
Anthropic’s S-1, filed with the SEC on 26 October 2026, is the most detailed public map yet of what frontier AI costs and who carries the risk. Against $16.3 billion of first-half revenue, Anthropic signed about $662 billion of new minimum compute commitments in 2026. And one risk factor says a US export restriction once switched off two models for every customer worldwide.
Most readings of the filing are valuation stories. I read it as a supplier’s architecture document: where are the single points of failure, and who can turn the service off? Disclosure: at Jaydus.AI and MediVox we use Anthropic models alongside other vendors, so this is my own supplier’s fine print. My thesis: for a European CTO, this S-1 is a free vendor-risk assessment, and five passages belong in your DPIA file and exit plan.
- Filing: Form S-1 for Anthropic, PBC, filed with the SEC on 26 October 2026; Class A shares to list on Nasdaq Global Select under the ticker “ANT”.
- Revenue: $16.3 billion in the six months to 30 June 2026 (unaudited), up from $1.2 billion a year earlier; $4.6 billion for all of 2025.
- New compute commitments: about $417 billion signed between 1 January and 1 June 2026, and about $245 billion more between 30 June and 26 October 2026.
- Take-or-pay contracts: Amazon at least $110.0 billion, Google about $114.5 billion and Microsoft at least $44.9 billion, each “regardless of actual usage”.
- Export shutdown: in June 2026 the US Department of Commerce imposed worldwide export restrictions on Fable 5 and Mythos 5, and Anthropic disabled both “for all of our customers”.
- Europe: a separate retail offering in unnamed EEA member states, but no European listing and no EU data centre named in the filing.
What does Anthropic’s S-1 say about compute commitments?
The numbers sit in the subsequent-events notes, not in the headline tables. Note 16 of the audited 2025 accounts says that after 31 December 2025 the company entered into agreements “with aggregate future minimum payment obligations of approximately $417 billion”, mostly leases for compute equipment and data centres plus purchase commitments with cloud providers, with terms “from less than one year to 20 years”. Note 17, which is unaudited, adds “approximately $245 billion” signed after 30 June 2026, with terms up to 20 years and commencement dates running to 2036.
These are two windows of newly signed deals, not a balance owed, and they overlap with the $265.6 billion of fixed or minimum commitments the filing reports as of 30 June 2026. At the end of 2025 the equivalent total was $54.6 billion. On top of that, Anthropic estimates $125 billion to $150 billion for expanded capacity delivered in 2028–2031, where “certain commercial terms have not yet been finalized”. The S-1’s own summary line, “an aggregate committed spend of more than $”, is still blank and will be filled in by an amendment.
Which contracts bill “regardless of actual usage”?
Three named contracts are take-or-pay. The Amazon agreement from April 2026 commits Anthropic to “at least $110.0 billion between May 2026 and April 2036, regardless of actual usage”, with a true-up if spend falls short and no cancellation except for Amazon’s uncured material breach. Google’s runs to July 2033 at about $114.5 billion; Microsoft’s to May 2033 at $44.9 billion. Other deals are softer: SpaceXAI is “up to $120.5 billion” with 90-day termination after initial periods; AMD has no fixed sum.
The mirror image: CNBC reported on 29 October 2026 that Amazon still expects about $220 billion of capital expenditure in 2026. A ten-year floor from one customer helps finance that. The risk is not removed, only spread between lab, cloud and your price per token.
Why read an S-1 as an architecture document?
Because it describes the physical system behind the API better than any whitepaper. Anthropic says it trains and serves Claude “natively across four major chip families (AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs)” through a proprietary orchestration layer that “reduces dependence on any single vendor”. That is multi-vendor strategy at the silicon layer, and bargaining power at the next negotiation.
The same document is candid about leaving a cloud: it “would require significant time and expense” and could mean having to “re-architect portions of our infrastructure, retrain or migrate models and large datasets”. If a lab with its own orchestration layer says that, a hospital that wired one model into a clinical workflow should assume its exit is harder than the slide deck says.
Then there is power. The S-1 notes that “certain of our planned data center facilities rely on dedicated on-site natural gas-fired generation”, with fuel costs borne “on a pass-through basis”. I argued in why power is the real constraint on AI that electricity sets the pace; here a lab confirms it is building around the grid.
What happened when the US restricted Fable 5 and Mythos 5?
This is the passage every European buyer should print out. The risk factor on trade controls says restrictions “have already affected our offerings directly. For example, in June 2026, the U.S. Department of Commerce imposed worldwide export restrictions on our Fable 5 and Mythos 5 models. To ensure compliance, we disabled Fable 5 and Mythos 5 for all of our customers; access to our other models was not affected.” They returned once safeguards were updated and the restriction lifted.
The filing does not say how long the outage lasted. CNBC’s live coverage on 26 October 2026 reports that access returned “after roughly two weeks of tense negotiations”. The S-1 warns that similar controls “could recur, could remain in place longer, or could extend to additional models”.
Two lessons follow. An availability risk you cannot control now has a dated precedent, which moves it from boilerplate into your risk register. And other Anthropic models stayed up: a same-vendor fallback would have covered this incident; a second vendor covers the next kind.
Which S-1 risk factors matter to a European enterprise buyer?
Most risk factors are boilerplate. These five are not, because each has already happened or is structural:
- Government-imposed unavailability. The June 2026 shutdown above. Your exit plan needs a tested path for “model withdrawn worldwide with no notice”.
