacity to issue and market appetite to absorb are separate questions.

Corporate supply on this scale is arriving in a bond market already absorbing heavy government borrowing. Both ultimately compete for duration demand. The Fed can influence the front end of the curve; the long end must still clear the volume.

This is not fiscal dominance by itself. It is where the AI capex cycle collides with it.

The AI trade is no longer only an equity story. It is becoming a bond-supply story, and one that could reach portfolios through yields, curves, currencies and asset allocation.

Four forces. One hour. Live.

Join Bruno Crastes, Co-Founder and Group Corporate and Market Strategy Director at H2O Asset Management, for:

Fiscal Dominance: The New Reality for Global Markets

🗓 Thursday, 30 July 2026
🕔 5:00 pm SGT/HKT | 2:30 pm IST | 1:00 pm GST

Register Now!

We will examine four forces reshaping the next market cycle:

• How debt burdens are influencing monetary policy

• Why a softer Fed may not mean lower long-term yields

• Whether inflation is a shock or a regime

• How AI capex looks against hard economic evidence

This is a live session with audience Q&A. Bring your questions and put them directly to Bruno.

Sources: Goldman Sachs credit strategy research led by Amanda Lynam, as reported by Yahoo Finance on 22 and 28 July 2026; Alphabet Q2 2026 earnings; Meta Platforms announcement dated 28 July 2026.

• The semiconductor sell-off was not a verdict on AI demand. It reflected a breakdown in the assumptions that scarcity, cheap capital, independent customer demand, and supplier pricing power would persist indefinitely.

• The central question is whether AI can become self-funding before it becomes overbuilt. Capacity is being constructed faster than enterprises can redesign workflows, establish trust, and convert technical capability into dependable revenue.

• The next AI trade will reward conversion, not construction. The winners will own durable bottlenecks, sell to customers with independently generated cash flow, and demonstrate rising economic returns as infrastructure scales.

On Tuesday, the S&P 500 closed slightly higher. Beneath that calm surface, fundamental long-short funds, systematic strategies and multi-strategy equity books all lost more than 1% on the same day, something Goldman had not recorded since March 2020. Technology portfolios were being de-grossed, memory stocks suffered exceptional long selling, and leveraged Korean products were forcing trades into a market whose liquidity was disappearing by the hour. The index was green. The AI trade underneath it was in liquidation.

The semiconductor index had fallen roughly 26% from its June high, while Samsung and SK Hynix suffered double-digit daily declines. Yet there had been no obvious collapse in AI usage. Access to leading models remained constrained, infrastructure plans continued to rise, and SK Hynix had just reported record quarterly revenue and operating profit. The industry was still describing demand it struggled to serve.

I began with the simplest explanation: a crowded trade had finally broken. That was true, but insufficient. Positioning explained why the decline became so violent. It did not explain why investors had suddenly become willing to sell every version of the same long-term AI story.

The deeper change was interpretive.

For the first phase of the boom, the market had a simple model. More capital expenditure meant more accelerators; more accelerators meant more high-bandwidth memory, advanced packaging, networking, power and semiconductor equipment; more demand against constrained supply meant stronger pricing and higher earnings.

Scarcity was the signal.

Over the past month, scarcity stopped being enough.

Thirty Days That Broke the Scarcity Trade

The AI trade rested on four assumptions: demand would outgrow supply; physical bottlenecks would remain scarce long enough to preserve pricing power; customers would eventually generate the cash required to pay for capacity; and capital markets would give monetisation time to arrive. Each came under pressure in a different way.

First, the cost of time changed. Renewed conflict between Iran and the United States pushed oil and inflation risk back into the market just as a new Federal Reserve chair was signalling less tolerance for persistent inflation. Within weeks, the debate moved from when rates might be cut to whether the Fed could raise them almost immediately.

Higher rates do not eliminate demand for intelligence. They shorten the period during which an unprofitable project can wait for demand to become revenue. Land, power and equipment must be secured before applications generate cash. A cash-rich hyperscaler can tolerate that gap; a specialist provider dependent on refinancing cannot. Macro determines how long the system has to answer the AI question.

Then intelligence itself appeared less scarce. Chinese open-weight models such as Kimi and Qwen suggested that useful capability could spread more quickly and cheaply than investors had assumed. The immediate interpretation was that cheaper Chinese models meant fewer Western chips. That skips a step.

Open models attack the pricing power of frontier laboratories more directly than physical consumption. Lower prices can attract more users, make lower-value tasks viable and enable several agents to work in parallel. If the price of a useful outcome halves and consumption more than doubles, physical demand expands; if consumption rises only modestly, users capture the benefit while capacity providers face weaker economics. Adoption can therefore remain strong while value migrates away from the laboratories supporting much of the buildout.

The third challenge was to the duration of hardware scarcity. Domestic Chinese immersion DUV production does not make those tools equivalent to ASML systems on throughput, uptime, overlay, yield or cost per good wafer. What changed was not current supply but the market’s estimate of how long Western scarcity rents might persist.

The fourth concern attacked the quality of demand. Bloomberg reported potential Nvidia-related arrangements worth more than $750 billion. The SK Group agreement includes reciprocal procurement, but the possible OpenAI arrangements are different: Nvidia was reported to be considering up to $250 billion of lease guarantees and financing roughly $350 billion of chip purchases. Google has reportedly backstopped Anthropic-related leases, while SoftBank has financed a large OpenAI commitment. The talks may change or fail, but their scale is difficult to dismiss.

None of this proves that demand is fictitious. Young industries often require coordination. A model laboratory needs compute before it can create the capability and revenue that make the compute affordable. A data-centre developer needs a committed tenant before lenders will fund construction. A supplier with superior information and a strong balance sheet may rationally bridge that gap.

Yet financing changes what an order tells us. A conventional purchase says the customer has cash, or can borrow against its own credit, and has chosen to spend it on the supplier’s product. A vendor-supported purchase says the supplier believes strongly enough in the customer’s future to help create the purchasing power. The second may be an excellent investment. It is not independent validation.

The market may not believe AI demand is fabricated. It may believe the timing of that demand is being financially engineered.

Then the trade broke. Crowded momentum exposure, high valuations, Korean leveraged ETFs, dealer hedging and systematic thresholds transformed a month of changing assumptions into indiscriminate liquidation. Fundamental concerns supplied the reason to sell. Market structure supplied the scale of the selling.

The Question Beneath the Sell-Off

At an inflection point, information does not disappear; its meaning becomes unstable.

Higher capex means more current demand and more future supply. Lower model prices mean weaker laboratory monetisation and broader adoption. Better efficiency means fewer chips per task and more affordable tasks. Vendor financing signals confidence and incomplete customer funding.

This is why capex up, down or flat can all be read negatively. The market has lost a stable framework for interpreting the data. The numbers had not merely changed. Their meaning had.

The question underneath the confusion is:

Can the AI boom become self-funding before it becomes overbuilt?

The fuller version is more precise:

Can falling costs and improving capability create enough broad, independently financed end-user demand before the physical capacity and financial obligations built in anticipation of that demand outrun the resulting cash flows?

The AI buildout moves through four stages:

The failure state is not necessarily a collapse. It is digestion: Sponsorship creates capacity faster than Conversion creates monetisable demand. Utilisation or pricing disappoints, financing tightens and new projects slow while underlying usage continues growing into what has already been built.

That is the mental model: Scarcity → Sponsorship → Conversion → Self-funding.

The market is not seriously questioning the first stage. It is asking whether the industry can complete the final two before the obligations created in the second come due.

Intelligence Has a Meter

Traditional software combines high fixed costs with almost free distribution. AI keeps the easy distribution but not the near-zero marginal cost. Every inference consumes accelerators, memory, networking and electricity; every improvement in capability must eventually support a recurring physical bill.

That means tokens, GPUs and announced megawatts are activity measures, not return measures. The better measure is revenue density: gross profit per deployed unit of capacity after utilisation, power and depreciation. AI consumption can rise rapidly while revenue density falls, leaving users with enormous value and the owners of capacity with inadequate returns. The handoff works only when cheaper intelligence creates enough additional paid use to raise, or at least sustain, the economic output of each new unit of infrastructure.

Two Clocks

The strongest bear argument does not require AI to fail. It only requires the construction clock to move faster than the conversion clock.

The construction clock is governed by capital, land, power, permits and equipment. Once a project is approved, the process is difficult but legible. The conversion clock is governed by organisations. An enterprise must connect models to proprietary data, establish permissions, build evaluations, redesign processes and assign responsibility. Technical capability can improve overnight; institutional trust moves at the speed of budgets and behaviour.

This difference is clearest in the divide between verifiable and non-verifiable work.

Code offers a cheap reward signal. A program can be compiled and tested against a specification. Mathematics can often be checked mechanically. A game or simulation has an objective score. The model can generate many attempts and identify the better result without placing an expensive human expert inside every iteration.

Most knowledge work is less accommodating. A legal argument can be persuasive while misunderstanding the client’s objective; a medical recommendation can be plausible while missing the decisive fact. The reward signal is expensive because improvement requires scarce human judgment and often a delayed real-world outcome.

That can create a fundamentally different cost curve. In a machine-checkable domain, another block of compute can generate many attempts that are cheaply evaluated. In an ambiguous domain, each equal improvement may require disproportionately more compute, context and expert feedback. Capability still rises, but the cost of each step can rise faster.

