Nvidia's $10B bet on Anthropic's IPO: what the token math says for EU users

What Nvidia is actually negotiating

This is not a product announcement. It is a financial engineering story with a compute bill attached, and the details matter more than the round number.

Anthropic filed a confidential draft S-1 with the SEC back in June 2026, which The Decoder also covered at the time. Since then, the reporting has shifted from "will they list" to "who gets shares before the public does". Anchor investors in an IPO of this size are not passive passengers — they get allocation certainty in exchange for signalling confidence, and they typically commit early.

The relationship between the two companies is already deeply entangled. Anthropic runs on Nvidia accelerators and has committed $30 billion to Microsoft Azure capacity powered by Nvidia hardware. Nvidia, on its side, has backed around $300 billion in guarantees to help data centres secure financing. An investment from Nvidia into Anthropic is therefore money moving in a circle: from the chip vendor, into the customer, back out as hardware and cloud orders.

The numbers, in euros

I prefer to see these figures side by side, converted, because USD amounts in the tens of billions stop meaning anything after a while. Conversion below uses roughly €1 = $1.08, and should be read as order-of-magnitude, not as an FX quote.

One ratio is worth doing in your head: a $2 trillion valuation on a $65 billion annualised run-rate is about 31× revenue. For a company that was running at roughly $9 billion annualised at the end of 2025, that is a sevenfold revenue climb in about seven months. It also means the listing price has to be defended by continued growth, not just by current numbers.

Claude is the expensive tier — here is the arithmetic

This is where the IPO talk actually touches your infrastructure budget. Claude Fable 5.1 is priced at $10.00 per million input tokens and $50.00 per million output tokens. That is not an accident; it is a premium position, and it is the pricing that produced the revenue chart above.

Let me put a realistic workload against it. Take a mid-sized European team running a document pipeline: 2 million input tokens and 500,000 output tokens per day. That is a fairly modest RAG-plus-summarisation setup, not a heavy agentic swarm.

The spread is roughly 75× between the top and bottom of that table for the same nominal token volume. Gemini 3.8 Flash is currently free or discounted through 31 December 2026, so its real cost today is lower still. This is why "which model" is not a taste question — it is a procurement decision with a two-order-of-magnitude range.

In AI Arena, our benchmark rig runs open-weight models locally on a single 16 GB card. The point is not that a local box beats Claude on hard reasoning — it usually does not — but that the per-token price for the easy 70% of a pipeline can drop to electricity and amortisation. That is the escape hatch from a premium API bill.

Europe: the AI Act paperwork is now real

The timing matters. Since 2 August 2026, the European Commission holds full enforcement powers over general-purpose AI providers: binding oversight, mandatory technical audits, model recalls, and direct financial penalties. Voluntary disclosure of synthetic media is gone too — Article 50 of the EU AI Act now requires providers and deployers to label and watermark AI-generated text, audio, image and video output.

Anthropic is a GPAI provider serving EU users. A stock listing does not exempt it from that regime, and being public actually adds a second layer: SEC risk-factor disclosure on top of European technical audit obligations. For an EU buyer, that is quietly useful. Procurement teams that have to justify a model choice under internal AI governance rules get more public documentation about model risk than they had two years ago.

What the listing does not do is change data residency. Claude is sold in the EU today through API and subscription tiers, and nothing in the reported plan suggests that changes. If your GDPR position depends on where inference runs and which sub-processors are listed in your data processing agreement, you still have to verify that yourself — a corporate event in the US does not reshuffle your Article 28 obligations.

The circular-financing question

I would flag one structural thing without predicting anything. Nvidia guarantees financing for data centres, sells chips to cloud providers, and is now reportedly prepared to take a $10 billion anchor position in one of its largest customers. In a market that keeps growing, that arrangement looks like genius. In a market that slows, every one of those links is exposed at the same time. Whether that happens is not something I can tell you — but it is the part of this story that a public prospectus will eventually have to describe in writing.

For a European developer or CTO, the practical conclusion is unglamorous: keep an abstraction layer over your model provider, benchmark the actual workload instead of the leaderboard, and price the top tier against the cheap tier for every task you route. A 75× gap funds a lot of engineering.