“Model companies are all buying share at a loss, and nobody dares raise prices”: does Bill Gurley's claim hold? Checking the receipts

① Who said it, and what was said

On July 31, 2025, Bill Gurley, a retired partner at the venture firm Benchmark, described model pricing on the venture podcast BG2 in a sentence with no exceptions in it: "no one's pricing the cost because they're pricing to take market share". Gurley added a rumor: "There's rumors of even some of the best known brands in AI having negative gross margin" (BG2 Pod). From there Gurley built a valve argument: price is what actually holds demand in check, and "the thing that throttles demand is price, but no one here is raising price".

How far the claim travelled. Among the 529 voices SecondSource tracks, this line came back 19 times in the July 31 round, more than any other line in the same window: 5 times in the daily brief, once in a deep dive, once on SecondSource's list of events to watch, and in 12 posts on X. A second named voice on the same list makes the same skeptical case. On July 11, Chamath Palihapitiya relayed the view of his company's CTO that the return on AI capital spending is overstated and the market will settle that account sooner or later. One more note on the count: this line of argument appeared another 5 times in SecondSource's own deep dives, but those were reprints of the same trend tree, not independent spread, so they are not in the count above.

What the claim is checked against. The official pricing pages of each model vendor, from late 2025 to today, each checked directly on August 2.

Which sentence could fall. "Nobody dares raise prices." Price-increase terms written plainly on the vendors' official pricing pages overturn it.

② The receipts

Whether vendors dare to raise prices can be read off their official pricing pages, and this is the hardest receipt. Gurley named Anthropic: "Anthropic doesn't need to throttle. They just raise price, but they're not willing to do that because they're afraid to lose share". A year later, price increases had happened in four forms.

The first is repricing the flagship line outright. OpenAI's flagship input price went from US$1.25 to US$5 per million tokens, 4x in a year. The second is a new tier above the flagship. Opus is Anthropic's main line and Sonnet 5 the lower-priced line; Fable 5 sits above the main line at twice the Opus price (US$10/$50), and its fast mode charges a further 2x premium. The third is a promotion expiring. Sonnet 5's current promotional price is US$2/$10, and the official page states that it ends on August 31 and that the standard US$3/$15 applies from September 1, a 50% increase. The fourth is the least noticed. Anthropic's official documentation states that its new-generation tokenizer produces about 30% more tokens from the same text (Anthropic's pricing documentation). The sticker price doesn't move a cent, yet the bill for the same document is 30% higher, and no price-list comparison shows it. A tokenizer is the ruler a model uses to cut text into billable units. Shrink the ruler and the same text counts as more units.

On the same pricing pages, price cuts are real too, and this is where Gurley still stands. Anthropic's main Opus line came down from US$15/$75 to US$5/$25, a 67% cut. On July 30, OpenAI cut the input price of the low-end model in its new flagship family from US$1 to US$0.2. Holding capability fixed gives the fuller measure. That annual decline doesn't measure the price war between these two vendors; it measures what the whole market pays for "the same capability", which falls between 9x and 900x a year, median 50x (Epoch AI). What is rising is the entry price for this generation's best capability. What is falling is the price of doing the same thing. The share-grabbing price cuts Gurley described are still running at the mature-capability tier, and harder than he described them.

On gross margins: negative margins really existed, but possibly a year earlier than Gurley said. In May 2026 SemiAnalysis gave a time series: "inference gross margins are now in the mid 60s, up from 38% in 2025 and -94% in 2024" (SemiAnalysis). The series gives Gurley half credit. The -94% year is 2024, and Gurley spoke in July 2025, when that year was already over. The positive 38% was published only in May 2026, looking back to fix the figure for the full year 2025. That figure did not exist when Gurley spoke in July 2025, so it can't count as something he could see at the time. Gurley also named no brand, so the rumor can't be pinned to any company. By the first half of 2026, the same analysis firm's conclusion read "The age of low gross margins for frontier model providers is over", and the same report put inference gross margins as moving from 38% to above 70% (SemiAnalysis). The revenue half can be checked against a founder's public remarks. Anthropic CEO Dario Amodei said at the company's developer conference on May 8, 2026, that annualized revenue had climbed to US$30 billion, "a three-fold increase compared to last year" (Fortune).

The cloud point is only half primary-source, and it is the weakest receipt in this column. SemiAnalysis says "AWS margins inflected this past quarter driven primarily by customer spending growth on Claude through Bedrock... EBIT margins increasing 213bp Q/Q". Amazon's official first-quarter 2026 results do show AWS operating income of US$14.2 billion, against US$11.5 billion a year earlier, and revenue that "increased 28% year-over-year" (Amazon's official results).

