OpenNodes for Europe
OpenNodes is an open protocol and an open-source registry for finding and verifying AI inference endpoints — from a university GPU cluster or an AI Factory service to a regional cloud. Providers publish a signed self-description; registries check those claims automatically; clients call the provider directly. OpenNodes is never in the data path.
Europe is building inference capacity in many places at once — EuroHPC AI Factories and Antennas, national centres, sovereign and regional clouds — but there is no open, vendor-neutral way to discover and compare it. The usual shortcut is a commercial broker in the middle of every request: prompts pass through a third party, often outside the EEA, and trust rests on the broker's own claims. ONP standardizes discovery and verification as an open layer, and leaves invocation (OpenAI-compatible), tools (MCP) and payment to existing standards.
| Need | How ONP addresses it | Status |
|---|---|---|
| Keep prompts with the chosen processor (GDPR, transfer minimisation) | Discovery is metadata only; clients call the provider directly. The hosted registry never receives prompts (one documented exception: the web advisor box). | Live |
| Choose providers by jurisdiction | Registered nodes declare the operator's country in their signed card; for imported listings of known European providers it is inferred and labeled as such. An EU/EEA filter works in registry search, the advisor, the MCP tools, the OpenAI-compatible gateway (onp/auto-eu) and the Ollama router (auto-eu). No country counts as outside. The registry also checks where each server is (IP geolocation and network, round-trip times from Nuremberg and Helsinki) and drops a server found outside the EEA from the filter — evidence, not proof. | Live |
| Know what you are using (AI Act transparency, deployer due diligence) | Signed, versioned Node Cards: model, operator, country, hardware, data policy (retention, training on inputs). | Live |
| Evidence instead of claims | Automated verification: serves the claimed model, context window holds (or is capped), measured latency and throughput under load, continuous blind probes. Claimed, measured and attested values are labeled and never blended. | Live |
| Auditability of use and cost | Every response carries a provider-signed receipt (usage × price pinned to a card revision); silent repricing is rejected. | Live |
| No lock-in, easy switching (EU Data Act) | OpenAI-compatible invocation; registries are a protocol, self-hostable, with a signed export feed for mirroring; open licences. | Live |
| Independent third-party assurance | Attestations as verifiable credentials: audits, benchmark runs, residency. | Planned |
| Federation across institutions | Several registries exchanging signed feeds (the feed itself is live). | Draft |
What it is not: OpenNodes does not certify compliance with the AI Act or GDPR. It makes the facts deployers need visible, signed and checkable; the obligations stay with providers and deployers.
A common discovery and verification layer for the inference services they expose — find, compare and call them through one open interface. Reference fingerprints for open-weight models could be computed on EuroHPC systems.
A free, neutral channel to be found and verified, with measured performance and disclosed hardware, without a broker in the traffic or taking a cut.
Signed self-descriptions plus verification are close in spirit to Gaia-X self-descriptions; a mapping is future work.
The registry lists European models wherever providers serve them — Mistral (370+ offerings) and Poland's Bielik today — and the EU filter keeps traffic with EEA operators. Our Finnish reference node serves Bielik 1.5B v3 (SpeakLeash, Apache-2.0), fully verified: identity matched and the whole 8,192-token context passed retrieval tests.
ONP Draft 0.2 (attestations, federation, composite nodes), a W3C Community Group for the specification, and pilots with European operators and AI Factory service teams. Interested? operators@opennodes.io