- Model rotation. Model versions released in the six months to 30 September 2026 “accounted for over 80% of our revenue” in the third quarter, and Anthropic says it has changed, and may change, “usage limits, rate limits, or model and product capabilities”. Budget re-validation several times a year.
- Cloud concentration and data location. Delivery depends on Amazon, Google and Microsoft. The filing names data centre leases in Australia and Canada and says Anthropic is “pursuing” other countries; no EU or EEA site is named. Where European inference runs is a contract question, which is why Norway’s data centre queue matters.
- Incident reporting. The S-1 notes that the EU AI Act requires providers of general-purpose AI models with systemic risk to report serious incidents “without undue delay to the AI Office”, alongside NIS2’s staged notifications. According to the European Commission’s GPAI guidelines, the Commission’s enforcement powers have applied since 2 August 2026. Ask how those reports reach you (context: what the AI Act requires as of October 2026).
- Uninsurable tail risk. “The insurance market for AI-specific risks is nascent and evolving,” the filing says, and Anthropic may be unable to obtain cover “at all”. If the vendor cannot insure it, your contract should not assume someone else has.
Why does a securities filing belong in a DPIA?
A DPIA assesses risks to people’s rights, and availability, data location and incident handling are part of that when a model processes personal data. Datatilsynet’s guidance stresses that organisations must assess the privacy consequences of the solutions they adopt, and Article 35 of the GDPR sets when a DPIA is mandatory. An S-1 is written under legal liability for omissions, which makes it better evidence than a trust-centre page.
For financial institutions the case is stronger. EIOPA notes that DORA, in application since 17 January 2025, targets “systemic and concentration risks arising from the financial sector’s reliance on a limited number of ICT providers”. An AI vendor that itself depends on three hyperscalers is a concentration risk stacked on another one.
What does the $80.7 billion loss actually mean?
Less than the headline suggests. Of the $80.7 billion first-half net loss, $72.4 billion is a change in the fair value of convertible notes and similar instruments, driven by the rise in Anthropic’s valuation. The operating loss was $693 million, down from $2.1 billion a year earlier. More useful for a buyer: second-quarter revenue grew 1,368% year on year while compute and infrastructure expenses grew 300%, and $13.1 billion of first-half revenue was consumption-based. That fits the frontier price curve: falling prices per token, fixed commitments underneath.
Can European retail investors buy shares in the Anthropic IPO?
Possibly, depending on the country. The S-1 says the offering “includes public offerings in Australia, Canada, certain member states of the European Economic Area, Japan, Switzerland, and the United Kingdom”, and that shares will go to non-qualified investors in “Covered EEA Countries” through a separate prospectus to be approved by a national competent authority. The filing does not name those countries, and Anthropic does “not currently intend to list” on any exchange outside the US.
Per Anthropic’s announcement of 26 October 2026, the share count and price range “have not yet been determined”. Norway is an EEA state, but as of November 2026 nothing I have read confirms it is covered. This is not investment advice.
What I would take from this
At Jaydus.AI the product is routing across many models, so the shutdown risk is part of the design rather than a footnote. At MediVox, where clinical transcription cannot simply stop, the rule is the one I set out in how I choose models for production: no clinical workflow should depend on a single model being available. The S-1 turns that from engineering caution into a documented precedent. I would ask every team running a frontier model to:
- Record the five passages above in the supplier assessment, with the S-1 URL and filing date as evidence.
- Keep at least one qualified fallback from the same vendor and one from a different vendor, and fail over to both in a drill each quarter.
- Write “withdrawn by government order” into the exit plan as a scenario with a target recovery time, not as force majeure.
- Pin model versions in contracts and budget re-validation for each rotation.
- Keep prompts, evaluations and logs portable in a layer you control; the neutral layer is getting bought.
Three predictions, dated so you can check them. The amended S-1 will fill in the blank aggregate committed spend with a figure above $500 billion. By 31 December 2027, at least one major European bank or health authority will publicly require government-restriction scenarios in AI supplier contracts. And Anthropic will announce EEA data centre capacity before the end of 2027, because the retail offering and the sovereignty risk factors point the same way.
Frequently asked questions
What does Anthropic’s S-1 say about compute commitments?
It discloses about $417 billion of new minimum payment obligations signed from January to 1 June 2026 and about $245 billion signed after 30 June 2026, mostly for compute and data centres. Contracts with Amazon, Google and Microsoft must be paid “regardless of actual usage”.
When is the Anthropic IPO and what is the ticker?
Anthropic intends to list its Class A shares on the Nasdaq Global Select Market under the symbol “ANT”. As of 1 November 2026 the price range and share count had not been set, so no date is confirmed.
Were Fable 5 and Mythos 5 really switched off for all customers?
Yes. The S-1 says Anthropic disabled both models for all customers in June 2026 to comply with US Department of Commerce export restrictions, and restored them after the restrictions were lifted. Other Anthropic models were not affected.
Is Anthropic building data centres in Europe?
The S-1 names data centre leases in Australia and Canada and says Anthropic is pursuing capacity in other countries outside the US. It does not name any site in the EU or EEA.
Why should an AI vendor’s S-1 be part of a DPIA or supplier assessment?
Because it is written under legal liability and discloses availability, concentration and regulatory risks that marketing material leaves out, and those risks affect continuity of processing and incident handling.