Amdahl’s law adds another constraint. Making one component of a workflow ten times faster does not make the entire workflow ten times faster if the remaining stages are unchanged. AI may generate code instantly, but integration, testing, security review and deployment can still determine the pace of the complete job. A model can become superhuman at one stage while end-to-end productivity remains limited by the slowest human or organisational stage.

This is the most credible capital-cycle bear case. AI can continue improving, enterprise adoption can keep expanding, and capacity can still arrive before enough workflows are ready to use it productively.

The bull case is stronger because verifiability is partly a property of how work is organised. Tasks can be decomposed, tools can check facts, simulations can test consequences, multiple agents can critique one another and humans can review exceptions.

Coding became an attractive AI market not only because models improved. Software development already had compilers, tests, repositories, permissions, version control and monitoring. The surrounding system made the model’s work legible. Insurance, law, medicine and operations can build their own equivalents, even if the process is slower and less complete.

The addressable market is also larger than the existing wage pool. Cheaper intelligence can create work that humans never performed because human cognition was too expensive: continuous monitoring instead of quarterly review, thousands of simulated experiments before a physical one, or several agents exploring alternative plans in parallel.

I think this broadening will happen. I do not think it will happen on the same clock as power contracts, chip orders and data-centre construction.

That mismatch is enough to create a cycle.

When the Seller Becomes the Lender

The financing structure tells us whether the industry is bridging the clock mismatch or hiding it.

Nvidia can use the profits from present scarcity to accelerate its market, secure memory supply and encourage Nvidia-based system designs. That may deepen its advantage, but it also changes its economic exposure.

A conventional semiconductor supplier ships a product, receives cash and leaves the customer responsible for earning a return. A supplier that provides equity, purchase financing or guarantees retains exposure to customer solvency, utilisation and refinancing conditions after delivery.

The useful distinction is not circular versus non-circular. It is acceleration versus validation.

Would the same capacity be built, at the same price and on the same schedule, without guarantees, strategic equity and financing from companies that sell to or benefit from the project?

If the answer is yes, supplier capital is accelerating an independently viable market. If the answer is no, future demand is being pulled into the present and the resulting order cannot be treated as independent evidence that the economics already work.

The healthy loop is straightforward: strategic capital builds capacity; capacity enables better models; better models create useful applications; applications generate customer cash; customer cash finances the next wave.

The fragile loop begins almost identically: suppliers support customers; customers commit to supplier-based capacity; those commitments support revenue, valuations and new borrowing; new capital finances more purchases.

The two loops look identical while facilities are being built. They separate when the bills arrive.

A useful measure would be a sponsorship ratio: the share of incremental capacity supported by vendor guarantees, strategic equity, cloud credits, government assistance or related-party financing rather than arm’s-length customer cash and credit. Its level will be hard to estimate and means little in isolation; early industrial systems often require support. Its direction matters.

If the first $100 of capacity needs $30 of strategic support and the next $100 needs only $15 because utilisation and customer cash flow have improved, the handoff is working. If the next $100 needs $45, the system is becoming less self-sustaining as it grows.

Financing is therefore both evidence of Nvidia’s extraordinary strength and a symptom of incomplete downstream capital formation. A weak company could not act as the balance sheet of an industry. A fully self-funded industry would not need it to.

One Income Statement, Two Memory Cycles

SK Hynix’s quarter shows why the framework matters at the company level.

The company reported record revenue of KRW79.3 trillion and operating profit of KRW60.5 trillion, yet both fell below elevated expectations. The market initially treated the miss as one signal about “AI memory.” The underlying disclosures pointed to two businesses moving on different economic paths.

Conventional DRAM remains exposed to familiar memory economics: common standards, switchable suppliers and pricing governed by capacity and inventories. HBM is different. Multiple dies must be stacked and connected, the logic base die is becoming more customised, qualification is lengthy and the customer’s roadmap becomes tied to the supplier’s process and packaging performance.

SK Hynix began mass shipments of HBM4 and disclosed long-term agreements with around ten customers. The arrangements include pricing structures designed to address volatility and financial mechanisms such as deposits intended to support fulfilment and improve demand visibility.

That is evidence of movement from Scarcity toward Conversion. Demand is shifting from opportunistic spot purchasing toward commitments embedded in multiyear roadmaps. It is not yet proof of Self-funding: a long contract can reflect strong downstream economics or strategic fear of being unable to secure capacity.

Still, treating HBM and commodity DRAM as one homogeneous cycle is a category error. One part of SK Hynix remains governed by traditional memory economics. The other is becoming a process-differentiated franchise embedded inside that cycle.

China adds another inversion. CXMT’s policy obligation to develop domestic HBM could force it to allocate more wafers to a product that initially yields fewer useful bits per wafer than commodity DRAM. The near-term result may be less conventional DRAM supply, even as the market reacts to Chinese equipment progress as though a flood were imminent. The long-term supply threat is real; the immediate allocation effect may run in the opposite direction.

The same distinction applies across semiconductors. ASML is valuable not because no one else can generate light, but because customers care about throughput, overlay, uptime and cost per good wafer at enormous scale. TSMC is not paid merely for owning factories, but for leading-edge yield and dependable integration. KLA benefits not simply from more wafers but from the rising economic cost of detecting defects late as HBM and advanced packaging become more complex.

The next semiconductor market will distinguish these structural capabilities from generic exposure to more capacity.

The Handoff Scorecard

The framework becomes useful only if it can be updated.

This scorecard matters more than whether one quarter’s Azure guide is two percentage points above consensus. A cloud beat can produce a squeeze. It cannot resolve the debate unless it reveals utilisation, demand breadth, revenue conversion and customer quality.

The next frontier-model release matters for the same reason. The market does not need another system that scores marginally better on familiar tests. It needs evidence that capability gains are opening new categories of paid work faster than capacity is arriving.

Credit should be watched alongside the models. Wider spreads, larger guarantees, shorter maturities or weaker lease protections may show that the handoff is failing before chip orders slow.

The Next AI Trade

The first AI trade rewarded exposure to scarcity. The next will reward proof of conversion.

For every company in the chain, investors should ask three questions.

What remains scarce when headline capacity becomes abundant? The answer may be leading-edge yield, HBM packaging, process control, dense power, proprietary workflow data or customer distribution. A temporary shortage is not a durable advantage.

Who pays the company, and from what independently generated cash flow? A hyperscaler monetising AI inside an existing cloud, advertising or enterprise business is different from a specialist provider dependent on one frontier laboratory. The same dollar of backlog can represent very different economic demand.

Does scale improve the economics, or require more sponsorship? A healthy system becomes easier to finance as it grows because utilisation, customer diversity and revenue density improve. A fragile one becomes larger while remaining dependent on favourable capital markets, rising asset values and new guarantees.

The sell-off became more violent than the change in end demand justified, and forced deleveraging can push durable companies below reasonable value. Yet buying the same broad basket at lower prices assumes the basis of the trade has not changed. That basis has changed.

The strongest bear case is not that AI fails. It is that AI succeeds too slowly relative to the physical and financial obligations made on its behalf. I think that case will be partly right.

My base case is a rolling digestion, not a broad collapse. The growth rate of capital expenditure peaks before useful AI consumption does. Some projects are delayed or restructured, generic capacity sees weaker pricing, and poorly financed operators struggle. Frontier-grade systems, leading HBM and the hardest manufacturing bottlenecks can remain scarce throughout the adjustment.

This view becomes more bullish if models begin completing long-horizon professional work reliably; if paid usage grows faster than price deflation; if revenue density rises as capacity opens; and if the sponsorship ratio falls. It becomes more bearish if each wave requires larger guarantees, credit weakens despite strong usage and strategic capital becomes more important rather than less.

The fibre buildout offers the useful historical warning. Investors were right that bandwidth demand would explode, but many were wrong about the sequence: capacity arrived, financial obligations followed, and sufficient cash flows came too late. The internet won; many of the companies financing its arrival did not. Technological inevitability is not the same thing as an attractive return for every owner of capacity.

The historic importance of this drawdown is not that the market has conclusively answered the question. It is that the old question no longer works.

The market has stopped asking whether AI will require more compute. It has begun asking whether the value created by that compute will reach the companies financing it soon enough.

Scarcity got the system built. The handoff will determine who earns the return.

The AI boom does not need to end for the AI trade to change. It only needs to start paying its own bills.

Subscribe now

• KLA’s structural position is getting stronger: HBM, advanced packaging and more complex logic are increasing the cost of late-stage defects, and the value of finding them earlier.

• The market moved ahead of the business: the stock priced in KLA’s acceleration before it appeared in reported revenue, then punished a quarter that was strong but below elevated buy-side expectations.

• Cash is now the key test: margins, memory exposure, packaging and services are tracking ahead of prior signposts, but free-cash-flow conversion must recover to validate the earnings trajectory.

In early May, when I published The Complexity Tax Gets More Real, KLA traded around $194. It reached $307.37 by the end of June, returned to $190.80 before the July 28 report, then fell to roughly $175 after hours. The stock travelled from doubt to certainty and back before KLA completed a quarter. The thesis did not make the same round trip.

My first KLA article argued that the company is best understood as a tax on semiconductor complexity. The second added a sequencing point: the inspector arrives after the builder. Lam Research and Applied Materials benefit first when customers install the tools that create capacity; KLA’s relative advantage becomes clearer when those fabs must turn capacity into yielding wafers and packages. The tax collector does not arrive first, I wrote. The tax collector always arrives.