Two things need saying first. One: the 213bp can only be checked on the same ruler, and SemiAnalysis measured quarter over quarter. On that ruler, using Amazon's official figures, AWS operating margin was 35.0% in the previous quarter (Q4 2025) and 37.7% this quarter (Q1 2026), an improvement of 270 basis points (our calculation). That doesn't match 213bp, and the gap is SecondSource's finding. The figures above, operating income of US$14.2 billion against US$11.5 billion and revenue up 28%, are year over year, a different ruler that can't check a quarter-on-quarter number. Two: Amazon has never officially said the improvement was driven by Claude. That attribution is SemiAnalysis's inference.

More important is the gross-margin series itself. Anthropic is a private company with no statutory filings, and the 38%, mid-60s and 70% figures come from SemiAnalysis alone. A recheck on August 6 found that the numbers cited by several outlets all trace back to the same report, and no second independent firm has reproduced them. The Information, meanwhile, puts Anthropic's overall 2025 gross margin at 40%, which comes from a different set of books from this series entirely. So the evidence that margins "turned positive" is clearly weaker than the evidence on price increases.

③ The receipts say: doesn't hold

(Re-ruled unresolved on October 3 under the condition in ⑤; see the correction at the top.)

What is being ruled on is the claim itself: "The whole industry is buying share at a loss, so price is the only valve on demand, and nobody dares touch it."

The first half gets only half credit. Negative gross margins did exist, but the visible negative figure, an inference gross margin of -94%, belongs to 2024. By the time Gurley spoke in July 2025, that year was over, and the finding that his timing was late rests on this point. The positive 38% was fixed for the full year 2025 only in May 2026, looking back, so Gurley could not have seen it when he spoke. Gurley also named no brand, so the rumor can't be pinned on anyone. The first half therefore comes out as: direction supported, timing late, target unverifiable.

The second half doesn't hold, and the break is clean. The hinge of Gurley's argument is that nobody dares touch price. From that, Gurley infers that the demand signal is watered down, and that this will show only when the capital window closes and forces a repricing. Within a year, prices were in fact adjusted four times, and the vendors moved by choice while the capital window was wide open and demand was surging, not under pressure. The company Gurley named as afraid to raise prices raised them the most. With the hinge broken, the chain of consequences after it doesn't connect.

To be clear about what has not been overturned: Gurley's price-cutting to grab share runs as before at the mature-capability tier, and the median 50x annual decline is the evidence. What is overturned is the universal "nobody dares raise prices", not the existence of subsidized pricing.

What would pull this call back to unresolved: ① The Sonnet promotion's September 1 expiry is not carried out as written, because the official page extends the discount or cuts the standard price outright. The hardest receipt is then void. ② Anthropic's first profitable quarter doesn't materialize, or later quarters slip back into losses. The turn to positive margins would then be only shortage-period rent, and the watered-down-demand argument comes back. ③ An Anthropic IPO filing, or a second independent firm, shows inference gross margins far below the mid-60s. The gross-margin receipt is then void in full.

How to use it. What this changes is the shape of a model-cost budget: a single line saying tokens keep getting cheaper no longer describes it, and it takes two lines. Costs for workloads that can move down to a good-enough tier fall every year. Costs for the share of work that competition locks to the flagship tier rise, and from four directions; the tokenizer change among them is invisible in the sticker price. The boundary, in one sentence: this column rules on "nobody dares raise prices", not on whether AI demand is real, and that second question is not one with an answer inside 90 days.

④ Where the claim sits

The claim's place on SecondSource's map of trends follows an existing tree; no new one is drawn. Two names in path B first. Evans is Benedict Evans, an independent tech analyst who argues that frontier models have become commodities. Narayanan is Arvind Narayanan, a Princeton computer-science professor and co-author of AI Snake Oil.

⑤ Checking the answers

This issue opens two short windows, and both test this column's call directly. First: a direct check of Anthropic's official pricing page after September 1, to see whether Sonnet 5 reverted to the standard rate as its terms say. If the price reverted, "they dare to raise prices" moves from a clause to an executed fact. If not, this column's hardest receipt is void and the call goes back to unresolved. Verdict date September 15. Second: whether Anthropic's first profitable quarter, the second quarter, materializes, with no return to an operating loss in the third quarter. This one tests whether the turn to positive margins is structural or only shortage-period rent. Verdict date October 31.

One older entry gets an update: the one opened in the previous column, on whether an analyst on CoreWeave's second-quarter call would press on how much of CoreWeave's unrentable compute Nvidia has committed to take. The triggering event has happened, with results on August 11, but the transcript of that call is not yet in hand. So the entry is neither closed nor guessed at. The standard and the September 15 due date stay as they are, and the call will be made once the transcript is available. The ledger records this as a revision.

The first verdict date is 2026-09-01. All judgments and their verdict dates → the judgment index