Weekly Chart Pick · 2026-10-30
AI Models — the pick with no chart yet: a frontier lab files to list
A maker of frontier AI models filed to list this week. There is no price chart to read, so this special edition reads the filing instead: quarterly revenue up thirty-fold in fifteen months, a first operating profit, far more cash than the loudest critics claim, and $266 billion of compute commitments.
- Revenue from $376 million to $11.6 billion a quarter in fifteen months; the June quarter grew 1,368% from a year earlier.
- Profitable in the June quarter: an 8% operating margin on $11.6 billion of revenue, yet a $693 million operating loss for the first half; the 85% gross margin counts serving costs only.
- Cash, not debt: $131.7 billion in cash and short-term investments after $141.1 billion raised in total; the $108.6 billion notes line is shares owed, expected to be settled in stock.
- The $80.7 billion first-half loss is mostly an accounting mark: $72.4 billion of it is the rising value of those notes, while operating cash flow was +$3.9 billion.
- Customers paying $100,000 a year went from about 950 to more than 11,900 in twelve months; two of our AI Users members are named in the filing.
- The risks are real: $265.6 billion of compute commitments, margins the company expects to compress, and founders holding 50.1% of the votes.
ANT Anthropic proposed symbol (not yet listed)
The Setup
The claim: this filing shows AI spending paying off at the receiving end, at a frontier model maker. This week the market asked whether the hundreds of billions going into data centres will earn a return (today's Daily Pulse). This filing answers from the company that rents much of that capacity. In the June quarter it sold $11.56 billion of AI against $5.56 billion of compute and infrastructure cost, and made an operating profit of $974 million.
Why we are bullish, reason one: growth that is still speeding up. Quarterly revenue went $376 million, $787 million, $1.38 billion, $2.04 billion, $4.73 billion and then $11.56 billion, from the first quarter of 2025 to the second quarter of 2026. Each step is bigger than the last. June-quarter revenue rose 1,368% from a year earlier, and first-half revenue reached $16.29 billion against $1.16 billion.
Reason two: each dollar of compute earns more. In the June quarter compute and infrastructure cost rose 300% from a year earlier while revenue rose 1,368%. Compute cost was 177% of revenue in the June quarter of 2025 and 48% a year later. Revenue passed the compute bill in the first quarter of 2026, and the gap widened in the second, as the chart shows.
Reason three: customers who keep spending more. Customers paying $100,000 or more a year rose from more than 950 to more than 11,900 in twelve months to 30 September. More than 1,900 pay $1 million or more and about 250 pay $10 million or more. The filing puts net revenue retention, what last year's customers spend now against what they spent then, at about 900%. More than half of the $100,000-plus customers use both the developer platform and one of the company's own apps, up from about 30% in the last quarter of 2025.
Two of those customers are members of our AI Users index, the group that still rises in a rocky market. Walmart, a customer since 2024, says weekly active users of its shopping agent grew 110% in a quarter, and orders placed through it are 35% larger. Eli Lilly has given more than 50,000 employees access and has expanded its use fivefold since the rollout. When AI users get paid, part of that money flows to the model maker.
Reason four: a balance sheet stronger than the headlines. At 30 June the company held $112.4 billion of cash and $19.3 billion of short-term investments after raising $141.1 billion in its lifetime. By our arithmetic about $9 billion of the capital raised is gone; Ball, counting restricted cash, puts it near $7 billion. The $108.6 billion line called convertible notes payable is mostly shares owed to early strategic investors, expected to be settled in stock, not cash. First-half operating cash flow was +$3.93 billion against $0.48 billion of capital spending, a free cash flow of $3.45 billion, 21% of revenue. After stock pay of $2.19 billion it is still about 8%.
Reason five: distribution already in place. The models are sold natively on all three large clouds, AWS, Google Cloud and Azure, and through systems integrators. Revenue per employee over the trailing twelve months was more than $6.6 million, against $1.4 million a year earlier.
Jamin Ball of Altimeter, an investor in the company, framed the debate in his Clouded Judgement letter of 30 October as three questions. Can the company stay at the frontier? Can anyone build a durable business on frontier models? And can that business pay for staying at the frontier? For now the filing answers the second. The first and the third are the bet, and they are what we score.
In the Closelooknet readings
Said plainly: this is Anthropic, PBC, the maker of the Claude models. It filed its S-1 on 26 October 2026 to list Class A shares on the Nasdaq Global Select Market under the proposed symbol ANT; no price range is set yet. It is in no Closelooknet index because it has not traded. Two members of our AI Users index, Walmart and Eli Lilly, appear as named customers in the filing's case studies.
