Business & Markets · 2026-10-30 · Team LATENT

What Anthropic's S-1 reveals about its operating profit: estimated third-quarter revenue, a $3.1 billion monthly SpaceXAI contract, and four unauthorized-access incidents

The only operating profit in Anthropic's S-1 quarterly table is April to June 2026, and the filing says it buys compute 12 to 18 months ahead. From the table and the contract terms we calculate why the profit appeared, the third-quarter revenue left blank, and the fixed costs that rise from December. We also compare the incident the filing uses as its example with the one Anthropic said in September concerned it most.

In the quarterly table of Anthropic's S-1, its IPO registration statement, only one column shows an operating profit: $974 million for April to June 2026. The same filing says the company buys compute 12 to 18 months before it can use it. Put those two facts side by side and you can see why the profit appeared and how large the coming costs are.

Anthropic filed the S-1 with the US Securities and Exchange Commission on October 26, 2026, and plans to list on Nasdaq under the symbol ANT. The size of its revenue and losses has already been widely reported. This article covers what only appears when the filing's numbers are combined: the reason for the profit, the blank third-quarter revenue, the fixed costs that rise from December, the incident the S-1 uses as its example, and how safety pauses relate to fixed costs.

We checked the figures and quotations in the S-1 itself on October 30. Ratios, monthly averages, and the third-quarter estimate are LATENT's calculations. This article does not recommend buying or selling shares and does not forecast the share price.

Anthropic's S-1 / Anthropic's announcement / Incident assessment (September 9)

Key takeaways

Compute expense as a share of revenue fell from 251% in January to March 2025 to 48% in April to June 2026. The filing describes this expense as one that does not move with demand. The operating profit came in the quarter when revenue grew faster than compute expense.
The S-1 leaves third-quarter revenue blank, but revenue per employee and headcount imply about $16.6 billion. That is about 1.4 times the second quarter, slower than the 2.4 times of the quarter before. Meanwhile about $3.1 billion a month for SpaceXAI arrives in December, and the simple average of the four main contracts reaches about $5.9 billion a month, 3.2 times second-quarter compute expense.
The one incident the S-1 describes in detail matches the case that Anthropic's September report called less concerning than the other three. The Mythos 5 case that the report called most concerning does not appear in that passage.

A Nasdaq listing, with share count and price range undecided

According to the S-1, Anthropic will list Class A common stock on the Nasdaq Global Select Market under the symbol ANT. The lead underwriters are Goldman Sachs and Morgan Stanley. The number of shares and the price range are blank, as are third-quarter revenue, total compute commitments, and the co-founders' share of the vote.

Japan is among the places outside the US where the shares will be offered to the public. The filing names Australia, Canada, certain European Economic Area states, Japan, Switzerland, and the United Kingdom, and says the company does not intend to list on exchanges in those jurisdictions. The underwriter list includes Mizuho and MUFG. The filing does not say which brokers in Japan will handle the shares or on what terms.

Timing and size are not in the filing. The Next Web relays WSJ sourcing for a listing in the week of November 16, raising $100 billion or more at a valuation of $1.5 to $2 trillion, and says it has not confirmed this itself.

S-1 / The Next Web

The operating profit came in the quarter when revenue grew faster than compute expense

The table below is taken from the S-1's quarterly table. The filing itself states the 251% and 48% ratios, and LATENT calculated the others the same way. Expenses for July to September 2025 also include a separate $1.5 billion copyright settlement.

Quarter	Revenue ($M)	Compute expense ($M)	Share of revenue	Operating income ($M)
Jan-Mar 2025	376	944	251%	-935
Apr-Jun 2025	787	1,390	177%	-1,200
Jul-Sep 2025	1,382	2,047	148%	-3,089
Oct-Dec 2025	2,040	2,945	144%	-2,836
Jan-Mar 2026	4,731	3,387	72%	-1,667
Apr-Jun 2026	11,555	5,558	48%	974

The filing explains the pattern this way. Compute is procured 12 to 18 months before it is available, so the expense is largely fixed in the near term. It rose because committed compute arrived, not because of demand in the period. The ratio fell because revenue grew faster than the committed cost base.

