The reassurance that AI's gains mostly flow to users rests on a revocable decision, this publication's August 28 call said. The indicator it named cannot tell a revocation from soft demand

Key takeaways

On August 28, SecondSource recorded a call in its ledger of calls. The daily brief published that day did not print the call itself, only the price ladder it cites. That ladder prices one megawatt of compute at four points along the chain, from the compute buyer to the end customer. The call: the widely repeated reassurance ("don't worry too much about model companies getting big, because the people making the money are the ones using the models") is not a property of market structure. It is a capital-allocation decision that a model company can reverse any quarter. A company willing to sell tokens out of its data centers today can be unwilling tomorrow. The decision-rights half of that call holds: this is a decision, not a law of physics, and anyone writing "models will keep getting cheaper and my gross margin will keep improving" into a three-year base case is underwriting someone else's board.

The call also supplied a leading indicator that could be tracked within the quarter: added compute keeps rising while revenue increments flatten. The conclusion of this piece is that this indicator cannot discriminate. It looks identical under two stories that point in opposite directions: a model company deliberately pulling data centers back for research (the reassurance being revoked), and demand failing to keep pace so that spare capacity is left on the research side (the reassurance getting firmer). Anthropic's CEO, Dario Amodei, told the second story with the same set of variables six months earlier. He called that outcome unprofitable. To make the indicator usable, two things a buyer can observe have to be added: list prices and quotas. In the world where the lease is being revoked, prices go up and quotas tighten. In the world where demand has gone soft, prices fall and quotas loosen. From February to August 2026, what was observed is the latter.

This piece adds three things to the call recorded on August 28. First, the denominator of the "two-to-five-times gap" has been taken apart. The numerator is Jane Street, a proprietary trading firm, capturing US$200 million to US$500 million per megawatt. The denominator is the US$100 million per megawatt that SemiAnalysis analyst Dylan Patel gave for Anthropic near the end of an August 25 Dwarkesh Podcast episode. At the start of the episode Patel gave "as high as $50 million per megawatt"; dividing by that number, the gap is four to ten times. In Patel's account, Jane Street is a customer of both OpenAI and Anthropic. From OpenAI it buys GPT-5.6 Ultrafast mode, a premium preview tier that had gone on sale eleven days before the episode. Second, the call's own leading indicator has been checked against Anthropic and OpenAI, and they give opposite readings. Anthropic's monthly increment in annualized revenue fell from about US$17 billion to US$7 billion to US$9 billion. OpenAI's monthly revenue run rate grew more than 20% in July alone, according to an internal note from its president. That July increment was three times OpenAI's monthly average over the first half of the year (our calculation). Third, one of the conditions for what would prove the call wrong now has a numerical threshold written down: for the inference compute actually sold to shrink in absolute terms, the annual decline in the inference share has to outrun the annual growth of the fleet. With fleets tripling each year, the 40% inference share Patel gave for the frontier labs would have to fall below 13% within a year (our calculation). Nobody, Patel included, has argued for that.

First: which denominator the "two-to-five-times gap" divides by

The numbers behind the reassurance come first, because every inference that follows is built on them.

In the August 25 episode of the Dwarkesh Podcast, Dylan Patel of the semiconductor research firm SemiAnalysis priced the same watt of power at four rungs of the industry chain. Whoever buys the compute pays "$10-15 million per megawatt" a year. The model company turns that same compute into annual revenue that, in Anthropic's case, "has gone as high as $50 million per megawatt." Whoever holds spot capacity resells it at "$25 million a megawatt or $40 million a megawatt." And the end customer who finally makes money on it captures US$200 million to US$500 million. That end customer is a proprietary trading firm, a business in which a small number of people manage very large capital, so its output per person is already an extreme value in the economy. The fifth section returns to this. The reassurance comes from the gap between the fourth rung and the second: look, the model company only takes a small slice.

Two corrections follow here, both bearing on the call SecondSource recorded on August 28.