Both claims look better after this quarter. What I got wrong was the stock implication. I had treated the handoff from capacity to yield as a business sequence that the market would recognise as it happened. Instead, the market priced the handoff while the builders were still working, crowded into the cleanest beneficiary, and then demanded that every quarter arrive ahead of an expectation that had already moved ahead of the evidence. KLA’s inspector is arriving. The market arrived first.

Scoring Our Own Work

The prior article left a clear scoreboard. Gross margin above 62.5% by September would show that the 2030 path remained open; June came in at 62.4%, and September is guided to 62.5%. Memory needed to recover above 25% of systems revenue by mid-2027; management now expects 27% in September, several quarters early. Advanced packaging needed to sustain more than $275 million per quarter; KLA now expects roughly $1.1 billion for calendar 2026, up more than 70%. Services needed to grow above 12%; it grew 17%.

The miss is cash. Free-cash-flow conversion was 59%, far below the 80% threshold I set last quarter. The relative growth gap with the capacity builders is narrowing, but has not disappeared. Four signposts moved in the right direction, one is partly confirmed, and the one that matters most for earnings quality remains unresolved.

The Unit of Failure Has Changed

The fundamental question is whether rising semiconductor complexity permanently increases the share of customer spending devoted to KLA, or whether logic, HBM and packaging are simply three spending cycles arriving together.

I think the structural answer has become stronger because the unit of failure has changed. A defect once destroyed a die. Increasingly, it can destroy a system containing a leading-edge logic chip, several HBM stacks, an expensive interposer and thousands of connections, after each component has absorbed dozens of costly manufacturing steps. The later the defect escapes, the more accumulated value disappears with it.

KLA sells earlier knowledge: its tools tell the customer that something has gone wrong before another ten, twenty or fifty process steps add more value to the defective object. The worth of that knowledge rises with the future cost it prevents the customer from wasting. KLA becomes more valuable because the cost of discovering a defect late rises faster than the cost of finding it early.

The mechanism compounds. More complex devices create more failure modes; those failures add to KLA’s application knowledge, defect libraries and field experience. Better detection improves yield learning, expands the installed base, and produces service revenue, customer contact and cash for the next generation of R&D. Scale does not push KLA toward commodity economics. It widens the experience gap between KLA and anyone trying to displace it.

This quarter shows that ratchet turning on three fronts.

Logic, Memory and Packaging

The logic arm appears in patterning, where revenue rose 61% year over year and 18% sequentially to $728 million. Patterning includes metrology and reticle inspection, the parts of KLA most exposed to tighter geometries, more EUV layers and harder alignment problems. Wafer inspection grew only 1%, so the quarter was not simply more of the same equipment sold into more wafers. Some of the 61% growth will be shipment timing. The direction is harder to dismiss.

The memory arm is more consequential. Memory represented 21% of semiconductor process-control systems revenue in June and should reach 27% in September, with DRAM making up about 90% of the category. CFO Bren Higgins said HBM intensity is “rivaling what we see in advanced logic.” HBM is turning memory from a relatively low-inspection process into a stacked yield problem where one bad die or bond can impair the whole assembly.

The packaging arm may matter most. KLA lifted its calendar 2026 advanced-packaging process-control target from high-50% growth to more than 70%, or about $1.1 billion. Rick Wallace said packaging roadmaps are moving toward “more sophisticated wafer front-end processes.” KLA is not forcing its way into an unrelated market; the customer’s back-end process is becoming precise enough to move into KLA’s established domain. PCB and component-inspection systems grew 96% year over year, while the broader segment grew 56%.

Services complete the loop. Revenue reached $820 million, up 17%, marking the seventeenth consecutive year of annual growth. Every new system sold during this buildout expands the recurring base that remains after equipment spending slows.

Why the Stock Fell

KLA reported $3.66 billion of June revenue against roughly $3.60 billion on the Street, but the reported buy-side bar was about $3.70 billion. The September guide of $4.0 billion cleared published consensus near $3.91 billion yet fell short of the roughly $4.05 billion bogey. Gross margin and EPS were strong; the two revenue numbers governing the acceleration story were merely good.

That explains why KLA gave sellers a reason. It does not explain the size of the move.

KLA entered the print in a semiconductor tape that had stopped rewarding good news. On July 28, before the release, KLA fell 6.2%; Applied Materials and Lam fell roughly 8%, ASML more than 4%, and Micron nearly 9%. KLA then lost another 8.4% after hours. The company-specific shortfall explains why KLA joined the selloff. The sector regime explains why a one-percent miss against an inferred bogey became a two-stage decline.

The call sounded far stronger than the price. Higgins disclosed backlog of about $12.5 billion, more than three quarters of current revenue, in response to Jefferies’ Blayne Curtis. Wallace said he had very few concerns about 2027, while Higgins said KLA was “sizing the company to serve the more bullish scenarios.”

I do not see a contradiction between that confidence and a $4.0 billion guide. Annual demand can be unusually visible while quarterly revenue remains dependent on tool completion, fab readiness, delivery and customer acceptance. The market wanted precision from a business that has visibility without perfect timing.

The Model May Already Be Too Low

KLA’s March Investor Day model targets about $26 billion of revenue and $8.40 of EPS in 2030, based on wafer-equipment spending of roughly $215 billion. It assumes process control grows faster than WFE, KLA gains more than 150 basis points of WFE share, services compound at 13% to 15%, and advanced packaging adds another source of growth.

Bernstein’s Stacy Rasgon pressed Wallace on the awkward implication. If calendar 2027 WFE is already approaching $190 billion, the model allows only about 4% annual growth from 2027 to 2030. Wallace’s answer was candid: “We are not very good at forecasting.”

If WFE reaches $190 billion in 2027 and grows at 8% to 10% through 2030, it lands around $239 billion to $253 billion, not $215 billion. KLA will not capture every incremental dollar, and a digestion period could sit between 2027 and 2030. Yet the company’s own framework assumes share gains, rising intensity, mid-teens services growth and packaging expansion. If the higher industry path holds, $26 billion of revenue and $8.40 of EPS begin to look more like a floor than a destination.

I would not promote the most optimistic sensitivity—$30 billion-plus of revenue and $10 to $12 of EPS—to the base case today. I would say something narrower: four months after Investor Day, the evidence beneath KLA’s 2030 model is running ahead of the assumptions used to build it.

That is the variant view. The market is treating KLA’s second wave as a discrete revenue event. I think logic, HBM and advanced packaging are adding process-control steps that later generations will inherit rather than reverse.

The Number That Still Does Not Fit

KLA generated $817 million of free cash flow, equal to 22% of revenue and 59% of non-GAAP net income. Receivables absorbed $586 million, inventory $212 million and other assets $423 million, partly offset by a $312 million increase in deferred system revenue.

The benign explanation is credible: KLA is funding working capital ahead of a shipment ramp, and rising deferred revenue supports the claim that customer commitments are real. Management says it is investing in inventory and facilities to meet demand.

Still, this is a business whose quality should improve as the installed base and services grow. Cash should eventually compound faster than revenue, not trail accounting profit quarter after quarter. I need free-cash-flow conversion above 80% by December. If it remains below 60%, I will reconsider how cleanly the operating model scales.

Margins deserve similar care. The prior 62.5% threshold has effectively been met, even with memory rising. That is better than I expected. But Higgins said memory-component costs of roughly 100 basis points may persist through next year, and he told UBS’s Tim Arcuri that existing backlog cannot simply be repriced after orders have been accepted. The margin risk has receded; it has not vanished.

Three Years From Here

The bear case assumes a 2028 digestion period, tighter China restrictions, lower process-control intensity and cash conversion that never fully recovers. The base case assumes KLA grows faster than WFE through share, intensity, packaging and services, while gross margin gradually approaches the 2030 path. The bull case requires all three arms of the ratchet to keep turning, a formal increase to the 2030 model, and cash conversion that validates the income statement.

Using the midpoints, the probability-weighted value is around $280. It is not a twelve-month target, and the terminal multiple contributes nearly as much uncertainty as the earnings forecast. It does tell me that the $175–190 range around the print is materially different from the prices at which I called KLA a perfect business at an imperfect price. I would begin adding here rather than demand a lower entry.

The signposts are clear: December revenue guidance above $4.25 billion; gross margin at or above 62.5%; memory holding near 27% or moving toward 30%; advanced packaging growing above 25% in 2027; services above 12%; and free-cash-flow conversion above 80% by December. If those conditions hold, the 2030 model will increasingly look conservative. If revenue accelerates while cash does not, the strategic thesis can be right and the stock thesis still fail.

The Market Arrived First

Three months ago, I argued that the inspector arrives after the builder. This quarter says the inspector is arriving: memory has recovered earlier than expected, packaging has become a billion-dollar business, patterning is surging, services are compounding and margins are holding.

What I missed was that the market would capitalise that arrival before it appeared in revenue, then sell the stock when the arrival was merely on schedule. The business question is now easier. The unit of failure is getting larger, the value of earlier knowledge is rising, and KLA’s process-control ratchet is turning across logic, memory and packaging. The stock question remains harder because valuation, timing and cash conversion can overwhelm a correct structural view for long stretches.

At the post-print price, I think the gap is wide enough to act. The thesis has advanced, the stock has retreated, and the one piece of proof I still need is cash.

KLA’s stock has completed a cycle. The business has not.

Subscribe now

• Agents are not replacing EDA tools; they are calling them more often. Core EDA grew 18%, recurring revenue approached 20% organically, and backlog reached a record $8.1 billion.