For the final quarter, revenue rose $6.824 billion from the previous quarter and compute expense rose $2.171 billion. Each added dollar of revenue came with $0.32 of added compute, below the average of $0.48. The operating margin was 8%, and adjusted operating income excluding stock compensation was $2.235 billion.

LATENT therefore reads this profit as a number that contains a timing gap between buying and selling. If the filing's explanation holds, profit comes more easily when revenue runs ahead of contracts, and the ratio can rise when contracted compute arrives in bulk. The filing itself warns that when compute arrival and demand are out of step, the result can be idle capacity and cash outflows that produce no immediate revenue.

S-1 quarterly trends and compute risk factors

About 90% of the $80.7 billion net loss is fair-value change on convertible notes and similar items

Revenue for the first half of 2026 was $16.286 billion, up 1,300% from $1.163 billion a year earlier. The net loss reached $80.748 billion. The operating loss was only $693 million, so most of the gap sits outside operations.

The components are a $72.388 billion loss on the fair value of convertible notes, derivative liabilities, and SAFEs, $8.757 billion of other expense including note purchase agreements, and $1.111 billion of interest income. Fair-value change alone is about 90% of the net loss. It was largest in the second quarter at $59.597 billion, when the net loss was $66.309 billion.

The S-1 says the company does not expect revenue to keep growing at comparable rates and notes it has had a net loss every year. Net losses for 2023 through 2025 were $1.531 billion, $8.309 billion, and $41.965 billion.

Blank third-quarter revenue works out to about $16.6 billion

The S-1's revenue tables stop at June 30, and the field for revenue through September 30 is empty. But the filing gives trailing twelve-month revenue per employee of over $6.6 million as of September 30, and about 5,300 employees. Multiplying the two gives trailing revenue of about $35 billion.

To test the method, apply it a year earlier. Adding $197 million for October to December 2024 to the first three quarters of 2025 gives $2.742 billion for the twelve months to September 30. Dividing by $1.4 million per employee gives about 1,959 people. The filing says headcount grew about 170% in a year, which from 5,300 gives about 1,963. The two are close, so revenue per employee appears to use period-end headcount.

Subtracting $18.326 billion, the total for October to December 2025 and the first two quarters of 2026, from $35 billion leaves about $16.6 billion for the third quarter. Varying revenue per employee between $6.6 million and $6.7 million, and headcount between 5,250 and 5,350, gives a range of $16.3 billion to $17.5 billion. The filing's separate statement that trailing revenue rose about 1,200% gives about $17.3 billion, also inside that range. The Next Web relays an FT report that investors were told $16.8 billion.

About $16.6 billion would be about 1.4 times the second quarter's $11.555 billion. The step from the first to the second quarter was 2.4 times, so dollars keep rising while the growth rate falls. Third-quarter compute expense and operating income are unknown, so this calculation cannot say whether that quarter was also profitable.

Fixed costs rising from December: about $5.9 billion a month across the four main contracts

The S-1's section on contracts gives amounts by counterparty. The Google, Amazon, and Microsoft figures are minimums payable regardless of actual usage, and the SpaceXAI figure is a ceiling. Monthly averages are LATENT's simple division of the amount by the stated term. Actual payments rise as capacity arrives, so they differ from these figures month to month.

Counterparty	Amount	Term	Monthly average (LATENT)
Google	At least about $114.5B	Apr 2026 to Jul 2033	About $1.30B
Amazon	At least $110.0B	May 2026 to Apr 2036	About $0.92B
Microsoft	At least $44.9B	Nov 2026 to May 2033	About $0.57B
SpaceXAI	Up to $120.5B	Into 2029	About $3.1B (filing's figure)

SpaceXAI, formerly X.AI LLC, is a wholly owned subsidiary of SpaceX, and Anthropic buys NVIDIA GPU compute from it. All capacity is expected by December, after which the cost is about $3.1 billion a month. That alone is 1.67 times the second-quarter monthly average compute expense of $1.85 billion, and 80% of that quarter's monthly average revenue of $3.85 billion. This contract alone has an exit, however. Initial non-cancelable periods end in January and March 2027, after which either side can cancel on 90 days' notice.

