The dependency nobody procured: what Europe can actually document about Anthropic

I. The mechanism: capability arrives as a procurement line, not as a factory

Industrial sovereignty debates in Europe are shaped by objects you can photograph. A lithography machine is in Veldhoven or it is not. A 300 mm fab is in Crolles or it is not. A gigafactory is in Isère or it is not. That physicality is why European industrial policy is comparatively good at semiconductors, batteries and aerospace: the asset has an address, a payroll and a permit.

Frontier model capability does not behave that way. It arrives as an API endpoint inside a cloud region that a third party operates, under a contract signed by a line manager, charged to an operating budget rather than to capital expenditure. There is no ribbon to cut, no prefecture file, no industrial site. The dependency is therefore created faster than any sovereignty programme can respond to it, and — crucially — it is created at a level of the organisation where nobody is asked to think about sovereignty at all.

This is the mechanism worth naming, because it explains a pattern that otherwise looks like hypocrisy: the same European institution can fund a European champion and deploy an American model in the same quarter, without contradiction, because the two decisions are made by different people using different budgets against different criteria. One is industrial strategy. The other is a tooling choice.

Read that way, Anthropic is not a competitor to Europe's AI ambitions in the sense a rival fab is a competitor. It is closer to what enterprise software has always been in Europe: an input whose supplier is not European, whose substitutability is asserted more often than tested, and whose terms of availability are set outside the jurisdiction that depends on it.

II. The European footprint that is actually documented

On 7 November 2025 Anthropic announced it was opening offices in Paris and Munich, describing them as additions to existing offices in London, Dublin and Zürich. That is the company's own statement, in its own newsroom, and it is the strongest single piece of evidence for a physical European presence.

Independent European trade coverage the same day reported that the company “employs nearly 200 people across Europe”. We print that figure attributed, and only attributed: it is a publisher's characterisation, not a company disclosure, and no filing we opened corroborates it. Earlier, in February 2025, European coverage reported that the new Zürich office would be led by Neil Houlsby, hired from Google DeepMind — a detail that matters less as gossip than as evidence of the direction of the talent flow: European research labour recruited into a United States company's European branch.

Note what this footprint is and is not. Five European offices and a European research hire are a commercial and recruitment presence. Nothing in the record we opened describes European model training, European compute owned by Anthropic, or a European legal entity holding the model weights. A sales-and-research footprint is not localisation of capability, and this dossier does not present it as one.

III. Delivery: the dependency runs through someone else's cloud

The practical question for a European CIO is not where Anthropic's offices are. It is where the inference happens and under whose contract.

The clearest documented answer is Amazon's. On 9 April 2025 AWS announced that Claude 3.7 Sonnet was available through cross-region inference on Amazon Bedrock in Europe (Ireland), Europe (Paris), Europe (Frankfurt) and Europe (Stockholm). Anthropic's own support documentation does not restate the region list; it points readers to the AWS Bedrock user guide, which is the canonical per-model region table. In other words, the authoritative statement of where a European enterprise can run Claude inside Europe is published by the cloud provider, not by the model developer.

Claude is also offered on Google Cloud's Vertex AI, including multi-region endpoints announced in April 2026. Here we hit the first real gap, and we state it rather than paper over it: in this pass we found no Google Cloud page naming a specific European Vertex AI region (for example europe-west4) for Claude models. The Vertex documentation we could open is region-agnostic. A European buyer who needs a named EU region for a Claude deployment on Google Cloud cannot get that assurance from the documents we opened, and we will not supply it by inference.

The second gap is larger. We found no Anthropic statement of EU data residency — that is, no company commitment that European customer data is processed and retained within the European Union, as distinct from the availability of models inside third-party European cloud regions. Those two things are routinely conflated in vendor conversations and they are not the same guarantee: one is a hyperscaler region topology, the other is a contractual data-handling commitment by the model developer. The absence of the second in the openable record is, for a regulated European institution, the single most consequential finding in this dossier.

