Anthropic Is Hiring Someone To Buy Europe

The research brief pointed me at a CNBC story this morning about an Anthropic job advert, which is not a document I would normally lead a deal note with. Job adverts are not filings. But this one is load-bearing, because Anthropic doesn't have a European data-centre footprint, and the advert, which I went and read on Anthropic's careers page before writing anything, is the public artefact that says they're about to.

The posting is for a Principal, Compute Partnerships, based in London, with a salary band of £200,000–£270,000 plus equity. The scope, in Anthropic's own words, is to "source, structure, and execute data center capacity agreements across Europe, including colocation, build-to-suit, and power purchase arrangements."¹ That is a full procurement mandate, not a BD role. "Build-to-suit" and "power purchase arrangements" are the terms you use when you're committing to gigawatt-scale, multi-year offtake. They are not the terms you use when you're renting a few racks in Dublin to reduce latency for enterprise customers.

The posting is for a Principal, Compute Partnerships, based in London, with a salary band of £200,000–£270,000 plus equity.

So the question is what this sits alongside, and the answer is: a great deal.

Earlier in April, Anthropic and AWS announced an expansion of their compute arrangement, reported variously at $100bn of commitments through the end of the decade, structured as Anthropic consuming AWS Trainium capacity and AWS continuing to invest in Anthropic.² Separately, Anthropic announced a deal with Broadcom for 3.5GW of custom-silicon capacity, to be deployed across partner data centres from 2027.³ The Broadcom release is worth reading because it does not say where those 3.5GW will sit. It says "deployed at partner data centers in North America and internationally," which is the first public confirmation that Anthropic's physical footprint is going outside the US.

Three-and-a-half gigawatts is a lot of gigawatts. For reference, the entire Republic of Ireland's data-centre sector drew about 1.8GW in 2023. Meta's largest announced campus, the Louisiana build, is planned at roughly 2GW. Anthropic is contracting for nearly double that in one deal with one silicon partner, on top of whatever Trainium capacity the AWS expansion implies. The London hire is the procurement arm for the "internationally" part of the Broadcom sentence.

The structural frame. Two of the lenses apply here and they pull in slightly different directions.

The inference economics reading is straightforward. Anthropic's constraint is no longer training, it's serving Claude to enterprise customers at margin, and the marginal customer for Claude is increasingly European, increasingly regulated, and increasingly unwilling to have inference happen on US soil. EU AI Act compliance, GDPR data residency, and the emerging pattern of sovereign-cloud procurement in France and Germany all make European inference capacity a revenue unlock, not a cost centre. The 3.5GW number tells you Anthropic is sizing for inference at scale, not training runs. Training is bursty; you can do it in Virginia. Serving is continuous and latency-sensitive; you do it close to the customer.

The AI performativity reading is the one I'd hold more loosely but worth stating. Anthropic is raising at a reported $300–400bn valuation in the round that's been circulating since late Q1.⁴ A company raising at those numbers needs to show that the capital has somewhere to go. Committing to 3.5GW of Broadcom silicon and hiring a European head of compute is, among other things, a demonstration that the round is backed by deployable capex rather than just sitting as a balance-sheet cushion. I am not saying the infrastructure push is performative, the demand-side case is real, but the timing of the public announcements, clustered in April around the round, is not accidental.

Where the frames fit and where they don't. The AI safety as market position frame would predict that Anthropic's European push leans heavily on the sovereign-compliance narrative, and it does: the careers page for the role specifically mentions "supporting our enterprise and public-sector customers in regulated European markets." That's the safety-as-moat play translated into procurement.

What the frame doesn't quite explain is why Anthropic is building its own European supply chain rather than riding AWS's European regions, which already exist at scale. AWS has Frankfurt, Dublin, Stockholm, Paris, Milan, Zurich, London. If the $100bn AWS commitment covers global Trainium capacity, Anthropic could, in principle, serve European inference through AWS European regions and be done with it.

The fact that they're hiring a principal to negotiate colocation and PPAs directly, bypassing AWS on the European build, tells you something about the AWS deal structure that isn't in the press release. Either the $100bn commitment is US-specific, or Anthropic wants optionality to run Broadcom silicon in facilities AWS doesn't control, or both. My read is both. The Broadcom chips need homes that aren't AWS data centres, because AWS data centres run Trainium. You don't put a competitor's silicon in your partner's building unless the partner has agreed to it, and nothing in either release suggests that agreement exists.

So what Anthropic is actually building is a second supply chain. AWS-Trainium for one set of workloads, Broadcom-custom for another, and the European data-centre footprint is where the Broadcom stack lives. That's a more interesting structural story than "Anthropic expands to Europe." It's Anthropic de-risking its compute dependency on a single hyperscaler while the relationship with that hyperscaler is, on paper, deepening.

What this is a case of. The pattern is the one OpenAI has been running for eighteen months: a headline partnership with a hyperscaler (Microsoft, in their case), plus a parallel infrastructure programme (Stargate) that is explicitly not on the hyperscaler's hardware. Anthropic-AWS-Broadcom is the same shape. The frontier labs are all, quietly, building out infrastructure positions that reduce their dependence on the cloud providers they are publicly aligned with. The AWS commitment makes the headline; the Broadcom deal and the London hire are what actually determines where Anthropic's compute sits in 2028.

What to watch.

• The first announced European site. My guess is the Nordics (cheap power, cold, sovereign-friendly) or Spain (renewables capacity, aggressive state incentives). Ireland is saturated. The UK is expensive and grid-constrained.
• Whether the hire ends up reporting to infrastructure or to policy. The job posting sits in the Compute org, but a principal doing European PPAs will be negotiating with governments as much as with developers.
• Whether Anthropic discloses the Broadcom deployment split between US and Europe. They haven't, and the silence is doing work.
• The next hyperscaler announcement. If Anthropic announces a European colocation deal with someone who isn't AWS, the second-supply-chain thesis hardens.

Footnotes

¶