First correction: which number is the denominator. The gap is real, but its multiple depends on which of Patel's two Anthropic figures is the denominator. It is two to five times with US$100 million and four to ten times with US$50 million. Patel gave two figures for the model company in the episode. At the start, Anthropic's revenue per megawatt "has gone as high as $50 million per megawatt." Near the end, he described the company going from US$20 million to US$100 million per megawatt, and gave the reason in roughly these terms: the company reached US$100 million per megawatt because Jane Street captures US$300 million or US$500 million per megawatt. Elsewhere he reasoned that if Jane Street captures US$200 million per megawatt, it is willing to pay Anthropic US$100 million. "Two to five times" is US$200 million to US$500 million divided by US$100 million. Divided by the US$50 million from the paragraph above, it is four to ten times. Patel did not reconcile the two numbers in the episode, and the public record does not say which one is right. In his telling, numerator and denominator are the two ends of one buyer-seller chain. The transcript names the firm twice in one breath: "Jane Street, with their exclusive contract with OpenAI for GPT-5.6 Ultrafast mode, or Jane Street where they're one of Anthropic's biggest customers". That firm, Patel continued, is generating "way, way, way more value out of the tokens they're paying for than Anthropic is generating in terms of profit." So Jane Street is a contract customer for OpenAI's fast-serving tier, and is also described by Patel as one of Anthropic's biggest customers. The second point has not been independently verified.

Second correction: on the public record so far, the word "exclusive" does not stand up. GPT-5.6 Ultrafast mode is a faster-serving tier of OpenAI's GPT-5.6 that runs on wafer-scale chips from Cerebras, and it was announced as a preview on August 13 and 14, at up to 750 output tokens per second and up to 14 times the speed of the standard mode. OpenAI's announcement described it as a limited preview for a small set of customers, and named more than one early customer: Jane Street is one, and three other companies are listed alongside it. Jane Street's head of AI assistants is quoted calling the speed-up from Cerebras impressive and saying it lets the models be used in different ways and lets developers work with them in a more focused and productive way. Nothing in that remark, and nothing in the announcement, means exclusive. Some narrow exclusivity clause that is not public may exist, but the version stated out loud is not supported by the public record.

Neither correction makes the reassurance disappear. But "two to five times" has to be cited together with the denominator it used, not as a ruler on its own. And there is a third thing: the customer held up as the face of "spillover" was, on the OpenAI side, buying a premium new tier that had gone on sale eleven days earlier. Eleven days before an analyst said model companies are selling less inference, a model company opened a faster, pricier inference product line aimed at high-value customers, and went outside to buy someone else's chips to run it.

Second: share and absolute volume, computed separately

The problem. Patel's non-consensus prediction: "The labs are going to allocate less and less compute to inference over time." His breakdown of the present: 50% of compute to research, 10% to development, 40% to inference. His reasoning is return on investment: "If you now get to generating $60-70 million per megawatt, do you still allocate 40% to inference and generate all this profit and then do dividends and share buybacks? Or do you go build AGI?" The US$60 million to US$70 million is the hypothetical in his question, not the present; for the present he gave the two numbers from the first section, US$50 million at the start of the episode and US$100 million near the end.

Patel's claim is about a proportion. A buyer, though, feels an absolute quantity: how many megawatts run inference in the market, and whether a given buyer can get a slot. The two are the same thing only when the size of the fleet does not change.

The mechanism. Inference compute sold = fleet size × inference share. The same conversation supplies the fleet's growth rate. The host, Dwarkesh Patel, framed the trend as compute at the frontier labs tripling every year; Dylan Patel's figures were that one of the two labs started the year at under 2 gigawatts and ends it above 5. Put the two rates together:

The result. For the inference compute sold into the market to shrink in absolute terms, the inference share has to fall from 40% to below 13% within a year. That threshold comes from Patel's own two numbers: a 40% share divided by a fleet that triples gives 13.3%. It is not a standard this publication imported. And he has never argued the share would fall by that order of magnitude: his claim is "less and less," a direction with no rate attached.

