Sovereign inference API · Early access
Sovereign inference.
European iron.
Zarx is an OpenAI-compatible inference API for open-weight frontier models, served from GPUs NPAW owns and operates in Spain. Your prompts, outputs and fine-tunes stay on our iron, in our IP space, under EU law.
- Early access · design partners
- OpenAI-compatible
- No US sub-processor in the data path
| name | key | limit | state |
|---|---|---|---|
| production | zx_live_····8f2c | daily tokens | active |
| evaluation | zx_live_····41d0 | monthly spend | active |
| time | model | endpoint | tokens | status |
|---|---|---|---|---|
| 18:04:12 | glm-5.2 | /v1/responses | 1 942 | completed |
| 18:04:09 | kimi-k2.6 | /v1/responses | 806 | completed |
| 18:03:51 | glm-5.2 | /v1/chat/completions | 3 118 | completed |
| 18:03:44 | gemma-4-12b | /v1/responses | 412 | completed |
The thesis
Inference has a jurisdiction.
Inference is infrastructure.
Models are becoming part of how banks answer customers, how telcos run networks, how newsrooms work. Infrastructure is judged on where it runs, who operates it and what law applies.
Infrastructure has a jurisdiction.
A prompt is data. Where it is processed decides which courts, which regulators and which foreign statutes can reach it. Most inference today is processed by a handful of US companies, or on servers they operate.
Zarx keeps inference on our iron.
Open-weight frontier models, served from GPUs NPAW owns and operates in Spain, inside our own IP space, under EU law. Prompts, outputs and fine-tunes do not leave.
What Zarx is
The API you already integrate, on iron you can point to on a map.
Four things that are true today, not on a roadmap.
OpenAI-compatible.
The official SDKs work by changing the base URL. Responses and Chat Completions endpoints, streaming, function calling, structured JSON output, idempotent retries.
Open-weight frontier models.
Models we run today include GLM-5.2, Kimi K2.6 and Gemma 4. Published weights, published licenses, pinned checkpoints. No remote model API behind the curtain.
Our own iron, our own network.
AMD Instinct MI300X accelerators in production on servers NPAW owns, with NVIDIA GB300 arriving. Traffic terminates in NPAW’s own IPv4 allocation in Spain.
Accounted for, per key.
Keys belong to accounts. Each request is recorded with its model and token counts; limits apply per key, per model or per account. You always know who used what.
In practice
Two requests, one session.
A streamed chat completion with curl, then the same with the openai Python SDK. Change the base URL, keep your code.
1 · a chat completion with curl, streamed
curl -N https://api.zarx.ai/v1/chat/completions \
-H "Authorization: Bearer $ZARX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5.2",
"messages": [
{"role": "user", "content": "Where is this request processed, and under which law?"}
],
"stream": true
}'event: response.output_text.deltamodel: glm-5.2
On GPU servers that NPAW owns and operates in Spain, inside its own IP space. Spanish and EU law apply, including the GDPR. No US provider sits in the data path for inference.
event: response.completedfinish_reason: stop · usage: 24 prompt + 41 completion = 65 tokensapi.zarx.ai · NPAW-owned GPUs, Spain · sample output
2 · the same, with the openai Python SDK
from openai import OpenAI
client = OpenAI(base_url="https://api.zarx.ai/v1", api_key=ZARX_API_KEY)
with client.responses.stream(
model="kimi-k2.6",
input="Draft the data-residency clause for our inference vendor.",
) as stream:
for event in stream:
if event.type == "response.output_text.delta":
print(event.delta, end="")event: response.output_text.deltamodel: kimi-k2.6
Vendor shall process all prompts, model outputs and fine-tuning data exclusively on infrastructure owned and operated by Vendor within the European Union, and shall not engage any sub-processor established outside the EU for such processing.
event: response.completedfinish_reason: stop · usage: 19 prompt + 47 completion = 66 tokensapi.zarx.ai · NPAW-owned GPUs, Spain · sample output
Sample output. The requests are in the shapes the API accepts; the replies are illustrative. The base URL is the design target for early access; access is by invitation.
Models
Models we run today include
Open weights, published licenses, pinned checkpoints. The launch catalog for design partners is confirmed with each partner.
Sovereignty
Where your data goes, and where it stops.
- Operated by NPAW.
- A Barcelona company that has run its own data infrastructure for streaming services since 2008, and runs this inference gateway in production for its own products.Who we are
- Your jurisdiction, by design.
- Spanish servers, a Spanish operator, EU law. Designed to help you meet GDPR data-residency requirements: inference does not involve a transfer outside the EU.The data path
- We publish what we can prove.
- No latency, throughput or price figure appears on this site without evidence you can check. Every statement traces to a source.What we do not do yet
Sources: NPAW infrastructure inventory, 2026-09-03 (server-by-server record of NPAW-operated GPU hosts); RIPE RDAP lookup of 193.111.209.0/24, 2026-09-04 (assigned to ES-NICEPEOPLEATWORK, country ES).
Early access is open to a small number of design partners.
Direct access to the engineers who run the platform, a say in the model catalog, and a data path you can describe to your regulator in one sentence. If inference has to stay in your jurisdiction, tell us what you run.