The latest print
Latest quarter in the filing: three months to 30 June 2026 (unaudited). Revenue was $11.56 billion, up from $787 million a year earlier. Compute and infrastructure cost was $5.56 billion and total operating costs $10.58 billion, leaving an operating income of $974 million.
Two margins, two periods, both true. The 8% operating margin Jamin Ball cites is the June quarter alone: $974 million on $11.56 billion, 8.4%. The March quarter before it lost $1.67 billion on $4.73 billion of revenue, a margin of minus 35%. Add the two and the first half shows an operating loss of $693 million on $16.29 billion, minus 4.3%. On the company's adjusted measure, which strips out stock pay and some other items, the margin went from minus 13% in the March quarter to plus 19% in June. So the business crossed into operating profit during the half, not across it, and the June quarter is the first profitable one in the table.
Table as text
| Quarter | Revenue | Compute and infrastructure | Operating income | Operating margin | Revenue per $ of compute |
|---|---|---|---|---|---|
| Q1 2025 | 376 | 944 | −935 | −249% | 0.40x |
| Q2 2025 | 787 | 1,390 | −1,200 | −152% | 0.57x |
| Q3 2025 | 1,382 | 2,047 | −3,089 | −224% | 0.68x |
| Q4 2025 | 2,040 | 2,945 | −2,836 | −139% | 0.69x |
| Q1 2026 | 4,731 | 3,387 | −1,667 | −35% | 1.40x |
| Q2 2026 | 11,555 | 5,558 | 974 | +8% | 2.08x |
S-1 filed 26 October 2026, unaudited quarterly statements of operations. Q3 2025 includes the $1.5 billion Bartz settlement. Revenue per dollar of compute matches the company's own chart (0.40x to 2.08x).
Where the 85% gross margin comes from. The statement of operations is organised by type of cost (compute, personnel, sales, other) and shows no gross margin at all. A note to the accounts, Supplemental Operating Expense Information on page F-39, regroups the same $16.98 billion of first-half costs by function: cost of revenue $2.38 billion, research and development $9.74 billion, sales and marketing $3.81 billion, general and administrative $1.05 billion. Revenue minus cost of revenue is $13.91 billion, 85.4% of revenue. That is the 85%. The note says substantially all of cost of revenue is compute and infrastructure for customer-serving workloads. On the same basis the margin was 12.4% in the first half of 2025 and 39.5% for the full year 2025.
Table as text
| Line | H1 2025 | H1 2026 | Share of H1 2026 revenue |
|---|---|---|---|
| Revenue | 1,163 | 16,286 | 100.0% |
| Cost of revenue | 1,019 | 2,380 | 14.6% |
| Research and development | 1,850 | 9,741 | 59.8% |
| Sales and marketing | 239 | 3,808 | 23.4% |
| General and administrative | 190 | 1,050 | 6.4% |
| Total operating costs | 3,298 | 16,979 | 104.3% |
| Operating result | −2,135 | −693 | −4.3% |
| Gross margin (revenue minus cost of revenue) | 12.4% | 85.4% | — |
S-1 filed 26 October 2026, note on page F-39 (Supplemental Operating Expense Information), unaudited. Cost of revenue is substantially all compute for customer-serving workloads; training compute sits in research and development, free-tier compute and channel partner costs in sales and marketing.
What the 85% leaves out. Three costs sit below that line. Training and research compute is booked in research and development: at least $6.6 billion of the $8.95 billion compute bill sits outside cost of revenue. Compute for free users is booked in sales and marketing. And 48% of first-half revenue came through the cloud marketplaces of AWS, Google and Microsoft, booked gross, so the partners' fees do not reduce revenue; the filing lists channel partner costs first among sales and marketing expenses. Counting all of sales and marketing as cost gives 62.2%. That is a floor for this adjustment, because the line also holds advertising and staff. Counting every dollar of compute, serving and training alike, gives 45.1% for the half and 51.9% for the June quarter. Each figure answers a different question. The 85% says serving a paying customer is very profitable; the 45% says the whole model factory is, too, but by much less.
Missing from the first draft: the September quarter's revenue and the latest annualised run rate are left blank. Audited numbers take time, and Ball points out that Cerebras and Figma also listed without their latest quarter. Press reports put third-quarter revenue near $16.8 billion. That figure is unaudited and not in the filing; the amended filing should carry it.