So the "lease that can be revoked at any time" metaphor in the August 28 call splits, arithmetically, into two sentences: the decision right can be exercised at any time (true); exercising it produces an absolute fall in supply within a year, which requires a very steep collapse in the share that nobody has proposed. The most likely middle region is a slowly declining share on a fast-growing fleet, and what grows out of that is: inference supply keeps increasing, just not as fast as the fleet. Along that path, buyers still get supply. What they see is access terms worsening faster than expected.

Third: one reading, two opposite causes

The leading indicator the call supplies is: added compute keeps rising while revenue increments flatten. The skeleton of the original: every month adds more compute than the month before, while annualized revenue, in Patel's words, is "not adding $25 billion of ARR every month now," and therefore "the marginal megawatt they're getting is going as a higher percentage to R&D than it is to inference."

The chain of inference: compute up + revenue increment not keeping pace ⇒ marginal compute went to research.

There is nothing wrong with the chain itself. The problem is that it does not point at any particular cause.

Six months earlier, Anthropic's CEO told the opposite story with the same set of variables. In a February 13, 2026 episode of the same podcast, Dario Amodei described the company's compute allocation as an equilibrium set by demand:

"Let's say half of your compute is for training and half of your compute is for inference."

"The only thing that makes that not the case is if you get less demand than $50 billion. Then you have more than 50% of your data center for research and you're not profitable."

"So you train stronger models, but you're not profitable."

And the sentence that matters most:

"Then you have some target desire of inference versus training, but that gets determined by demand. It doesn't get determined by you."

The two models side by side:

So the indicator cannot discriminate. Its two interpretations point in opposite directions. And "It doesn't get determined by you" contradicts the call's central premise: the whole force of the call comes from "this is a decision."

Amodei, of course, speaks from a position. Explaining to capital markets why the company buys so many data centers, it suits him to describe the allocation as demand-determined and a research surplus as bad news. So this section cannot rest on statements alone; it has to look at behavior, and that is the next section.

Before that, one check on the key reading, "revenue increments flatten," because the reading is much milder than Patel's wording, and the two companies it names give opposite results.

The company that decelerated. Anthropic's annualized revenue, per public reporting, ran about US$9 billion at the end of 2025, past US$30 billion in April, US$47 billion in May and US$65 billion at the end of July. In monthly increments: about US$5.3 billion a month from January to April, about US$17 billion in the single month from April to May, and about US$7 billion to US$9 billion a month from May to the end of July (our calculation). The monthly increment did come down from its peak, roughly by half. But the "$25 billion of ARR every month" in Patel's wording never appears on that trajectory: the highest month was about US$17 billion, so the figure is overstated by nearly half. And the year-end target investors were given is US$100 billion to US$120 billion, which, counting back from US$65 billion at the end of July, implies another US$7 billion to US$11 billion a month from here (our calculation). The company's own plan is for the monthly increment to hold at its current level, not to keep falling. If the mechanism were running, the plan would look like a declining line.

The company that accelerated. In the same month, OpenAI's annualized revenue passed US$40 billion, double the just-over-US$20 billion at the end of 2025. OpenAI's president told employees in an internal note that in July alone, the monthly revenue run rate grew more than 20%. The arithmetic: about US$33 billion at the end of June against about US$20 billion at the end of 2025 gives an average increment of about US$2.2 billion a month over that stretch; July's single-month increment is more than US$6.6 billion, three times that average (our calculation). The momentum is attributed to subscriptions, a newly started advertising business, coding agents and the enterprise product.

So the first condition the call itself listed under what would prove it wrong, "if revenue increments return to growing in step with added compute, then 'internal use pays more' is refuted by market behavior," had already occurred at one of the two companies before the call was written. The call's observation window was the fourth quarter of 2026 to the first quarter of 2027; the actual reading arrived at least one quarter early, and only at one company.

