← Blog

your prompts don't belong in virginia — the case for eu-sovereign inference

Think about what your organization actually sends to an LLM API. Draft contracts before they're signed. Patient notes being summarized. Customer complaints, source code, board memos, HR cases. Prompts are not telemetry — they are a live feed of the most sensitive text your company produces, serialized into JSON and shipped to whichever region your provider terminates in. For most European companies, that region is Northern Virginia.

That should bother you more than it does.

two laws that cannot both be obeyed

The structural problem is well known to every DPO and still routinely waved away. The US CLOUD Act compels US providers to hand over data in their possession, custody or control regardless of where it is stored. A US inference provider's Frankfurt region is a comfort blanket: the data sits in Germany, the legal obligation sits in Washington. The GDPR, meanwhile, prohibits exactly that kind of disclosure to third-country authorities outside proper legal channels (Article 48).

A US-controlled processor handling EU personal data lives permanently between those two obligations. When they conflict, you — the controller — carry the compliance risk.

The Court of Justice already told us how this ends. Schrems II (2020) invalidated Privacy Shield precisely because US surveillance law offers EU data subjects no effective redress. Its successor framework is under the same legal attack, for the same reasons, and every legal team knows a transfer mechanism that has been struck down twice is not a foundation — it's a countdown. Standard contractual clauses don't cure it either: paper commitments can't override a statute that binds the provider at home.

Prompts raise the stakes further. Unlike a database backup, a prompt is plaintext by necessity — the GPU must read it to answer. You cannot encrypt your way out; you can only choose whose jurisdiction the plaintext passes through.

what "eu-sovereign" means here — precisely

"Sovereign" has become a sticker US hyperscalers put on a region. We use it to mean something checkable. On ElephantPool's EU-Sovereign tier:

  • Attested EU machines only. Your requests route exclusively to machines physically located in the EU — attested or under direct contract — never to the general pool. Routing is enforced by the scheduler, not by a promise in a PDF.
  • No transfers, because there's nothing to transfer. Prompts and completions are processed in RAM, never written to disk, never used for training, and never leave EU territory. No transfer mechanism is needed when no transfer occurs.
  • European operator, European law. The network is operated by De Buck Technologies, a Belgian company, under Belgian jurisdiction. No US parent, no CLOUD Act exposure by construction.
  • A standard DPA, ready for your Article 28 checklist and your next audit — not a bespoke negotiation.

And because everything runs on open-weight models — Qwen3.8, GLM-5.2, DeepSeek V4-Flash and friends — there's no black-box model provider upstream of your data either. You can inspect exactly what processes your text.

who actually needs this

If you're building a hobby chatbot, honestly, run it anywhere. EU-sovereign inference is for the workloads where "probably fine" is not an answer:

  • Healthcare — clinical notes and patient communication are Article 9 special-category data; a US processor is a finding waiting for an auditor.
  • Legal — privilege does not survive contact with a subpoenaable third party in another jurisdiction.
  • Finance — DORA and national regulators increasingly ask exactly where inference happens, not just where data is "stored".
  • Public sector — several member states now bar non-EU processors for citizen data outright. This isn't a trend to hedge; it's procurement reality.

sovereignty without the sovereignty tax

The usual objection is price: European sovereign clouds charge a hefty premium for the flag. ElephantPool's distributed pool inverts that — efficient community and fleet hardware across Europe, an 80/20 split instead of datacenter margins, and per-token prices that undercut the sovereign clouds at roughly half their rates. Compliance and cost pointing the same direction, for once.

Your prompts are the most honest record of your business that exists. Keep them under a legal system that answers to you.

run it in europe

Get API access, flip your traffic to the EU-Sovereign tier, and request the standard DPA from your dashboard.

Get API Access Attested EU routing · zero retention · standard DPA

Read next

one model, every card: how a 27b runs on 8 gb

Qwen3.8 27B ships in five builds, from Q5_K_M on a 24 GB card down to IQ2_XXS on 8 GB. Why we split models from builds, what quantization actually costs, and how a quantized KV cache doubled our context overnight.

we stopped projecting your earnings and started measuring them

Most GPU marketplaces advertise what a card could earn. Our dashboard shows what yours measured, and refuses to show a number until it has one. Here is why that choice cost us a nicer-looking page and was still right.

Model catalogue · Knowledge base · Run a GPU