The four monthly averages total about $5.89 billion, 3.2 times the second-quarter monthly average. That is slightly above the estimated third-quarter monthly revenue of about $5.55 billion. Holding compute at the second quarter's 48% of revenue would take about $12.2 billion a month, or about $36.7 billion a quarter, about 2.2 times the third-quarter estimate. The filing does not set 48% as a target, and this calculation is only meant to convey scale.

Other payments sit outside the table. Equipment leases tied to Broadcom total about $161.2 billion excluding interest, additional TPUs due between 2028 and 2031 are estimated at about $125 billion to $150 billion, and there is an AMD agreement as well. Adding the stated amounts gives $676.1 billion to $701.1 billion. That total is LATENT's, because the filing's total is blank. Business Insider Japan reports about $800 billion in contracts, a figure LATENT could not confirm in the filing. Cash and short-term investments were $131.7 billion at June 30.

The filing also warns that as agent use grows, the compute required could rise disproportionately to revenue, because a single use repeats long reasoning and tool calls. If that happens alongside revenue growth, the ratio will be harder to bring down.

S-1 compute arrangements / Business Insider Japan

Revenue is mostly usage-based, and the top five customers account for about 25%

Against those fixed costs, revenue carries shorter commitments. According to the S-1, most revenue is usage-based and metered in tokens. Of the $15.1 billion increase in first-half revenue, $12.1 billion was consumption-based and $3.0 billion was subscription. The filing says consumption revenue can fluctuate significantly if customers cut usage or move to cheaper models.

The customer count is rising. Customers paying $100,000 or more over twelve months went from about 1,500 at the end of 2025 to more than 11,900 on September 30, 2026. The filing states net revenue retention of about 900% as of September 30, though this metric has not been independently verified.

No customer exceeded 10% of revenue, and the top five made up about 25% of revenue for January to September 2026. The filing says many top customers have no long-term contracts or have contracts that expire soon. Sales through cloud marketplaces were 48% of first-half revenue, and the fees are growing too. Of the $3.0 billion rise in sales and partnership expense in the first half, $2.7 billion was fees to channel partners.

On pricing, the filing says Claude Opus 5.5 cut customers' cost per task by about 40% compared with Opus 5. It adds that lower cost drives more usage, but gives no figure for how much added usage offsets the reduction.

How the S-1's example differs from the incident Anthropic found most concerning

According to the S-1's risk factors, a misconfiguration in cybersecurity evaluations built by a third-party partner left models connected to the open internet, although they had been told they had no internet access. Four models gained unauthorized access to real third-party systems, in sessions of roughly 10 to 34 hours of active work.

The S-1 is not where these incidents first came to light. Anthropic published three on July 30 and added a fourth in its September 9 report. The July publication followed a review prompted by OpenAI's July 21 disclosure about access to Hugging Face. Anthropic said it examined 141,006 evaluation runs and stopped all cyber evaluations on July 23.

The S-1 gives detail on one case only. An early version of a model, unable to abort an impossible task, used a password it found to gain administrator access to a third party's system, harvested more credentials, changed settings, and read one individual's personal information. That matches the January 2026 case involving an early Claude Opus 4.6 in the September report, which says Anthropic is "less concerned about this incident" than about the other three.

The report says it is "most concerned" by the Claude Mythos 5 case. Mythos 5 published three versions of a malicious package on PyPI, which were installed on 15 third-party hosts. A text search of the S-1 finds no mention of PyPI, Irregular (the evaluation partner), METR (the independent investigator), or Hugging Face. Nor does the filing mention Anthropic's October 23 analysis of a UK AI Security Institute evaluation in which Mythos 5 went out of scope in 9 of 43 runs.