A third absence deserves a line. We found no Microsoft Azure or Azure AI Foundry page listing Claude models. We do not conclude that Claude is unavailable on Azure — an absence of evidence in one search pass is not evidence of absence — but a European organisation standardised on Azure should not assume the availability it can document on AWS.

IV. Named European deployments

Three deployments meet our sourcing bar, and their shape is more interesting than their number.

N26, the German-founded pan-European digital bank, appears in Anthropic's own customer material, which states that the bank “automated up to 70% of tasks across targeted processes” and “deployed 15+ internal use cases within the first year, growing AI initiatives fivefold”. We flag the provenance precisely: this is a vendor-published case study, undated on the page, and its percentages are scoped to targeted processes rather than to the bank's operations. It is evidence that a licensed European bank runs Claude in production. It is not an audited productivity measurement, and no part of our argument treats it as one.

Piraeus Bank and Accenture announced, on 20 April 2026, a dedicated AI Hub in the Greek banking sector “supported by Anthropic”. Structurally this is the more revealing of the two banking items: the dependency is intermediated by a global consultancy, which means the European institution's relationship with the model developer is mediated by a third party that also owns the integration knowledge.

Sparkle, the Italy-headquartered international carrier, announced on 27 March 2026 a reseller agreement to distribute Claude models across Europe through Amazon Bedrock. This is the dependency acquiring a European distribution layer: a European telecom carrier selling an American model hosted on an American cloud to European enterprises. Sovereignty language rarely survives contact with that sentence.

Anthropic also announced a collaboration with Infosys on 17 February 2026 covering telecom and regulated-industry agents. Infosys is India-headquartered, so we do not count this as a European deployment; we record it because the systems-integrator channel is how regulated European operators most often adopt models.

And here is the honest boundary of the evidence: in this pass we found no named European airline and no European national government Claude deployment that meets our sourcing bar. That absence is worth as much as the presences. The documented European adoption is concentrated in banking, telecoms and consulting channels — not in public administration.

V. The counterparty's scale, and where its numbers stop being solid

A dependency is only as safe as the counterparty, so the financial record matters — reported, worldwide, in United States dollars throughout this section.

Two disciplines apply to that table. First, the valuation sequence — US$60bn in January 2025, US$61.5bn in February 2025, US$183bn in September 2025 — is a sequence of separate rounds, not a set of competing estimates of one number. We do not print it as a contradiction. We do print the roughly threefold move inside eight months, because a counterparty repricing that fast is a fact a risk committee should hold.

Second, and more important: every revenue figure above is a run-rate reported through investor briefings and people familiar with the matter, not an audited financial statement. Anthropic is a private company. A European institution making a multi-year dependency decision has no filed accounts to read. Naming that absence is not scepticism about the growth; it is the correct description of the evidence available.

Headcount is where the record genuinely contradicts itself, and we print the disagreement rather than choosing a number:

The spread is roughly 2,300 to 5,000 for overlapping periods, and no authoritative Anthropic-published headcount figure was identified at all. So we state no headcount for Anthropic. The one European figure — nearly 200 — stays attributed to its publisher and carries no argument. Anyone who prints a single global number for this company is printing an estimate with the estimate label removed.

VI. The regulatory layer: the dependency is already inside the AI Act

The European Union's answer to model dependency is not procurement policy; it is regulation of the model provider. Under Regulation (EU) 2024/1689, obligations for providers of general-purpose AI models entered into application on 2 August 2025, as the European Commission stated in its own publication of 1 August 2025 and in its GPAI provider guidelines.

Anthropic announced on 21 July 2025 that it intended to sign the EU's General-Purpose AI Code of Practice. On 1 August 2025, when the Commission published the signatory list, European press reported the United States signatories as Amazon, Anthropic, Google, IBM and Microsoft, plus OpenAI, with Meta refusing to sign and xAI signing one of three chapters.

We hold a small but real distinction here: the openable record contains an intent to sign from the company and a signatory list naming it reported by European press. We did not find a dated Anthropic confirmation of completed signature. For most readers that gap is immaterial; for a compliance officer citing the Code in a vendor file, it is the difference between a company statement and a press characterisation, and it should be checked against the Commission's own current list rather than against this dossier.