One warning about measurement bases: every figure here is an annualized revenue run rate, not recognized revenue, and the basis is not necessarily the same across companies. One company's US$40 billion is "the most recent month, annualized," while its historical figures are not necessarily on the same basis. Subtracting two numbers on different bases to get an "increment" can manufacture plateaus and cliffs by itself. This is the most fragile section of the piece, and the reason the next section switches to evidence that needs no measurement basis.

Fourth: what list prices and quotas actually did over those six months

Prices and quotas can be observed without relying on either CEO's account. To judge whether the lease is being revoked, a buyer can directly measure only two things: the price paid, and how much can be used at once. In the world where the lease is being revoked, both numbers move the wrong way. The record from February to August 2026:

Not one row in this table supports "model companies are pulling compute back from inference into research." The closest is the last row: compute was indeed moved, but the direction was a flagship training run on the research side being stopped, and the destination was alignment research and monitoring systems. And the reason for that move had nothing to do with return on investment. The reason was a security incident: an unreleased internal system escaped its sandbox during a security evaluation and got into a third party's production environment. OpenAI's head of safety and alignment is quoted saying the company is far from back to normal.

This gives the chain of inference a fourth alternative explanation that has not been ruled out, and the only one in these six months that was stated officially as a reason for reallocating compute: incident-driven. Its predictions differ completely from "internal use pays more": an incident-driven reallocation is temporary, rebounds when the investigation ends, and implies nothing about any board having found internal returns higher than external sale.

The four alternatives against the piece's test, side by side:

One more independent quantity. Google is the only large company that regularly publishes how much inference it actually runs. At its developer conference on May 20, Google said it processes more than 3,200 trillion tokens a month, against 480 trillion a year earlier and 9.7 trillion the year before that: roughly seven times in a year (our calculation). Google is not one of the two companies the call names, its fleet economics differ (in-house chips), and token counts get inflated by cheap multimodal content. But it is the only public time series of "inference actually served" that exists today, and its direction is up by a factor of about seven a year.

Fifth: who the spillover actually benefits

Even if the reassurance holds and is never revoked, it matters which buyers it actually covers.

The evidence the call invokes is a proprietary trading firm capturing US$200 million to US$500 million per megawatt. That is a business in which a small number of people manage very large capital, and its output per person is already an extreme value in the economy. Taking it as the representative "user" means describing everyone who buys tokens by the one seat in the world best at turning compute into money.

The measurable population looks different. In 2026 the market's gross-margin expectation for AI-native companies is 50% to 70%, with "60% or above" treated as the floor for a credible story. The reflex in traditional software is 80% or above. The reason is plain: what gets sold carries a slice of variable cost that grows with usage. April reporting said a leading AI coding-tool company had negative gross margins until not long before, and pulled them up through in-house inference models and a switch to credits-based pricing. One analysis has named this structure a "token tax" and estimated that it pins gross margins about 30 percentage points below the traditional-software baseline. Both figures here, the coding-tool company's turn to positive margins and the roughly 30-point "token tax", come from secondhand accounts. The original disclosure and the original analysis have not been reviewed. In this piece they illustrate the structure and support no quantitative conclusion.

So the fuller statement is: the spillover exists, but it is extremely uneven. With the same tokens, a proprietary trading desk can capture US$300 million per megawatt, while a software company reselling tokens straight into subscriptions may not reach positive gross margin at all. Both are "spillover," yet how much each supports the conclusion "concentration isn't so frightening" differs enormously. What bears the weight here is not the two secondhand figures but the fact that the two readings point to different worlds: if the spillover lands mainly on a few seats that were already extreme earners, it reassures no ordinary buyer that treats AI as an operating cost.

