According to the S-1 that Anthropic made public on October 26 (a registration statement filed with the U.S. Securities and Exchange Commission for a U.S. initial public offering), the company posted an operating profit of $974 million for April–June 2026, an operating margin of 8%. We read this profit as a snapshot taken before the large, newly contracted blocks of compute have started to hit the income statement. The S-1 itself says operating margin may compress in periods when large amounts of compute come online. The contract with SpaceXAI, which provides NVIDIA GPU compute, comes to about $3.1 billion a month once fully ramped, or about $9.3 billion a quarter. That is roughly 1.7 times the company's entire compute and infrastructure cost for April–June. The basis for judging this is the amount, term and termination conditions of the seven contracts the S-1 lists by counterparty.
A $11.5 billion quarter, and the only operating profit in the six quarters the S-1 shows
In the S-1's quarterly income statement, revenue grew from $376 million in January–March 2025 to $11.555 billion in April–June 2026. The April–June figure is up 1,368% from $787 million a year earlier. GAAP operating income for the quarter was $974 million, the only operating profit among the six quarters for which the S-1 gives quarterly figures. Adjusted operating income was $2.235 billion (a 19% margin); the $1.261 billion gap with GAAP is stock-based compensation.
The net loss for the first half reached $80.748 billion, of which $72.388 billion came from changes in the fair value of convertible notes and similar instruments, mainly valuation losses that follow a rise in the company's value. It involves no cash outlay.
The filing was submitted with the amounts left blank for July–September revenue, run-rate revenue as of October 31, and total contract value. Under Rule 3-12 of the SEC's Regulation S-X, a registrant that is not an accelerated filer must update to newer interim financial statements if the financial statements are 135 days or more older than the expected effective date. The 135th day after June 30 is November 12 (our calculation), so if the registration becomes effective after that date, financial statements as of September 30 would need to be added.
For Japanese readers, April–June revenue is about ¥1.819 trillion and operating profit about ¥153 billion (¥157.45 per dollar, ECB reference rate, October 26, 2026). The offering will be made in the U.S. and also in Japan, but the shares will not be listed on a Japanese exchange. Mizuho and MUFG are among the underwriters.
$321.4 billion in leases not yet commenced shows when compute becomes a cost
The S-1 describes its own financial model this way: compute is procured and contracted in advance, and it becomes an expense when it goes live and joins the shared internal pool of compute. Costs often rise in large steps. Most long-term contracted compute is procured 12 to 18 months before it is put to use, the company says.
Because of the lag between committing to investment and earning enough revenue, the company expects that "operating margin and Adjusted Operating Margin may compress" in periods when large amounts of compute come online. If demand exceeds contracted capacity, it will fill the gap with higher-priced short-term compute, and heavier use of that is also expected to compress margins in the near term.
The scale of that lag shows up in the notes to the financial statements. Leases not yet commenced (estimated undiscounted future lease obligations) grew from about $50.2 billion at the end of December 2025 to about $321.4 billion at the end of June 2026. They begin between 2026 and 2028, with terms from under one year to 20 years, and some depend on contractual conditions being met. Compute and infrastructure costs for April–June rose 300%, from $1.39 billion a year earlier to $5.558 billion. That increase is already in the income statement, but the leases not yet commenced are not yet in this expense.
Gross margin also makes this burden hard to see. First-half cost of revenue was $2.38 billion, which against revenue of $16.286 billion implies a gross margin of about 85.4% (our calculation). But most of cost of revenue is compute and infrastructure cost for serving customers, with some allocated personnel and overhead. That $2.38 billion is only about 26.6% (same calculation) of total first-half compute and infrastructure costs of $8.945 billion. The rest of those costs is booked mainly in research and development, for compute used in training and research, and in sales and marketing, for free-tier usage.
Seven compute contracts by counterparty; the SpaceXAI deal is about $9.3 billion a quarter at full ramp
The S-1 leaves the total contract value blank but describes the compute contracts individually by counterparty. The seven for which amounts are given are as follows.