• Cadence has begun monetizing the agent layer through incremental licences and add-ons. However, it has not yet shown that revenue scales with design iterations or customer productivity.

• The next test is economic, not technical. Rising revenue per customer and margins above 45% in FY2027 would support the throughput-toll thesis; slower growth and weaker margins would suggest customers retain most of the AI productivity gains.

Ten days before Cadence reported earnings, the stock fell almost 10% after Moonshot AI’s Kimi reportedly designed and verified a chip using open-source tools.

The conclusion seemed obvious. If a general-purpose model could design a chip, the expensive software beneath chip design was about to matter less. Cadence had spent years presenting AI as a source of demand; Kimi appeared to turn the same technology against it.

Then Anirudh Devgan explained what the demonstration had shown.

The design was a small block built on a 45-nanometer process, technology roughly two decades old and, in Devgan’s telling, running at a frequency 20 to 30 times slower than current leading designs. Even there, the agent still needed EDA tools. It could generate intent, coordinate tasks and propose a design, but the result still had to be simulated and verified.

The agent did not bypass the physics. It called the engine.

That should have settled the question of whether Cadence remains technically necessary. It did not settle the more important question for the stock.

Six months ago, in Design Throughput Toll, we argued that Cadence’s historical growth was tied to engineering headcount. More engineers meant more seats, which meant more revenue. Agents offered a way out of that constraint: one engineer, assisted by autonomous design tools, could explore vastly more possibilities, and every possibility would call Cadence’s simulation, verification and sign-off engines.

The thesis was attractive because the technology apparently threatening Cadence could become the mechanism that accelerated its growth.

It was also too neat.

We treated greater tool consumption and greater value capture as if they were the same step. They are not. An agent can run one hundred times more design iterations while the customer remains on a contract negotiated before the agent existed. Cadence’s software becomes more useful. The customer becomes more productive. Whether Cadence gets paid in proportion to either outcome depends on how the commercial terms evolve.

We were right about the direction of demand.

We were too quick about the direction of value.

That is what Q2 clarified.

The Question Has Changed

The old question was whether agents would reduce the need for Cadence’s tools.

I think that question is largely answered.

At advanced nodes, an agent cannot hallucinate its way to a manufacturable design. Timing, power, electromagnetic interference, thermal behaviour, signal integrity and foundry rules are physical constraints. They have to be computed against process-specific models correlated with real silicon over many years.

The number of trusted tools capable of doing this remains very small.

Needham’s Charles Shi pushed Devgan on the extreme version: could a future model generate tape-out-ready files without commercial EDA at all?

Devgan’s answer was blunt:

“The ground truth will prevail.”

The new question is harder:

As AI multiplies the number and complexity of designs engineers can attempt, can Cadence make revenue scale with every additional iteration—or will customers capture the productivity while Cadence remains tied to contracts built for a seat-based world?

This is the Big Fundamental Question because the bull case and the bear case begin with the same event: agents make design exploration abundant.

In the bull case, every additional exploration creates demand for agent licences, more engine capacity, more verification hardware, more IP and eventually higher renewal value. Revenue detaches from human headcount.

In the bear case, those explorations run inside fixed agreements. Cadence becomes more indispensable, but the productivity accrues mainly to the customer. Revenue continues to grow with contracts and renewals rather than with the amount of design work performed.

The moat survives in both cases. The economics do not.

The Sentence That Mattered

The most revealing moment on the call came near the end, when Wolfe Research’s Joshua Tilton asked how Cadence was capturing financial value from becoming more central to its customers.

CFO John Wall replied:

“Customers continue purchasing our underlying EDA software. What they do is they purchase Cadence agent licenses that orchestrate engineering workflows. Generally, our economics scale with customer adoption.”

This is the first real commercial answer Cadence has given.

The base tools continue. The agent is sold on top. Cadence is not simply giving away every productivity gain and waiting years for contracts to renew. There is already an incremental product, an incremental selling motion and at least some incremental revenue.

That matters.

It also leaves plenty unanswered.

Wall did not say revenue scales with the number of iterations. He did not disclose agent revenue, renewal uplift, revenue per customer or pricing linked to design output. Management gave engagement counts—more than 20 for ChipStack and more than 25 for ViraStack—rather than dollars. That usually means the commercial contribution is real but still small.

Cadence has shown that agents can be monetized. It has not shown whether monetization rises in proportion to the value the agents create.

The flywheel has a meter.

We still do not know how sensitive it is.

The Evidence Appears

The quarterly numbers matter only because they tell us whether the throughput thesis has started to appear in the business.

It has.

Recurring revenue grew about 24%. Hexagon contributed roughly four points, leaving normalized growth in the high teens to around 20%. Core EDA grew 18%. Wall attributed the strength to share gains, healthy renewals and add-on demand from both “design for AI” and “AI for design.”

This occurred in what management described as a low year in Cadence’s three-year renewal cycle. First halves also usually draw backlog down as revenue is recognized. Instead, backlog increased to a record $8.1 billion.

That is the first convincing financial evidence that growth is no longer being driven mainly by engineers renewing conventional licences.

The rest of the company broadened with it. Semiconductor IP grew more than 40%, with management saying most of the increase was organic. System Design and Analysis grew 37%. Hardware delivered another record quarter and remains, in Wall’s words, “supply-constrained by customer demand.”

These are not four unrelated businesses.

They are four ways of getting paid for the same increase in design activity.

More custom chips require implementation and sign-off. More architectures require interface and memory IP. More alternatives require verification capacity. More chiplets create packaging, thermal and structural problems.

The design problem keeps expanding outward.

Cadence keeps following it.

The annual guidance raise confirms that this was not a backward-looking quarter. FY2026 revenue guidance moved from $6.125–$6.225 billion to $6.260–$6.340 billion. The midpoint increased by $125 million despite a quarterly revenue beat of only about $9 million against the Bloomberg estimate we used before earnings. Q3 revenue guidance of $1.595–$1.625 billion also replaced the feared second-half slowdown with sequential acceleration.

I had expected a good quarter and a modest raise.

I underestimated both the breadth of demand and the second-half setup.

The decoupling is no longer just a product argument. It is beginning to appear in recurring revenue.

The Twenty-Year Gap

The Intel agreement deserves more attention than a normal customer announcement.

Devgan described Intel as a “10 or 20-year-old problem.” Cadence had long been strong in the TSMC ecosystem. Its relationship with Samsung has improved. Intel remained the gap.

The new multi-year agreement covers Intel 14A, design-technology co-development, agentic EDA and Cadence IP. Management said the arrangement is entirely incremental and that most of the benefit remains ahead.

The near-term revenue is less important than what the agreement says about the scarce layer beneath the agent.

Models can proliferate. Customers can build their own orchestration. Interfaces will change. But leading-edge design still depends on process-specific rules, qualified flows and sign-off tools created through years of work with foundries.

Kimi can choose which engine to call.

It cannot create twenty years of correlation knowledge on demand.

Intel deepens Cadence’s process-specific position and raises the probability that Cadence participates in more designs built around Intel processes. It does not guarantee exclusive routing; Synopsys remains formidable, and sophisticated customers mix tools. But it closes a historical weakness at the exact moment physical truth is becoming more valuable.

The technical position is stronger than it was three months ago.

Which makes the pricing question more valuable, not less urgent.

The Cost of Capturing It

Cadence is not converting all the growth into near-term margin.

Q2 non-GAAP operating margin reached 45.5%, but Q3 guidance falls to 43.5%–44.5%. FY2026 non-GAAP expenses are expected to grow about 20%, while the full-year margin midpoint of 44.25% remains below 2025’s 44.6%. Management says it is investing in Intel, Hexagon integration, IP and the broader System Design and Analysis business.

That explanation is credible.

It also creates a promise.

If Cadence is becoming central to a wider set of customer decisions, the broader portfolio should eventually produce higher revenue per customer and better margins. If margins fail to recover in 2027, Cadence may be adding faster-growing businesses that lack the economics of Core EDA.

The accounting gap deserves attention as well. Stock compensation reached $147 million in Q2, more than 9% of revenue. Acquired-intangible amortization rose to $80 million from $24 million a year earlier. The gap between GAAP and non-GAAP operating margin now exceeds 17 percentage points.

Free cash flow softens the concern. Cadence generated $582 million in Q2, up 74% from a year earlier. The growth is producing cash.

Still, the 2027 test is clear.

Revenue growth above 14% with non-GAAP margin recovering beyond 45% would suggest the wider portfolio is improving both strategic relevance and economic quality. Strong growth with margins stuck below 44% would suggest Cadence widened the business without improving the revenue model.

What We Thought, and What We Think Now

Consensus can reasonably view Cadence as a high-quality EDA franchise enjoying an unusually strong AI design cycle. Hardware, IP, acquisitions and custom silicon are lifting 2026 growth toward 19%, after which revenue should normalize toward the low teens.

Our prior view was different. We thought agents could break the link between revenue and engineering headcount by turning design throughput into a new unit of demand.

Q2 makes that view more credible, but also more precise.

The decoupling may already be starting through recurring add-ons, agent licences and increased engine consumption. The market may still spend too much time asking whether agents replace EDA and too little time asking how much more EDA an agent consumes.

But agent adoption alone is no longer the variant perception. It is now visible.

The variant view has narrowed:

Cadence may be able to convert rising design activity into higher revenue per customer before the next renewal cycle fully arrives.