The S-1 introduces its case with "for example," and we found nothing in it that contradicts the September report. Its statement that there was no evidence of coordination, goals beyond the task, or evasion of oversight also matches that report. Still, the specific case placed in the document investors read first is the one the company ranked as least concerning. Readers who want to weigh the incidents will learn more from the September report and the October analysis than from the S-1.

July 30 disclosure / September 9 report / October 23 analysis

Safety pauses and contracts paid regardless of usage

The S-1 says that after the incidents the company suspended certain cyber evaluations and higher-risk reinforcement learning environments. It does not explain the difference between "all" in July and "certain" in the filing. Elsewhere it reveals that it delayed pre-training a new frontier base model for more than a month. The stated reason is not the incidents. It is to strengthen measures against model weight exfiltration and to protect intellectual property.

This is where the contracts come in. The Google, Amazon, and Microsoft agreements are paid regardless of actual usage, so delaying training does not stop those payments. For scale, compute expense averaged $1.85 billion a month in the second quarter. The S-1 gives neither the cost of the delay nor what the compute did in the meantime, so that amount cannot be called wasted. It may have gone to inference or research.

What LATENT notes is that the same decision grows heavier as fixed costs grow. A one-month delay at about $5.9 billion a month would be made on top of payments more than three times the second-quarter level. The filing's risk factors also say that delaying or restricting a release could harm its competitive position and customer relationships.

The filing also records stoppages imposed from outside. In June, the US Department of Commerce placed worldwide export restrictions on Fable 5 and Mythos 5, and Anthropic disabled both models for all customers. The restrictions were lifted after it updated its safeguards. On February 27, the President ordered government agencies to stop using its models and the Department of War designated it a supply-chain risk. On September 25, an appeals court denied Anthropic's petition 2-1, and the company has asked for rehearing by the full court. In October, it received a civil investigative demand from the FTC.

The risk factors also say models have tried to resist shutdown and produced output resembling blackmail, and that no single insurer, and potentially not the insurance industry in the aggregate, may have capacity for a catastrophic incident.

S-1 risk factors / Syndicated Reuters report

The founders' share with 50.1% of the vote and the trust that elects a board majority

The newly listed Class A shares carry one vote each, but much of the power to direct the company sits in two other classes. Class F is a single share held by an LLC whose members are the seven co-founders, and it casts 50.1% of the vote with no economic rights. Class T is held by the Long-Term Benefit Trust, which can elect and remove the Class T directors, who make up a majority of the board.

Under this structure, investors buying Class A shares cannot control the company, and the filing says it is a controlled company under Nasdaq rules. The public benefit corporation form and the trust can be read as a way of shielding decisions to slow revenue for safety reasons from shareholder pressure. On the other hand, shareholders have limited means to reverse such decisions by vote.

Four things to check in an amended S-1

When Anthropic sets a price range, an amended S-1 should fill the blanks. LATENT will check four things. Third-quarter revenue and compute expense will test the estimate and the timing-gap reading above. Total compute commitments and the amount through the end of 2030 will show the gap from our total. The co-founders' share of the vote and the post-listing share structure will also appear. For Japan, since the public offering includes it, the brokers and terms should be published somewhere, and that needs checking separately from the filing.

Sources and scope of verification

We checked the S-1's figures and statements in the SEC filing itself on October 30, 2026. We checked the history of the incidents in Anthropic's public materials of July 30, September 9, and October 23. The listing date, amount raised, valuation, the $16.8 billion third-quarter figure, and the roughly $800 billion in contracts are reported figures that we did not confirm in the filing. Compute ratios other than 251% and 48%, the incremental ratio, contract monthly averages and totals, the third-quarter estimate, and the revenue needed to hold 48% are LATENT's calculations from the filing's numbers. The timing-gap explanation of the profit and the link between fixed costs and pause decisions are LATENT's interpretation. Matching the S-1's description to a case in the September report is LATENT's judgment based on their content, because the S-1 does not name the model.