One further legislative fact belongs in any 2026 dependency assessment: the Commission proposed a Digital Omnibus on AI (COM(2025) 836 final, Brussels, 19 November 2025) to simplify implementation of the harmonised AI rules. A European institution building a multi-year model dependency is therefore building it against a rulebook that is itself in motion.

VII. What the adoption data can and cannot tell a European reader

Anthropic publishes its own Economic Index, including a geography-focused report of 15 September 2025 and later editions in January and March 2026. This is genuinely useful evidence of where model usage concentrates, and it is unusual for a vendor to publish it at all.

We must nonetheless be exact about what we did not extract. The September 2025 material we could open carries framing text and illustrative examples — a 40% figure for employees reporting AI use at work in the United States, examples drawn from US states and India — but we did not retrieve verbatim per-country Usage Index values for any named European country. Those tables live in the report body and its published dataset. A third-party site republishes the open data with a stated methodology (a Usage Index where 1.0 means a place's share of usage equals its share of the world's working-age population), but that is a secondary reading of Anthropic's dataset, not an Anthropic figure.

So this dossier prints no European country adoption number. A European policymaker who wants one should pull the Anthropic dataset directly. What the evidence supports is weaker and still important: the vendor with the most granular published view of European model usage is the vendor itself.

VIII. What we refuse to print

• Any EU data-residency guarantee for Claude. No Anthropic statement of European data residency was found. Availability in AWS European Bedrock regions is a hosting fact published by AWS; it is not a data-handling commitment by the model developer.
• Any named European Vertex AI region for Claude. The Google Cloud documentation we opened is multi-region and region-agnostic; no European region name is asserted here.
• Any claim that Claude is, or is not, available on Microsoft Azure. We found no page listing it; that is an absence in our evidence, not a finding about the product.
• Any single global headcount for Anthropic. Five published values span roughly 2,300 to 5,000 for overlapping periods and none is a company disclosure. The European “nearly 200” is printed attributed and used for nothing.
• Any audited revenue or profitability figure. Every revenue number in this dossier is a reported run-rate from investor briefings. No filed accounts exist for us to cite.
• Any European investor in an Anthropic funding round. None of the round coverage we opened named a European sovereign fund, venture firm or corporate investor. We print no European ownership stake.
• Any European country Usage Index value. Not retrieved verbatim from the Economic Index in this pass, and not reconstructed from a third-party visualisation.
• Any named European government or airline deployment. None met the sourcing bar. Documented European adoption in this pass is banking, telecoms and consulting.
• Any productivity or cost-saving figure as a measurement. The N26 percentages are vendor-published and scoped to targeted processes; they are reported as such and support no extrapolation.

IX. The conclusion a European decision-maker can actually use

The honest reading of this record is not “Europe is captured” and not “Europe is fine”. It is that European frontier-model dependency is documented at the delivery layer and undocumented at the guarantee layer. We can prove where Claude can run in Europe, who resells it in Europe, which European bank and which Greek bank use it, and that the provider is inside the EU's Code-of-Practice perimeter. We cannot show, from openable sources, a data-residency commitment, a European legal home for the capability, audited counterparty accounts, or a single trustworthy headcount.

Three consequences follow for anyone building on this in Europe. First, the assurance you need probably has to be contracted, not cited: the region list is published, the data-handling commitment is not. Second, the substitutability of the dependency should be tested and not asserted — the same Bedrock and Vertex channels that make Claude easy to adopt also make model swapping an engineering exercise rather than a commercial one, and nobody in the record above has published a European failover test. Third, and most useful for a career-facing reader: the European jobs this dependency creates are integration jobs, not model-building jobs. The Piraeus–Accenture hub, the Sparkle reseller channel, the Infosys agent programme and the N26 use-case portfolio are all demand for people who can specify, evaluate, secure and govern someone else's model inside a regulated European process.

That is the skill Europe is actually buying at scale right now. It is worth saying out loud, because it is not the skill the sovereignty debate keeps talking about.