Wall’s description of agent licences suggests the commercial structure is moving in that direction.

What remains unproven is the capture rate.

Cadence does not need to adopt a literal per-call tariff. It could capture the value through premium agent tiers, larger enterprise agreements, more engine capacity, verification hardware, project-based pricing or higher renewals. The test is simpler than the contractual form:

Does revenue per customer rise as agent-enabled design output expands?

Three Possible Cadences

The scenarios are not really about semiconductor demand. They are three different answers to the same value-capture question.

The Essential Utility

Agents raise customer productivity, but pricing remains attached to seats and conventional renewals. Revenue compounds at about 9% through 2029, non-GAAP margin settles near 43%, EPS reaches roughly $10.50, and the market pays 26x.

Implied price: approximately $275–$290.

The Premium Add-On

Agent licences, IP, hardware and system analysis lift revenue per customer, but the basic subscription model remains intact. Revenue compounds around 13%, margin reaches about 46%, EPS approaches $13, and Cadence retains a 35x multiple.

Implied price: approximately $450–$460.

The Throughput Toll

Commercial terms begin to reflect agent workflows, engine capacity and the value of design output. Revenue compounds around 17%, margins approach 48%, EPS reaches roughly $15.50–$16, and Cadence earns a low-40s multiple.

Implied price: approximately $650–$670.

At a reference price around $339, the three-year annualized returns are roughly:

• Bear: -6%

• Base: 10%

• Bull: 24%

The probability-weighted value is around $460, or roughly 11% annualized over three years. That is attractive, but not overwhelming, because the bull case requires commercial evidence Cadence has not yet provided.

What Would Change the View

The next evidence should be hard to fake.

Normalized recurring growth above 15% would support the decoupling thesis; below 13% would suggest 2026 was mainly a strong cycle.

Core EDA growth above 15% would show that the main software engine remains healthy. Agent revenue, renewal uplift or rising revenue per major customer would move the thesis from usage evidence to pricing proof.

FY2027 guidance matters more than the next quarterly beat. Revenue growth above 14% with non-GAAP margin above 45% would show that Cadence can sustain growth while absorbing its investments. Growth below 12% or margin below 44% would suggest the current expansion is less durable or less profitable than it appears.

Backlog should remain near or above $8.1 billion despite revenue recognition. Stock compensation moving beyond 10% of revenue would weaken the per-share economics. DSO should return below 60 days rather than remain in the mid-60s.

These are the markers.

The Next Turn

Six months ago, we asked whether agents would create enough additional design activity to break Cadence’s dependence on engineering headcount.

Q2 answered the first half.

Agents are being deployed. They are calling the underlying engines. Customers are buying add-ons. Recurring software growth is approaching 20%. The base tools remain intact, and the foundry relationships beneath them are getting stronger.

The second half is harder.

Cadence must turn abundant design exploration into higher revenue per customer before customers learn to treat the productivity as part of the old contract. Agent licences are the first answer. Renewal pricing, revenue per customer and margin recovery will tell us whether it is enough.

The flywheel is real.

The open question is where the value goes every time it turns.

Subscribe now

• BESI’s earnings inflection no longer depends entirely on HBM. Logic, co-packaged optics, panel packaging, power management and even wearable devices are already driving advanced-bonding demand.

• The real opportunity is architectural, not cyclical. As performance moves from individual dies to the connections between them, bonding equipment becomes embedded earlier in product design and harder to replace.

• The process advantage is strengthening, but customers still control the clock. BESI can win qualifications, expand margins and deepen its installed base, yet semiconductor capacity cycles will continue to create sharp swings in orders and valuation.

In 1971, Intel introduced the 4004. The achievement was not the package around it. It was the fact that the processor itself fit on one piece of silicon.

That model defined the next five decades of computing. Performance came from putting more transistors onto a single die, then making those transistors smaller, faster and cheaper. Packaging protected the result, connected it to a circuit board and carried heat away. It mattered, but it came after the real work.

AI has broken that hierarchy.

The most capable systems now combine accelerators, CPUs, memory, optical links, power components, interposers and substrates across multiple dies. Their performance depends less on what any one chip can do alone and more on how closely, quickly and efficiently those chips work together.

The boundary between chips has become part of the computer.

BESI sells the machines that build that boundary.

That is the historical change behind the investment case. BESI is still an equipment supplier. Its customers still buy in waves. Capacity will still overshoot demand. Yet the role of bonding is changing. As more performance moves from inside the die to the connections between dies, bonding becomes less like a final assembly step and more like part of the product architecture itself.

The second quarter mattered because it showed this change already creating demand across several applications, before hybrid bonding reaches volume production in HBM.

BESI did not prove that memory has arrived.

It proved that the company can begin growing before memory does.

A Miss That Wasn’t

BESI reported second-quarter revenue of €249.9 million, slightly below Bloomberg consensus of €251.6 million. EPS of €1.11 also missed the €1.14 estimate.

A miss is a miss.

It just was not the right way to read the quarter.

Orders reached €292.9 million, above both the published estimate and the real bar I had set before the print. Book-to-bill held at 1.17x even though revenue grew 35% sequentially. BESI shipped far more equipment and still added backlog.

Then management guided third-quarter revenue up another 10–15% sequentially, roughly €275–287 million against a pre-print estimate near €254 million.

That changes the interpretation. Reported revenue tells me what customers ordered months ago. Orders tell me whether they are still expanding. Guidance tells me whether those orders are becoming a higher revenue base rather than one unusually strong shipment quarter.

The answer was yes.

I had been treating BESI’s near-term earnings path as a countdown to HBM hybrid bonding. The quarter showed why that framing was too narrow.

The Big Question

The single question I would use to judge BESI is this:

Can BESI turn the movement of system performance from inside chips to between chips into a self-reinforcing process advantage, or will customer-controlled capacity cycles keep revenue too uneven to justify the valuation?

The mechanism is straightforward.

As packages become harder to build, BESI becomes involved earlier in product development. Earlier involvement gives the company more chances to define the production process around its equipment. More qualified systems create more operating knowledge. That knowledge raises the odds of winning the next design.

The loop strengthens when the same underlying skill applies across several products.

Logic requires accuracy and throughput. Photonics adds alignment and new materials. Memory adds heat and density. Panel packaging changes the physical format and cost equation. The applications differ, but the core problem is the same: join more components with tighter tolerances at a yield and cost that production managers will accept.

The loop breaks if older bonding methods remain good enough, customers fund credible alternatives, or qualification takes so long that BESI’s technical position never turns into large commercial deployments.

This quarter supported the loop.

It did not prove that it is permanent.

HBM Is Still Waiting

The easiest BESI story begins with HBM.

Memory stacks become taller and hotter. Thermal compression bonding approaches its limits. Hybrid bonding creates direct copper connections, improves performance and reduces heat. BESI supplies the tools.

That story may still be right. It is just early.

When Needham’s Charles Shi asked what was holding back qualification at the three major memory producers, CEO Richard Blickman gave a direct answer:

“The major hurdles are simply cost and yield.”

That sentence separates a working technology from a working factory.

The bonding step may perform better, but materials, surface preparation, throughput and process integration all have to work together. Development teams can prove the physics. Operations teams must prove that millions of devices can be produced at acceptable yield and cost.

Blickman pointed to the leading Taiwanese logic customer, where qualification took more than three years before hybrid bonding became a mainstream production process. The large US logic customer followed a similar path. Memory is now passing through the same uncomfortable middle stage: the technical case is strong, while the factory economics are still being proven.

All three major memory companies have BESI systems. One is moving faster than the others and Blickman told us how close the decision is:

“We will shortly find out, do we pass the test for full production at this moment, or does it take a bit more time?”

That is not a man holding a signed purchase order. That is a man waiting for a verdict. Management expects clarity in the second half of 2026 on whether hybrid bonding enters HBM4E, HBM3 or another early application. The October call is the natural disclosure point.

Still, no volume-production decision.

I do not see that as a thesis break. I see it as a timetable reminder.

The Glasses

The more revealing part of the quarter was that BESI’s latest hybrid-bonding orders did not depend on HBM at all.

Three orders. Three different end markets. A repeat logic customer. A repeat co-packaged-optics customer. And a new hyperscaler that, per Blickman, is “probably for wearables... the next step in wearables is those glasses.”

Read that again. A hyperscaler ordered the most advanced chip-packaging equipment in the world to make glasses.

Nobody was modeling AR wearables as a 2026 hybrid bonding application. Not the sell side, not the delay-narrative bears, not me. And the order arrived as a purchase order, not a roadmap slide, which means someone has a product far enough along that they need production-grade packaging tools now. That is not speculative demand. That is a bill of materials.

This is the strongest single piece of evidence for the loop. The market spent July pricing BESI as a binary bet on memory timing. The stock fell 23% from its June peak on reports that Samsung and SK Hynix were pushing out hybrid bonding decisions. Then the quarter delivered new hybrid bonding demand from a direction no one was watching.

BESI added customers and orders before memory arrived.

Many Versions of the Same Change

BESI ended 2025 with 15 hybrid-bonding customers. It had 21 by the end of the second quarter.

The number matters. The spread matters more.

Hybrid bonding is now being worked on across logic, memory development, co-packaged optics, consumer devices and other chiplet designs. BESI is also benefiting from the same architectural shift through products that are not hybrid bonders.