| Counterparty | Amount | Term | Key terms |
|---|---|---|---|
| Microsoft | $44.9 billion or more | Nov 2026–May 2033 | Payable regardless of usage; generally non-cancelable |
| About $114.5 billion or more | Apr 2026–Jul 2033 | Shortfall paid as difference; generally non-cancelable | |
| Amazon | $110 billion or more | May 2026–Apr 2036 | Shortfall paid as difference; generally non-cancelable |
| SpaceXAI | Up to $120.5 billion | Through 2029 | Cancelable on 90 days' notice after initial period |
| Nscale | Up to about $44.6 billion | Through 2033 | Billed only for delivered capacity; generally non-cancelable |
| Lambda | About $41.3 billion | Through 2034 | Billed only for delivered capacity; generally non-cancelable |
| Broadcom (equipment lease) | About $161.2 billion (excluding interest) | 5 years | Non-cancelable |
The figures in the table are not all the same kind of number. Microsoft, Google and Amazon are minimum commitments payable regardless of usage, SpaceXAI and Nscale are caps, and Broadcom is equipment lease payments. Except for SpaceXAI, the six contracts cannot be canceled other than in limited cases such as a material breach by the counterparty. The seven contracts for which the S-1 gives amounts by counterparty total $637 billion; adding estimates for additional TPU procurement from Google and others, whose amounts are not yet fixed ($125 billion to $150 billion), and AMD (more than $20 billion), brings the total to about $782 billion to $807 billion. Because the AMD figure is "more than $20 billion," even the upper end is not a ceiling.
This total is not a figure the S-1 itself presents. It covers a different scope from balances in the notes, such as the leases not yet commenced mentioned above, and partly overlaps with them. In yen, $637 billion is about ¥100 trillion.
The SpaceXAI contract is where the timing and amount of costs are described most concretely. Full ramp is expected in December 2026, with different pricing applying during the ramp-up period until then. After full ramp, the contract runs at about $3.1 billion a month, or about $9.3 billion a quarter, roughly 1.7 times the total compute and infrastructure cost for April–June 2026 ($5.558 billion). The former is converted from a contractual monthly amount; the latter is the actual income-statement figure across all suppliers. April–June costs may include part of the SpaceXAI ramp-up period, and accounting expense can differ from cash payments.
Even with those caveats, the point stands: a single supplier's contract, converted to a quarter, exceeds the entire compute and infrastructure cost for April–June.
Of the seven in the table, the Google and Amazon terms already began in April and May 2026, and part of the Broadcom lease is also already running in 2026, so these appear to be partly in April–June costs already. By contrast, the Microsoft contract begins in November, SpaceXAI is expected to reach full ramp in December, and additional Broadcom capacity starts in 2027. Nscale and Lambda are billed as capacity is delivered. Considering also the roughly $321.4 billion in leases not yet commenced at the end of June, we read the 8% for April–June as a figure taken before the large additional blocks of contracted compute have become costs.
If the 8% margin holds in January–March 2027, this reading is wrong
The S-1 itself supports the opposing view. The company writes that it has operated in an environment where customer demand exceeds available compute, so additional compute can convert directly into revenue. On margin compression, it says that margins can fluctuate from period to period even as the underlying economics of the business keep improving. It also states that operating margin and adjusted operating margin for January–September 2026 expanded substantially from a year earlier (the July–September amounts alone are left blank).
Calculating from the S-1's quarterly table, the ratio of compute and infrastructure cost to revenue has fallen steadily across the six quarters, from 251% in January–March 2025 to 48% in April–June 2026 (our calculation). It is a fact that revenue growth has outpaced costs even as costs have risen ahead of it.
Demand figures are also rising. Customers with trailing-12-month revenue of $100,000 or more number over 11,900, up from over 950 a year earlier. July–September revenue was reported to be $16.8 billion, up about 45% from the prior quarter, according to attendees at an investor meeting (this is not in the S-1).
The means of stopping costs are limited, but the company has financial capacity. The SpaceXAI contract can be canceled on 90 days' notice after initial non-cancelable periods ending in January and March 2027. But that is a contract with a cap of $120.5 billion, only about 19% (our calculation) of the $637 billion total for the seven. The other contracts cannot be canceled except in limited cases such as a material breach by the counterparty. First-half operating cash flow was a positive $3.9 billion, and as of June 30 the company held $112.4 billion in cash and cash equivalents and $19.3 billion in short-term investments.
The opposing view comes down to two points: revenue can grow faster than costs, and the SpaceXAI portion can be canceled to stop costs. Our reading denies neither. What can be said is that the 8% is a figure from before the large compute blocks become costs, and there is no guarantee it will hold in January–March 2027, when that is under way.
Whether this reading is right will be tested by results for January–March 2027, which would include the period after SpaceXAI's full ramp. If compute and infrastructure costs exceed $9 billion a quarter and GAAP operating margin is still 8% or higher, the reading is wrong. The $9 billion mark is a benchmark we set, slightly below the roughly $9.3 billion for the SpaceXAI contract.
A leading indicator will be full-year 2026 results, which include December, the expected month of full ramp. The announcement of January–March 2027 results also serves as a reference point for the final release of the 180-day lockup. The release comes two trading days after that announcement or 180 days after the prospectus date, whichever is earlier.
The path the S-1 itself lays out is one in which demand keeps exceeding supply and compute that comes online converts directly into revenue. If that holds, revenue growth would absorb the cost of contracted compute, and the 8% in April–June would be a figure that persists after those costs arrive.