The numbers in the prospectus
Anthropic, the company behind Claude, is preparing an initial public offering in the United States. Details from its S-1 were reviewed and reported in late September 2026, and the figures are unusual even by the standards of this industry. Revenue for 2025 came to $4.59 billion, roughly twelve times the previous year. The listing target has been raised to $2 trillion, against a private market valuation of $965 billion in May 2026, $183 billion in September 2025 and $61.5 billion in March 2025. The Economic Times summarised the disclosures.
Divide the target by trailing revenue and the multiple lands near 436. Nothing about that number is normal, but it is also not the number that should worry a European buyer of Claude.
What the $42 billion loss really contains
The GAAP net loss of roughly $42 billion is mostly accounting. About $34 billion of it is a non-cash charge from the rising value of convertible debt held by Google and Amazon. Strip that out and the operating loss for 2025 was $8.06 billion. Against $4.59 billion of revenue, that works out to roughly $1.76 lost for every dollar earned. Compute and infrastructure spending alone reached $7.33 billion, or about 1.6 times revenue.
Those are real cash costs, not paper entries. They are also the reason the listing matters commercially: public markets demand a path to positive cash flow, and Anthropic has to explain how it gets there while paying for capacity it has already signed for.
$518 billion of contracted cloud capacity
The prospectus lists $518 billion in committed cloud infrastructure obligations for the coming years. That is about 113 times 2025 revenue. The commitments are spread over multiple years and are not a single-year bill, but the ratio is the point. Anthropic has booked future spending on a scale that assumes revenue keeps growing at a pace closer to its 2025 trajectory than to anything steady.
Revenue concentration makes that assumption less comfortable. Between 24% and 25% of 2025 revenue came from two customers, according to the filing, and those agreements are not locked in long term. An eight-figure dependency on two unnamed buyers is the sort of thing European procurement teams already screen for when they evaluate a supplier's stability.
Claude's price card, and what we can verify
Claude Sonnet 5.5, released on 28 September 2026, is listed at $2 per million input tokens and $10 per million output tokens on the vendor's own pricing page. The figure below is a list price in USD, checked on 3 October 2026 against that page; the euro column converts at roughly 0.92 EUR per dollar, so treat it as an order of magnitude rather than a quote.
| Model (list price, as of 3 Oct 2026) | Input $/M | Output $/M | 1M in + 1M out, $ | ≈ EUR |
|---|---|---|---|---|
| Claude Sonnet 5.5 (Anthropic official pricing page) | 2.00 | 10.00 | 12.00 | 11.00 |
Comparing Claude against rival flagship models is the obvious next step, and it is also where the numbers stop being checkable. We only publish list prices we can trace to a vendor's own pricing page; for competing flagships, that check was not completed at the time of writing, so there is no comparison column here. An unsourced price table is worse than no price table. What is verifiable is the shape of a monthly bill: a pipeline that burns a million input and a million output tokens costs around $12 on Sonnet 5.5. At a few hundred million tokens a month, that is a budget line, not a rounding error — and it is the number a European procurement team can put on the table in a vendor negotiation.
Legal exposure and the EU rulebook
Anthropic is defending author copyright class actions over training data in the United States, and it has been in a dispute with the US Pentagon over safety guardrails on military use. The prospectus adds a further disclosure: the company warns investors that advanced AI systems could pose catastrophic or existential risks to humanity, and flags liability from rogue agents acting on their own.
Inside the EU the timeline is easy to conflate, so it is worth separating the two dates. General-purpose AI model-provider obligations have applied since 2 August 2025: technical documentation, a copyright policy, systemic-risk mitigation and a real sanctions framework. Enforcement is the second date, not the first. The AI Office's supervision and enforcement powers became exercisable on 2 August 2026, and Article 50 transparency duties have applied since the same day — AI interactions must be disclosed and synthetic media marked. The Digital Omnibus on AI, as reported, would push Annex III high-risk obligations to 2 December 2027 and Annex I product-integrated systems to 2 August 2028; those are proposed changes, not law in force. A US listing does not change any of that, though it does add quarterly disclosure about the provider behind the model.
What a European buyer can do with this
Three practical points. First, ask for contractual exit terms and model-version guarantees, because a listed company answers to shareholders when it reprioritises a roadmap. Second, keep a second provider warm. That can be a competing API, or an open-weight model you host yourself. On our RTX 5060 Ti rig in AI Arena we keep local Ollama deployments running for exactly this kind of fallback, and MIT-licensed open weights are candidates once your volume justifies the hardware. Third, if the workload involves personal data or public-sector material, check where inference actually happens. Sovereign European deployments on local infrastructure exist for that reason.
Is Claude available in the EU, and in what currency is it billed?
Yes, Claude is available across the EU through claude.ai and the API. Anthropic lists prices in US dollars, and European business customers typically pay in USD with local VAT added, which means your effective cost per token moves with the exchange rate.
Who carries the AI Act obligations, Anthropic or the company using Claude?
Both, in different roles. Anthropic, as a general-purpose AI model provider, owes technical documentation, a copyright policy and systemic-risk mitigation — duties that have applied since 2 August 2025. You, as the deployer, owe the transparency duties that have applied since 2 August 2026: telling users they are talking to an AI, and marking generated content where Article 50 requires it.
Is $2 trillion the final listing price?
No. It is the target reported in the reviewed prospectus. The actual valuation is set during bookbuilding, when institutional investors place orders. The private figure it replaces was $965 billion in May 2026.
