Skip to content
NLEN
Illustration: Anthropic's S-1/A: $17 billion Q3 revenue, $895 billion in commitments and an RSP warning

Anthropic's S-1/A: $17 billion Q3 revenue, $895 billion in commitments and an RSP warning

By Ivo Donker · — compiled with AI assistance (DeepSeek V4.1 Flash)

On 2 November 2026 Anthropic filed Amendment No. 1 to its Form S-1 registration statement (the S-1/A) with the US Securities and Exchange Commission (SEC). The original S-1 was filed on 26 October 2026 and contained financials only through the first half of 2026; the amendment adds the third quarter, the three months ended 30 September 2026. The proposed Nasdaq symbol is 'ANT'. The price range and the number of shares are still blank in the document. CNBC reports that the company 'prepares to kick off its roadshow'; no IPO date or price is known.

For builders and IT decision-makers, the S-1/A is above all a supplier-risk document. It shows how fast revenue is growing, how much compute the company has tied up, which supplier contracts are running, and which risks Anthropic itself names. This article lines up the third-quarter figures, separates what is new in the amendment from what was already in the first S-1, and explains what to watch if you build on Claude or have Anthropic as a supplier.

What the S-1/A adds

The amendment adds the third-quarter period: three months ended 30 September 2026, with the corresponding figures and risk factors. The proposed Nasdaq symbol 'ANT' is also included. The price range and the number of shares remain blank, and there is still no listing date or price. According to CNBC, the company is preparing for its roadshow.

New in the amendment are, among other things, the updated RSP risk factor and the higher minimum amount in the Google agreement: at least about $121.3 billion, against about $114.5 billion in the S-1 of 26 October. The S-1/A contains more than 90 pages of risk factors, including political headwinds, complex infrastructure needs and balancing safety and commercial objectives. On that, the co-founders say: 'We consider it our personal responsibility to help the world understand these risks, to prevent or mitigate them when they arise from our AI models, and to work with governments and the rest of the industry to address them at a societal level.'

The third quarter in figures

Revenue in Q3 2026 came in at $16,956 million, or just over $17 billion. The filing's narrative says '$17.0 billion'; CNBC says 'roughly $17 billion'. The second quarter had revenue of $11,555 million. Compute and infrastructure expenses were $9,863 million in Q3, against $5,558 million in Q2.

For the nine months ended 30 September 2026, revenue came in at $33,242 million, about $33.2 billion, an increase of roughly 1,206% year over year according to the filing; CNBC says 'roughly 1,200%'. The net loss over those nine months was $104,039 million, about $104 billion. According to CNBC, that is mostly due to the change in fair value of convertible notes and derivative liabilities. In the same period in 2025, Anthropic posted a loss of $25.1 billion (CNBC). Compute and infrastructure expenses for the nine months increased by $14.4 billion, up 329% year over year (CNBC, citing the filing). The filing says: 'As our compute and infrastructure capacity grows, the associated compute and infrastructure expenses have increased.'

Revenue versus compute: two ways of counting

The filing compares year over year itself: revenue grew 1,127% against compute and infrastructure expenses that grew 382%, for the three months ended 30 September 2026. That is the document's own comparison.

Quarter over quarter, there are no percentages to be found in the filing. If you work through the filing's quarterly table yourself, you arrive at roughly +47% revenue (16,956/11,555) and roughly +77% compute (9,863/5,558). That is a calculation by llmnet based on the table, not a percentage from the filing. CNBC writes that third-quarter revenue increased 'around 47%' compared with the second quarter, 'a steep drop from 144% growth in Q2 and 132% in Q1'.

That difference is practically relevant. If compute grows faster than revenue, the margin comes under pressure, even if absolute revenue keeps rising. Why labs buy so much compute is explained in our piece on scaling laws and compute.

Operating loss is not net loss

The operating loss in Q3 2026 was $567 million. According to the filing, that loss was 'primarily driven by $5.0 billion of compute under short-term commitments and $2.0 billion of non-cash charges related to our donation matching program'.

The net loss in Q3 2026 was $23,291 million, about $23.3 billion; CNBC says $23.29 billion. That is something quite different from the operating loss, and the two should never be confused. The operating loss sits above the line: it concerns the operating business itself, including the compute and donation-matching items mentioned. The net loss sits below the line and also includes items that have nothing to do with operating the business. For the nine months, CNBC attributes most of the net loss to changes in the fair value of convertible notes and derivative liabilities. So if you look only at the net loss, you see a much larger amount than if you look at operating performance.

For supplier risk, the distinction matters: the operating loss shows how much money the operating business costs per quarter, while the net loss is partly determined by financing structures.

Nearly $900 billion in commitments

Anthropic has entered into purchase commitments and lease agreements with an aggregate committed spend of about $895.4 billion, 'substantially all of which is related to compute and infrastructure'. About $447.5 billion of that falls due before 31 December 2030, and about $69.8 billion in the 12 months after 30 September 2026.