Management said CoWoS-like equipment is already a meaningful part of orders. Demand began in the fourth quarter, continued through the first and second quarters, and should expand further. BESI has orders for panel-level packaging using the 310-by-310 millimetre format selected in Taiwan. Those production flows can require flip-chip, thermal compression and hybrid systems at different stages.

Then there is AI power management.

Management called out new orders from several customers for power-related applications used in datacenter compute modules. These rely on more conventional products, including flip-chip and multi-module attach.

ING’s Marc Hesselink observed that BESI tends to name a new category only when it expects the contribution to become material. Blickman agreed:

“We only mention when we feel with high certainty that that could be a meaningful mainstream contribution.”

This is less exciting than HBM.

It may be more useful over the next two years.

An AI system needs accelerators, but it also needs memory, power delivery, optical links, substrates, chiplets and packaging capacity. BESI can earn revenue at several of those boundaries.

These are not separate growth stories. They are different expressions of the same architectural change.

System performance is being created through assembly.

Becoming Part of the Architecture

The strongest BESI thesis is not that advanced packaging grows.

Many companies benefit when a market grows.

The stronger claim is that bonding becomes part of the architecture itself.

That changes when the customer chooses the equipment. A conventional assembly tool can be selected late in the manufacturing flow. A process that affects interconnect pitch, thermal behaviour, package thickness and yield must be considered much earlier.

Earlier involvement creates a deeper relationship.

Once a production line has been qualified around a specific machine, replacement becomes expensive. A customer can choose another supplier for the next programme, but changing an existing process means taking on new yield risk, retraining operations teams and reworking parts of the production flow.

That does not make BESI impossible to displace.

It makes displacement costly.

The quarter offered some evidence of that mechanism: repeat orders from logic and co-packaged-optics customers, a wider customer base, and strong orders even as shipments accelerated.

The missing disclosure is scale.

BESI does not tell us how many hybrid bonders it shipped, the order value by application, or how many of the 21 customers are building production capacity rather than running evaluation systems. Deutsche Bank’s Robert Sanders asked for unit numbers. Management declined.

That gap matters.

Twenty-one customers may represent the start of a broad production standard. It may also represent many experiments and a few large commercial programmes.

I do not know yet.

The most useful metric is not customer count.

It is production systems per customer.

The Margin Before the Standard

BESI’s second-quarter economics give the company time.

Gross margin reached 65.7%. Operating margin reached 43.5%. Net margin reached 35.6%. Revenue increased by €65 million from the prior quarter while operating expenses rose by less than €2 million.

The operating leverage is already visible.

BESI does not need HBM production to earn exceptional margins. Logic, photonics, advanced assembly and conventional systems are already producing an attractive mix.

The third-quarter gross-margin guide falls to 63–65% because revenue will contain a less favourable product mix. I do not see evidence of pricing pressure. The business is broadening into more mainstream equipment.

I would take that trade.

A lower-margin system can deepen a customer relationship, expand the installed base and place BESI inside another AI production flow. The immediate gross margin may be lower while the long-run commercial position improves.

The margin ceiling may also be higher than I first assumed.

Management’s long-term operating-margin range is 45–55%. BESI reached 43.5% at roughly €250 million of quarterly revenue, below the level implied by its third-quarter guide. If annual revenue moves above €2 billion and hybrid bonding becomes a larger part of sales, a peak operating margin in the high 50s is possible.

Peak is the key word.

A larger company will need more engineers, service capacity, spare parts and supply-chain redundancy. Customers will push on price. Competitors will spend more.

I would model high-50s margins in an unusually strong year, not as a permanent state.

Process Without Schedule

BESI’s advantage and its weakness come from the same place.

The company may become more deeply embedded in customer production. It still does not control when customers add capacity.

Robert Sanders asked Blickman about the danger of overinvestment in advanced packaging. The CEO did not pretend AI had repealed semiconductor history.

“In this cyclical world, it remains cyclical.”

Blickman said prior upcycles often lasted six to eight quarters. New fabs, packaging plants and OSAT capacity eventually create excess supply. High returns invite more investment. AI may extend the current cycle, but management does not claim to know by how much.

That answer captures the stock better than any total-addressable-market slide. The current upturn began in the third quarter of 2025. We are in quarter four.

BESI can control part of the process.

It cannot control the schedule.

The positive loop is:

Harder packages → earlier involvement → more qualification wins → more installed equipment → better process knowledge.

The negative loop is:

High packaging returns → more capacity spending → equipment orders surge → capacity catches demand → orders fall before revenue → estimates and valuation reset.

Those loops can operate at the same time.

BESI may enter the next downturn with a stronger position than it had entering the last one. The shares can still fall sharply when orders turn.

That is not a contradiction.

It is the investment case.

Variant Perception

The common view is that BESI is the leading supplier of hybrid-bonding equipment and will benefit when HBM adoption accelerates.

I think that view is too narrow in the near term and too relaxed about cyclicality in the long term.

The near-term earnings path is less dependent on HBM than it appears. Logic, co-packaged optics, CoWoS-like assembly, panel packaging, power management and mobile can carry revenue higher before memory reaches production.

The long-term value is more dependent on the duration and repetition of customer capacity waves than the bull case admits. Several applications do not automatically create recurring demand. They could all be different expressions of one large AI packaging expansion.

My variant view is:

BESI can begin its earnings inflection before HBM, but it can sustain a premium valuation only if the shift from monolithic chips to assembled systems creates repeated production cycles rather than one front-loaded capacity build.

The quarter strengthened the first half.

The second half remains open.

What Could It Be Worth?

The key assumptions should move together.

If hybrid bonding becomes a standard production process, revenue grows faster, margins rise and the stock may retain a high multiple because the opportunity still looks early.

If the current order strength reflects capacity pulled forward, revenue slows, margins fall and the multiple contracts at the same time.

I use roughly €1.02 billion of 2026 revenue as the starting point. I stress-tested these assumptions through an adversarial review arguing each variable from both sides, twice and the framework below reflects where the numbers settled.

Bear: The Capacity Cycle Wins

Logic and CoWoS-like spending peak before HBM reaches volume. Customers digest equipment, thermal compression remains adequate for longer, and hybrid bonding stays weighted toward development systems.

Revenue reaches about €1.25 billion by 2029, but the path includes a downturn. Operating margin remains better than in past cycles, though far below the current bull expectation. The multiple falls because earnings are judged as cyclical.

Value: about €125.

A deeper interim trough could take the shares below €80 even if the long-run process position remains intact. BESI’s stock has halved twice in the past four years. It can do so again.

Base: Several Bridges to Memory

Logic hybrid bonding grows. CoWoS-like, photonics, power management and panel packaging expand. HBM begins to contribute in 2028 or 2029, but does not dominate the model.

Revenue reaches around €1.75 billion. Operating margin approaches 49%. EPS reaches roughly €9. The market keeps a premium multiple because growth remains healthy and the next trough appears structurally higher than the last.

Value: about €320.

This is my central case.

Bull: Bonding Becomes Part of the Standard

One memory producer moves into volume production and another follows. Orders shift from individual machines to lines. Logic, CPO and panel packaging continue growing alongside memory.

Revenue reaches about €2.25 billion. Operating margin rises into the mid-50s. EPS reaches €13–14. The stock retains a low-to-mid-40s multiple because adoption still looks early.

Value: about €590.

A true high case could exceed €700 if revenue passes €2.5 billion, margins reach the high 50s and growth remains above 20% in 2029.

We are not there yet.

What I Am Watching

I would raise my assumptions if a memory producer completes volume qualification, orders move from individual systems to full-line purchases, repeat orders broaden across logic and CPO, and CoWoS-like demand remains strong without HBM.

I would also want to see book-to-bill remain above 1.0x, gross margin hold in the mid-60s as revenue scales, and operating expenses continue to rise far more slowly than sales.

I would cut the numbers if hybrid bonding remains concentrated in development through 2028, AP7 or other capacity plans are delayed, CoWoS-like customers pause after the current build, or a credible rival wins a major production qualification.

The earliest warning would be orders falling below revenue for several quarters.

That is when process strength starts losing the argument to schedule.

Between the Chips

The 4004 mattered because it put the computer on one chip.

AI is creating the opposite problem. The computer is once again spread across many components, and the hard part is making them behave like one.

That is BESI’s opportunity.

The second quarter did not prove that hybrid bonding has become a mass-production standard. It showed that the same architectural change is already producing orders across logic, optics, power and advanced assembly, including, remarkably, a pair of glasses.

I entered the quarter thinking BESI needed HBM to validate the earnings ramp.

I was wrong.

It needs HBM to reach the most ambitious outcomes. It does not need HBM to begin growing.

At €232, the market is paying for the base case and getting the memory decision as an option. I would take that trade, the distributed demand evidence from this quarter makes the option worth more than the market is pricing, and the SK Hynix verdict lands within two quarters. But I would size the position for the cycle, not the thesis. The interim path in the bear case runs through €150 or lower, and no qualification flywheel spins fast enough to outrun an order downturn.

BESI may be gaining control over the process.

Its customers still control the clock.

$BESI

Subscribe now

• Intel’s CPU engine is working again. DCAI revenue rose 59%, operating margin approached 40%, and better yields turned additional factory output into unusually strong profit growth.

• TSMC’s capacity discipline is creating demand for alternatives. Customers do not need Intel to surpass TSMC; they need a credible second source that reduces concentration and secures capacity.

• Technology proof is no longer enough. 18A, 14A and EMIB-T show meaningful progress, but major external contracts will determine whether Intel has built a real foundry business—or merely a better internal manufacturing system.