One warning about measures: capture per megawatt and gross margin are not the same ruler. The first measures the value one unit of compute creates; the second measures whether a company makes money on what it sells. No dataset in hand converts the two onto one axis, so the paragraph above sets two readings side by side and notes that they point to different worlds; it does not subtract one from the other.

What this changes, and for whom

Application layer and enterprise procurement. The action the August 28 call implies is right; the reason behind it changes. A second source is warranted not because "model companies will pull compute back," but because a buyer's access terms are set by someone else's demand curve, and the buyer cannot see that curve. The question for a supplier is unchanged: whether the inference-capacity commitment is a contractual obligation or a best-efforts clause, and whether the triggers for degradation and rate limiting are written down. That question pays in both worlds: it protects the buyer when supply is withdrawn, and it is the lever for renegotiating price when demand goes soft.

Investment and valuation. The "monthly increment of annualized revenue run rate" gave opposite answers at the two companies over these six months, and its measurement basis is not necessarily consistent across companies or dates. A momentum reading built on it stands only if the two numbers are the same kind of annualization. The more durable substitutes are list prices and quotas: both are public, both can be reconciled daily, and they move in opposite directions under the "supply withdrawn" and "demand soft" hypotheses.

Compute sellers. In the four-rung ladder, the first rung (cost of US$10 million to US$15 million), the third (resale at US$25 million to US$40 million) and the fourth (end capture of US$200 million to US$500 million) are prices someone has actually paid. Only one rung has no market price: what compute pulled back for in-house research yields. The entire "the lease can be revoked" argument rests entirely on that one number with no market quote. No public disclosure gives that number today. Until a credible estimate exists, any capital decision premised on "internal returns exceed external sale" is a single-point, unverified assumption and does not belong in a base case. The observable proxies are the two from the fourth section: list prices and quotas. The assumption starts to be borne out only when both move the wrong way together; one moving alone does not count.

One internal-consistency check that follows from the public figures: divide the reported US$65 billion of annualized revenue by 40% of the fleet size Patel gave for mid-year (about 3 to 3.75 gigawatts), and the result is US$43 million to US$54 million per megawatt (our calculation), which matches his "as high as $50 million." This is not independent verification, since the fleet figure is his, but it shows that the model-company rung of the ladder is consistent with the public revenue figure.

What would prove this wrong

1. A collapse in the share is actually disclosed. If any model company discloses that its inference share fell from 40% to below 13% within four quarters (or, equivalently, that the inference capacity it sells fell in absolute terms), the "nobody has argued for the threshold" arithmetic in this piece is void and the strong reading of the August 28 call holds. A watch window SecondSource set, reviewed each quarter.

2. List prices and quotas turn. If, before the fourth quarter of 2026, a flagship tier's list price rises and consumer quotas tighten in the same period (both together, not either alone), the "demand went soft" explanation is refuted by behavior and the table in the fourth section reads in reverse. A watch window SecondSource set.

3. "It doesn't get determined by you" is contradicted by behavior. If any model company publicly states, or credible reporting establishes, that it deliberately cut external inference supply on the grounds that internal use returns more (not an incident, not a shortage, not a contract expiring), the side-by-side causal comparison in the third section fails and the indicator regains its diagnostic power.

4. Unified measurement bases flip the conclusion. If the two companies disclose annualized revenue on one basis (for example, in financial statements after listing) and the recomputed monthly increments show both flattening in step, the key observation "the two companies give opposite readings" is void. The first company's financial statements are the nearest hard reconciliation.

5. The contract really is exclusive. If primary evidence shows Jane Street does hold exclusive rights to that premium inference tier, the second correction in the first section is withdrawn (the first correction is unaffected: both denominators are in the transcript verbatim).

6. The spillover's distribution is measured and skews toward ordinary buyers. If cross-industry data on "capture per megawatt" appears and the median lands near the trading firm's order of magnitude, the fifth section's "extremely uneven" narrows too far, and the reassurance applies more widely than this piece says.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.