Those are commitments, not costs that are already in the income statement. They say something about how securely future compute has been purchased and how much money has to be reserved for it. The commitments run years ahead of the revenue they have to earn back. For a customer, this means that a large part of capacity is contractually locked in; the question is whether that capacity comes online in time and whether the pricing remains sustainable. Background on how compute networks are put together is in our piece on decentralized AI compute networks.

Google contract: at least $121.3 billion

Under the cloud services agreement with Google, Anthropic's minimum commitment is at least about $121.3 billion over the period from April 2026 to July 2033, including a true-up mechanism. The S-1 of 26 October stated about $114.5 billion. So the amount has been revised upwards in the amendment.

CNBC reports that Anthropic gets much of its capacity from Amazon, Google and SpaceX, and additionally depends on neoclouds for part of its compute. That means the company does not lean on a single supplier. For customers that is relevant: a broader supplier base can help availability, but it also means that contract terms and capacity planning differ per party.

The Google contract is one of the largest items in the committed spend. If you want to assess a supplier's risks, it is best to look at contract duration and minimum amounts, not just at revenue growth.

RSP: Anthropic warns of accelerated AI progress

In the new risk factor on its Responsible Scaling Policy (RSP), Anthropic writes that it 'may imminently meet' the threshold for 'dramatically accelerated progress attributable to AI'. No model has yet been designated as meeting that threshold, and the threshold has not been reached. This is a forward-looking statement based on internal estimates. The RSP is Anthropic's own framework that ties safety and security measures to the capability thresholds of its models.

That distinction matters: 'may imminently meet' is not a confirmation that a threshold has been reached, but a warning that, according to the company, it could happen soon. For teams that assess models, it is wise to read such statements alongside the benchmarks themselves; how to do that is in our guide to reading safety benchmarks. Background on the intersection of AI and security is in our dossier on AI and cybersecurity.

Pre-training delayed for security

The filing describes that Anthropic delayed the pre-training of a new frontier base model by more than one month in order to strengthen security against exfiltration of model weights. So this concerns countering the theft of model weights and intellectual property. Whether this passage is new in the amendment or was already in the S-1 of 26 October cannot be established; so do not call it new in the amendment.

What the filing does show is that security considerations intervene directly in the training schedule. A delay of more than a month is an operational choice with costs and slippage. For builders, this means the release calendar of frontier models is determined not only by compute and research, but also by security measures around model weights. How labs detect misuse of their models is in our piece on misuse detection at OpenAI and Anthropic.

What was already in the first S-1

The filing states: 'over 90% of the contributions to our codebase in September 2026 were written by Claude'. That is more than 90% of the contributions to the codebase (contributions to our codebase) in September 2026. This was already in the original S-1 of 26 October and is therefore not news from the amendment. It is background information, not a new disclosure.

Mind the wording: it concerns contributions to the codebase, not all code or research. Anyone who cites this sentence incorrectly turns it into a stronger claim than it is. For teams choosing models for code audits and security, it is useful to know how such figures end up in a prospectus: they say something about internal adoption, not about the quality or safety of the output. More on that is in our overview on choosing models for cybersecurity and code audits.

What this means for builders and European customers

If you build on Claude or have Anthropic as a supplier, the S-1/A is mainly a document for weighing supplier risk. Revenue is growing fast, but on the quarter-over-quarter comparison compute and infrastructure expenses are growing faster. The committed spend of about $895.4 billion runs years ahead. That says something about expected demand, but also about financing needs.

CNBC points to market jitters around the AI buildout: rising interest rates could make debt unaffordable for data-center developers and neoclouds, hyperscaler AI spending is eating into their cash flows, and some investors worry that computing capacity may not come online quickly enough for Anthropic. For you as a customer that means: follow not only model releases, but also capacity planning and contract structure. Forecasting your own API spend is a separate exercise; for that we have a guide on token usage and API spend per use case.

Anthropic is also about five years old; CEO Dario Amodei co-founded the company with researchers who left OpenAI. OpenAI confidentially filed its prospectus in June and its executives signal a 2027 debut.

What to watch

The S-1/A of 2 November 2026 makes Anthropic's third-quarter figures public: just over $17 billion in revenue in Q3, an operating loss of $567 million and a net loss of $23,291 million. The amendment also shows how large the committed spend has become (about $895.4 billion) and that the minimum commitment towards Google has risen to at least about $121.3 billion.

The RSP risk factor is the most forward-looking part: Anthropic writes that it 'may imminently meet' the threshold for 'dramatically accelerated progress attributable to AI', without a model having been designated or the threshold having been reached. For builders and IT decision-makers, the core point is to read the figures as they are meant: the filing compares year over year (revenue +1,127% against compute +382%), the quarter-over-quarter figures come from our own calculation on the quarterly table and from CNBC, and the operating loss is something other than the net loss.

Checked on 2026-11-04.

This article was compiled with AI assistance and editorially reviewed; sources are listed below.

Read also: OpenAI plans new security sprint and raises bug bounties

Sources