From Intel CFO David Zinsner’s prepared remarks on the second quarter earnings call:

Q2 revenue of $16.1 billion was up 19% sequentially and 25% year on year, exceeding the top end of our guided range. This was $1.8 billion above the midpoint of our guide, driven by strength across the board... Despite exceeding our expectations for wafer outs in the quarter, strengthening demand continues to outstrip our growing supply.

Intel’s quarter was plainly excellent: non-GAAP EPS was $0.42, DCAI revenue rose 59%, and the segment’s operating margin reached almost 40%. The real surprise, though, was not demand. Intel already knew customers wanted more server CPUs. The surprise was that its factories produced far more usable output than management expected, customers absorbed all of it, and supply remained tight.

My read is that Intel now has two distinct theses. The near-term thesis is a CPU recovery: AI infrastructure demand, higher core counts, pricing and improving yields are generating earnings and loading the fabs. The long-term thesis is a second-source opportunity: TSMC’s cautious capacity expansion, customer efforts to diversify, and the collapse of the old boundaries between compute, memory and packaging may create a place for another integrated manufacturing system.

Eighteen-A and 14A show whether Intel can build that system. External contracts will show whether anyone else wants to use it.

The CPU Buys Time

The most revealing number was not DCAI revenue. It was DCAI operating profit.

Sources: Intel earnings materials and the prior Q1 analysis.

Intel added roughly $1.2 billion of sequential DCAI revenue and nearly $1 billion of segment profit. That sort of fall-through will not repeat every quarter, but it shows what drove this one: more units, higher core counts, richer mix, pricing and manufacturing improvement.

Agentic AI belongs in this discussion, but not as a made-up revenue bridge. Intel cannot tell us how much of the 59% growth came from agents, and neither can I. Today’s drivers were hyperscaler construction, enterprise demand, scarcity and better factory output. The agentic thesis explains why CPU demand may remain stronger for longer.

What makes that thesis more credible is that TSMC is now making the same argument. C.C. Wei described AI moving from “query mode” to “command and action mode,” driving another step-up in token consumption and compute. He also acknowledged that CPUs are becoming more important in AI data centres, even though TSMC cannot yet separate AI-related server CPUs in its reporting.

That is the right distinction: agentic demand is visible in system architecture before it is visible in segment accounting.

For Intel, CPU demand does more than increase revenue. It trains the factories. Eighteen-A output rose more than 50% sequentially and came in around 25% above target. The cost of the primary Panther Lake SKU fell roughly 50% year-to-date, with another 20% reduction targeted this year. Better yields and shorter cycle times also helped narrow the Foundry loss.

The loop is straightforward:

Intel spent a decade running that loop backwards. Late processes weakened products; weaker products reduced utilisation; low utilisation made the fabs more expensive. Q2 was the clearest evidence yet that integration can transmit improvement again.

But CPUs are still the funding mechanism, not the final destination.

The TSMC Brake Creates a Market

For years, using TSMC was the low-risk choice. It had the best process, the deepest production knowledge and the best record of bringing difficult chips to volume. Moving a flagship design to Intel or Samsung meant porting expense, schedule risk and the possibility of an ugly yield surprise.

That calculation changes when TSMC capacity itself becomes uncertain.

TSMC’s capital spending was broadly flat after its 2021 jump and declined in 2023 and 2024, just as hyperscaler AI spending accelerated. TSMC is now raising 2026 capex toward $52–56 billion and adding N3 capacity in Taiwan, Arizona and Japan, but much of that production arrives in 2027–28. Wei has been candid about the reason for caution: committing tens of billions today against uncertain demand several years out can become “a big disaster” if the cycle breaks.

TSMC is behaving rationally. The risk does not disappear; it moves to customers.

A hyperscaler unable to obtain enough silicon delays services and forgoes revenue. The old choice was certain TSMC execution versus uncertain second-source execution. The new choice is TSMC execution with uncertain capacity versus a less proven supplier that can reserve capacity.

That is why customer behaviour beyond TSMC matters. Tesla has disclosed a “landmark” Samsung deal to facilitate future chip production, while Samsung and AMD signed an MOU to expand collaboration across next-generation AI memory and computing. These are not equal forms of proof—a production arrangement matters more than an exploratory MOU—but they show customers widening their manufacturing and technology options.

Intel does not need customers to conclude it is better than TSMC. It needs them to conclude that dependence on TSMC alone is more dangerous than helping Intel become viable.

That is a much lower bar. It is also a real market.

When Capacity Is Not Enough

If the problem were merely a shortage of leading-edge wafers, Intel’s opportunity would be a conventional second-source story.

AI is creating a harder problem. Power has crossed conventional cooling capacity; HBM bandwidth is straining standard interfaces; and accelerator packages are outgrowing the geometry of traditional interposers. These constraints interact: adding memory enlarges the package, more routing raises power, and more power forces cooling changes that affect the chip and package design. The bottleneck is moving from the individual component to the boundary between components.

This is why customers may need more than spare wafers. They may need another integrated path across process technology, custom silicon, packaging, power delivery and high-volume manufacturing.

Intel has relevant pieces: x86 CPUs, purpose-built silicon, networking IP, EMIB-T, 18A, 14A and a large US factory network. The purpose-built silicon business nearly tripled year over year and is approaching a $2 billion annual run rate; management sees a route to $4 billion. EMIB-T has a growing backlog and is aimed at customer ramps in 2027.

That does not mean Intel controls the system. Nvidia still defines much of the high-end AI architecture. TSMC owns the deepest external manufacturing and packaging learning loop. Intel has breadth, but breadth can be either an advantage or an expensive collection of businesses.

The realistic pitch is not “we are better than TSMC.” It is “use us where dependence on one process and packaging platform has become too concentrated.” A second platform with adequate performance and guaranteed allocation can be more valuable than waiting in TSMC’s queue. Yet EMIB-T remains option value until Intel proves volume yields; the warpage observed on its large panel test vehicle goes directly to the company’s historic weakness in turning technical demonstrations into reliable manufacturing.

This is the structural opening: TSMC’s restraint creates demand for alternatives at the same time the three walls increase the value of an integrated alternative.

Now Intel must earn it.

Process Proof, Customer Proof

Intel’s disclosures were precise where the evidence was strongest and vague where it was weakest.

Sources: Intel earnings release and call.

On 18A, management gave output, yield and cost figures. On 14A, it gave PDK timing, risk-production dates and a high-volume ramp target. When Bank of America’s Vivek Arya asked when customer enthusiasm would become announcements, management returned to those process milestones. When Melius’s Ben Reitzes asked whether higher capex reflected hard external orders, the answer broadened to confidence across Intel’s businesses.

This is not necessarily evasion; foundry agreements can remain confidential for years. It does tell us where the burden of proof sits.

Eighteen-A and 14A are evidence that Intel may possess competitive technology. A signed, high-volume external customer is evidence that Intel has a foundry product.

The distinction matters because Intel is spending ahead of that proof. The company now expects more than $20 billion of 2026 capex and materially more in 2027. CPU demand makes the near-term investment rational. The long-term return requires outside volume, packaging economics and customer commitments that prevent Intel from carrying the entire risk itself.

Intel therefore has two clocks.

The CPU clock is running quickly: demand, earnings, yields and internal utilisation are improving.

The second-source clock has barely started: the industry opening is clearer, but Intel has not yet shown that major customers will design around 14A or trust EMIB-T at volume.

The CPU keeps Intel in the race.

TSMC’s brake and the three walls create the opening.

External contracts will decide whether the opening belongs to Intel.

Subscribe now

• Demand is no longer the debate. Gas and Electrification orders, pricing and backlog confirm that the power shortage is real and global.

• The conversion arrives later. Today’s spending supports equipment revenue from 2028–2030, while the higher-margin service economics follow after delivery.

• Margins now matter more than orders. The next phase of the thesis depends on GEV expanding output, fixing Wind and preserving high-teen to low-20s profitability.

“We now see further opportunity to serve this growing demand with 30 gigawatts of annual output in 2030…all funded by customer down payments.”

— Scott Strazik

Three months ago, after GE Vernova’s first-quarter results, we wrote down a test for the thesis:

“The thesis strengthens if Electrification grows faster than Power while services backlog compounds. It weakens if orders stay high but margins plateau—that would signal scarcity without conversion, which is a cycle, not a franchise.”

On July 22, GEV reported $24.2 billion of orders, up 88% organically, and left its adjusted EBITDA margin guidance at 12% to 14%. The stock fell 8.7%.

By our own test, the thesis appeared to weaken.

I do not think it did. I think the test was wrong.

We assumed that the orders being signed today and the margins being reported today belonged to the same economic period. They do not. The gas equipment GEV is now pricing more than 20% above its late-2025 order book will largely reach revenue from 2028 through 2030. The revenue reported this quarter came from older contracts, while the labor, machinery, inventory, research spending, and supplier commitments needed to deliver the new backlog are being paid for now.

We asked a multi-year mechanism to prove itself in a ninety-day result.

That does not mean the financial miss was irrelevant. It means the order book has finished doing its first job. It proved that demand was real. From here, the stock will trade on whether those orders convert into revenue, margins, and cash at the pace investors already assume.

The queue created the rerating. Conversion will determine what comes next.

The Test That Could Not Pass

The headline result looked weaker than the business underneath it.

