Change one line — the baseURL — and your prompts stop leaving the EU.
Open-weight models on European bare metal, priced per token in euros. No US-owned
service ever touches your data — and we publish the proof, not just the promise.
Like the cloud you already know — a single account, a single console, a single bill — except every layer of it is European. Inference is live first; storage and dedicated compute join the same account, not a new sign-up.
Drop-in replacement for the OpenAI API, served from EU bare metal with continuous batching. Streaming first, deterministic errors, no proxying to any proprietary model.
Sovereign inference is worthless if the documents, vectors and logs behind it sit in an American bucket. Storage lands next to the GPUs that read it — same account, same keys.
Dedicated GPU capacity in European data centres for fine-tuned and private models — operated by us, with no hyperscaler underneath and no foreign jurisdiction reaching through the floor.
const client = new OpenAI({
- baseURL: "https://api.openai.com/v1",
+ baseURL: "https://api.nativekloud.eu/v1",
apiKey: process.env.NATIVEKLOUD_API_KEY,
});No sales call, no CAPTCHA, no credit card to try it. This is the whole integration, and it is what your key will unlock the day your cohort opens.
curl https://api.nativekloud.eu/v1/chat/completions \
-H "Authorization: Bearer $NATIVEKLOUD_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "nativekloud/mistral-7b",
"stream": true,
"messages": [{"role": "user", "content": "Where do my prompts go?"}]
}'
# → nowhere. Processed in an EU data centre, never written to disk.Built for agents as much as for people: machine-readable pricing, llms.txt, deterministic OpenAI-shaped errors, and docs that work without JavaScript.
Per-million-token pricing for every model, in euros, on this page and in pricing.json. No "contact us" tier, no negotiated opacity. Prepaid credits; you can never wake up to a surprise bill.
| Model | Context | Input · €/1M tokens | Output · €/1M tokens |
|---|---|---|---|
| nativekloud/mistral-7b | 32k | €0.20 | €0.50 |
| nativekloud/llama-3.1-8b | coming with GA | ||
| nativekloud/qwen-2.5-14b | coming with GA | ||
Early-access pricing — locked for early-access customers through GA. Token counts are metered per request and visible live in the console; we bill token counts and metadata only, because content is never retained to bill from.
Most "EU AI" offerings are an American platform with a European region selected and a DPA stapled on. Ownership beats geography: if the provider is US-controlled, the CLOUD Act reaches the data wherever the disks spin. So we built on EU-owned everything — hosting, GPUs, DNS, auth, billing, email — and the few dependencies that never see your data are named, not hidden. The full argument, with the subprocessor list, the data flow and our sovereignty self-assessment, is in the trust centre.
Prompt and completion content is never written to disk, logs or third parties. We keep token counts, timestamps and key IDs — enough to bill you and stop abuse, nothing that contains your data. Verified by a zero-retention audit of our own source that we publish and you can read, not asserted in a policy page.
Hosting, GPUs, DNS, auth, email, billing — even this page: EU-owned, checked for who owns it, not just where it runs. No US cloud, no CDN, no analytics, no remote fonts, no Stripe. The handful of dependencies that never see your data — our certificate authority, the GPU silicon — are named on the trust page rather than rounded down to zero.
Sign-in is an emailed one-time link. No passwords to breach, no social login, no third-party identity SaaS — auth runs on our own EU infrastructure.
Your CISO's checklist, answered before it arrives: DPA, subprocessor list and technical measures on request, plus a published self-assessment against the European Commission's own sovereignty framework — including the parts we haven't reached yet. Vendors who state their limits are the ones worth believing.
Mistral, Llama, Qwen — never a proxy to a proprietary US model API behind the curtain. Parquet, Iceberg, Postgres and S3 on the data side, so nothing holds your data hostage.
Wire-compatible with the OpenAI SDKs: migration is one line, and so is migrating away. We would rather earn the renewal than trap it.
Six principles is the summary. The trust centre is the evidence: the precise retention boundary, the CI audit that enforces it, a data-flow diagram, every subprocessor with what it can see, our self-assessment against the European Commission's Cloud Sovereignty Framework, and a plainly-worded list of what we do not claim.
We would rather show you the real state of the build than a launch date we invented. Everything below is either running, being written now, or honestly labelled as later.
The parts that decide whether this is a real cloud or a wrapper. An OpenAI-compatible gateway with key auth, token metering and SSE streaming, and a Go control plane that treats compute backends as interchangeable — so the same gateway serves a laptop in development and European bare metal in production by configuration alone.
A single GPU in an EU data centre running vLLM behind api.nativekloud.eu,
with the zero-retention claim proven by a CI audit rather than asserted in a policy page.
Llama and Qwen join the catalogue, and so does something no American provider will serve you: PLLuM, the Polish open model family built by a NASK-led public consortium and released under licences written for the AI Act. A Polish model, on Polish-governed infrastructure, priced in euros.
Today no provider offers GPU-capable AI cloud inside Poland. Changing that — through Tier III/IV Polish data centres — is the point of the company, not a footnote on the roadmap. Dedicated nodes for your own fine-tuned weights land alongside it.
Dates are deliberately absent. A one-person company that promises quarters it cannot control is a company that lies to you twice: once about the date, then again about why it slipped.
Early access opens in cohorts. Leave your email and you'll hear from a person — not a drip sequence — when there's a key waiting for you.
Your address is used to email you about early access, and for nothing else. No tracking pixels, no ad networks, no newsletter you didn't ask for.
Would rather just talk? Email Radek directly — it's a small company, so that reaches the person building it.