The fundamentals
The $80.7 billion loss line. First-half net loss was $80.7 billion, of which $72.4 billion is the rise in the fair value of convertible notes, derivatives and SAFEs. The notes are marked at today's valuation, so the loss grows as the company becomes more valuable. Another $8.7 billion comes from other financial instruments, which Ball traces to forward commitments to sell preferred shares to Google and Amazon at an earlier round's price. None of it is cash. The filing says these marks continue until the notes convert into equity, which its pro forma figures tie to the offering.
Valuation. The filing sets no price range yet. The latest private round, Series H-1 in May and June 2026, sold 77.6 million shares at $589.01 for $45.7 billion. Media reports put the target valuation at $1.5 trillion to $2 trillion and the raise near $100 billion, with a listing expected in the week of 16 November. None of that is confirmed. Our arithmetic: the June quarter's revenue, annualised, is $46.2 billion, so $1.5 trillion to $2 trillion is 32 to 43 times that figure. On the reported September quarter it would be 22 to 30 times. Ball's high-growth software group trades at a median 18.7 times next-twelve-month revenue, so the premium is paid for growth no software company has shown.
The chart, read by hand
Trend: There is no trend to read: the shares have not traded. The chart above replaces the price chart with the filing's own lines, revenue against the compute bill by quarter, and the operating result under each quarter.
Key levels: Series H-1 private round: $589.01 a share (May–June 2026, from the filing) · IPO price range: not set in the first filing · Scoring anchor: the first close after listing
The risks, equal billing
What would prove us wrong. The company itself says margins will compress. Compute is bought 12 to 18 months before it arrives, and its cost lands before the revenue it will support. It also expects to buy more short-term compute at a higher unit cost in the near term. If compute cost outgrows revenue again for two quarters, the efficiency case behind reason two is broken. And the 85% gross margin is the easiest of the margins to flatter: half of revenue comes through cloud partners whose fees sit below that line.
The commitments. Fixed or minimum compute commitments were about $265.6 billion at 30 June, with terms of one to 20 years, against $54.6 billion at the end of 2025. That is almost six years of June-quarter revenue, annualised. If demand slows, the bills stay. This is the same question the market put to Microsoft, Alphabet and Oracle this week, seen from the other side.
Control. Public holders will own Class A shares with little say. Class F stock, held by a founders' company, carries 50.1% of the votes and no economic rights, and the Long-Term Benefit Trust can generally elect and remove a majority of the board. As a public benefit corporation, the company balances profit against its stated mission.
Price and supply. At the reported valuations the stock would start at 32 to 43 times annualised June-quarter revenue. A raise near $100 billion, and later lock-up expiries, add a lot of stock. There is legal risk too: the company booked a $1.5 billion settlement in a copyright class action (Bartz) in 2025. And the frontier race is costly; one weaker model generation would hit growth and pricing together.
Our forecasts
Into the listing: The next facts are the amended filing with a price range and the audited September quarter. We expect third-quarter revenue well above the June quarter's $11.6 billion; press reports put it near $16.8 billion. A figure below the June quarter would break the bull case before the first trade.
First weeks of trading: A deal this size tends to trade in wide swings at first. We score from the first close, not the offer price, and expect the first report as a listed company to set the tone more than the debut.
12 months: Our view: revenue keeps outrunning the compute bill, even as the operating margin dips while new capacity comes online, as the company warns. If revenue stays above twice compute cost and the $100,000-plus customer count keeps compounding, we stay bullish. If compute again outgrows revenue for two quarters, we will say we were wrong.
Diary views, not advice — held publicly alongside the scored record.
Special edition, written 30 October 2026 from the S-1 filed 26 October. There is no price yet: the score starts from the first close after listing and runs at one week, one month, three, six and twelve months. This edition was drafted with the help of Claude, the company's own AI model; every number is from the filing or a named, dated source, and the view is Closelooknet's. Diary, not investment advice.
Closelooknet publishes an investment research diary, not investment advice. Every pick is scored publicly against the tape at 1w/1m/3m/6m/12m from publication — the record, not the single call, is the product.
A note from the editor: the Weekly Chart Pick is free. Its price is one click — add Closelooknet as a preferred source on Google, and you support this work in Google Discover and Search every time you read.
Google opened in a new tab. Add Closelooknet there, then come back to this tab — the Chart Pick unlocks the moment you return.
The preferred-source setting lives in your Google account — if Google asks you to sign in first, you will land on the selection page right after. Google does not tell us whether you saved the selection. Clicking “Done” unlocks the Weekly Chart Pick on this device for 12 months.