Revenue of $11.1 billion beat expectations, but adjusted EBITDA of $1.25 billion came in slightly below consensus. The company raised annual revenue guidance by $1 billion and almost doubled free cash flow guidance, yet kept its company-wide margin range unchanged.

That was the table investors read, and it was the right table to read. GEV is valued on the assumption that a large backlog will become a much more profitable company. More orders without a higher margin outlook do not automatically raise that future value.

Yet the segment results did not show a failure to convert. Power’s EBITDA margin reached 18.8%, up 240 basis points. Electrification’s organic margin reached 19.1%, up 700 basis points. The two businesses at the center of the thesis were already operating around the level that GEV targets for the whole company.

The gap sat elsewhere. Wind lost $275 million in the quarter, corporate costs remained meaningful, and equipment grew much faster than services. Equipment revenue rose 32% to $6.5 billion, while services grew 10% to $4.6 billion. Equipment increased from 53.7% of revenue a year ago to 58.2%.

That mix does not mean equipment is unattractive. Power and Electrification show that well-priced equipment can earn high-teen margins. It means that the higher-margin service revenue created by these shipments arrives later, while the cost of producing the equipment arrives first.

The core franchises converted. The portfolio did not.

This is a more demanding conclusion than blaming the market for being short-term. The market is already paying GEV for much of the future service value. It is entitled to ask when that value becomes visible in consolidated earnings.

Can GEV Scale Scarcity?

The new question raised by the quarter is not whether demand remains strong. It is whether GEV can double output without building away the shortage that created its pricing power.

Industrial booms usually contain the mechanism that ends them. Demand rises, prices increase, manufacturers add capacity, supply catches up, and returns return to normal. The producer benefits from the shortage and then helps solve it.

GEV is trying to reverse that sequence.

Customers are reserving capacity before GEV builds it. They are accepting higher prices and paying deposits that help fund machinery, inventory, and supplier commitments. The company then adds output inside an existing factory network rather than building speculative plants and hoping the demand arrives.

Gas production is moving from roughly 15 gigawatts annually to a 20-gigawatt rate now, 24 gigawatts in 2028, and 30 gigawatts in 2030. GEV expects to be mostly sold out through 2030 and to have more than half of 2031 capacity contracted by the end of this year.

The commercial evidence remains unusually strong. GEV signed 20 gigawatts of new gas contracts and reservations during the quarter while shipping 3 gigawatts. Contracted capacity rose from 100 to 116 gigawatts, and management raised its year-end target to at least 125 gigawatts. First-half gas equipment orders were priced more than 20% above fourth-quarter 2025 orders.

The demand is also broader than the easy AI framing. Roughly 80% of contracted gas capacity comes from traditional power customers, with data centers accounting for 20%. The order book spans about 100 customers across 26 countries. Management discussed demand in Taiwan, Saudi Arabia, Mexico, Qatar, Southeast Asia, Brazil, and the United States. AI accelerated the shortage; it did not create the entire need.

The capacity plan may therefore strengthen the loop rather than end it. Much of the output is sold before the capital is committed, customers provide part of the funding, and every turbine delivered enlarges the future service base.

Still, the risk has changed. It is no longer mainly demand discovery. It is execution.

GEV must train workers, secure castings and forgings years ahead, preserve quality as factory output rises, and coordinate with engineering contractors, pipelines, permits, and grid connections. A reservation only becomes valuable when the rest of the power project is ready.

The move from slot reservations into firm backlog is therefore more important than another record order headline. GEV converted 10 gigawatts of reservations into orders during the quarter, lifting firm gas backlog from 44 to 53 gigawatts. Management expects firm backlog to exceed reserved capacity as more projects mature.

That is the first test of whether GEV is scaling scarcity or simply collecting options on a distant future.

When Orders Stop Telling the Truth

Orders have been the clearest evidence of GEV’s change since the spin. Quarterly orders rose from $9.7 billion in early 2024 to $24.2 billion in the latest quarter. Total RPO increased from $116 billion to $176 billion, while equipment RPO more than doubled.

That metric will become harder to read.

Once 2030 is mostly sold and 2031 begins filling, GEV has less near-term capacity left to book. Orders may decline because there are fewer available slots, not because customers have disappeared. Reported order growth can weaken at the same time that shipments accelerate, backlog pricing improves, and the installed fleet expands.

This matters because the next bearish argument is easy to anticipate. At some point, quarterly gas orders will fall from today’s exceptional level. That decline will be presented as evidence that the cycle has peaked.

Sometimes it will be.

But the correct dashboard is changing. The questions should become: are contracted gigawatts rising after accounting for shipments? Are reservations converting into firm orders? Is pricing per kilowatt holding? Are backlog margins intact? Is the production ramp on schedule? Are Power and Electrification retaining the economics as volume rises?

The order book now has an asymmetric role. Another record quarter may add little because strong demand is accepted. A sharp slowdown in commitments, weaker pricing, or stalled conversions would matter a great deal because those developments would challenge the duration supporting the valuation.

Orders have become necessary for maintaining the thesis, but insufficient for raising the stock.

The market has already paid GEV for winning the queue. From here, it will pay or punish the company for converting it.

Honest in Both Directions

The call contained an unusual contrast.

GEV generated $5.1 billion of free cash flow in the quarter and $9.9 billion in the first half. Ken Parks could have allowed investors to treat that as a new earnings run rate. Instead, he explained that customer deposits and slot reservation payments drove much of the result, and that second-half cash generation would be far lower.

Scott Strazik then spent the call describing 2030 production, 2031 contracting, and service capacity needed in the middle of the next decade. He did not offset the caution on cash by manufacturing a near-term margin reset.

A more promotional management team could have narrowed the margin range to 13% to 14%, allowed the market to capitalize the first-half cash result, and dealt with the consequences later. GEV did neither.

That does not mean management communicated poorly. It means the company itself is operating on the longer clock. It appears more interested in being right about the capacity plan than in producing the cleanest possible quarterly reaction.

The 8.7% decline was partly the price of that choice.

The guidance arithmetic explains the disappointment. Revenue guidance increased by $1 billion, while the midpoint of the margin range stayed near 13%. That implies roughly $130 million of EBITDA on the additional revenue, an incremental margin close to the corporate average. A stock valued on fast margin expansion wanted more.

Yet the guidance table describes revenue being recognized in 2026, not the economics being contracted for 2028 through 2030. It can be accurate about this year and incomplete about the business being built inside the backlog.

Both clocks are real.

The Core Converted. The Portfolio Did Not.

Electrification now deserves to be treated as a second earnings engine rather than a supporting business.

Organic revenue grew 29%, organic EBITDA margin reached 19.1%, and equipment backlog passed $40 billion. Data-center orders exceeded $5 billion in the first half, more than twice the total booked during all of 2025. None of that requires solid-state transformers or medium-voltage uninterruptible power systems to succeed. Those products remain future possibilities rather than current earnings.

Power also showed that higher output need not destroy margin. Revenue grew 14%, EBITDA rose 31%, and the segment margin reached 18.8% while GEV continued hiring, adding machinery, and spending on capacity and research.

Wind remains the counterexample inside the same company.

First-half Wind losses reached $657 million, yet management retained its full-year expectation of roughly $400 million of losses. That requires about $257 million of positive EBITDA in the second half. Third-quarter Wind EBITDA is expected to be around breakeven, leaving a demanding fourth-quarter bridge.

Management has an explanation: more second-half deliveries, better tariff protection, improving Onshore service profit, and lower Offshore project costs. The plan is plausible. It is not proven.

Wind matters beyond its direct earnings drag. It shows how physical complexity can consume the value of a good market through poor contracts, policy changes, installation delays, and project costs. It is a reminder that backlog does not become shareholder value by itself.

The next quarter will therefore test two types of credibility at once. Gas and Electrification must show that GEV can expand capacity while retaining price. Wind must show that management can contain the part of the portfolio where physical execution has repeatedly overwhelmed the plan.

The Cost of the Long Clock

The business thesis strengthened during the quarter. Customer diversity is broader than expected, gas pricing is better, reservation conversion is progressing, and Power and Electrification are already producing high-teen margins.

The stock thesis is less straightforward.

At $1,017, GEV is already valued on the assumption that the long clock works. Bloomberg’s pre-print estimates expected adjusted EBITDA margin to rise from about 13.7% in 2026 to 18% in 2027 and 21.5% in 2028. The Street is not waiting to discover conversion. It is underwriting a fast version of it.

My base case sits between that optimism and the view that equipment mix delays the margin story until well after 2029.

The base case assumes GEV broadly meets its long-term plan, Wind becomes immaterial, and Power and Electrification convert higher-priced backlog near low-20s margins. Even then, the expected return from $1,017 is moderate because investors have paid in advance for much of the improvement.

The bear case does not require an electricity-demand collapse. It requires conversion at ordinary industrial margins. The bull case requires evidence that services, Electrification, and customer-funded capacity have changed the earnings quality of the company, rather than simply enlarged its equipment business.

That is why margin matters more than another order record. Revenue tells us how large the cycle becomes. Margin tells us what kind of company emerges from it.

In April, we thought strong orders without faster margin expansion would tell us that scarcity was failing to convert. The second quarter showed that conversion is not absent. It is arriving on a different schedule.

GEV may have found a way to add supply without immediately destroying scarcity: sell the capacity first, use customer cash to help build it, and turn each shipment into future service demand. But the spending comes first, the equipment revenue follows, and the full service economics arrive later still.

Management is building for 2030.

Every ninety days, the stock asks to see it now.

Subscribe now