Sovereign AI: build your own cluster, or use an API?

“Sovereign” covers three different questions

Vendors use the word for very different promises. Split it into three questions before comparing routes: each has its own answer, and no route wins all three.

• Custody — who physically holds your prompts, files and answers while they are processed, and could hand them over.
• Jurisdiction — which country's courts can order the companies holding that data to produce it. Under the US CLOUD Act, a provider must disclose data in its “possession, custody, or control”, wherever the data is stored (18 U.S.C. § 2713). Where a server sits and who controls its operator are two separate facts.
• Dependency — whose software, hardware, updates and support you need to keep the system running and improving.

A risk analysis should treat the three apart. Evanesca's sovereignty page applies the same split to its own route.

What the Palantir and NVIDIA blueprint describes

The most detailed public answer to “what would it take to do this ourselves?” is the Palantir Sovereign AI Operating System Reference Architecture, published by Palantir and NVIDIA in March 2026. It specifies a complete stack — Palantir's software on NVIDIA HGX B300 servers with eight GPUs each, linked by NVIDIA Spectrum-X Ethernet, plus CPU servers that run the platform — and states that it is designed for air-gapped and on-premises deployments, with full autonomy after the initial installation.

GPU servers come in “scalable units” of four and are ordered in multiples of four. The CPU “platform nodes” are fixed for each of the three qualified sizes:

The smallest setup, server by server

At its floor, one scalable unit, the “Small” size is already a data-centre project:

• 4 GPU servers, each with eight B300 GPUs carrying 288 GB of HBM3e memory apiece: 2.30 TB of GPU memory per server, 32 GPUs in all.
• 12 CPU servers — three for the Kubernetes control plane, nine workers for the platform's data services — each specified with at least two 32-core Intel Xeon 6530P processors, 1 TB of RAM and ten 7.68 TB NVMe drives (Dell PowerEdge R670 in the reference configuration).
• Three separate networks: a compute fabric giving each GPU 800 Gb/s, a converged network for storage and users, and a physically isolated management network for the servers' baseboard controllers. The design is Ethernet-only; InfiniBand is not supported.

Note the ratio: twelve of the sixteen servers never run a model. They carry orchestration, storage, monitoring and access control — the machinery every AI service needs around its GPUs. A managed provider builds it once for many customers; when you build, it is yours alone, sized before your first request.

Installation then follows six steps in a fixed order, from hardware procurement to platform services and activation with your own data; models come last. Rack power is a constraint of its own.

Why 32 GPUs is a floor, not a luxury

The blueprint itself warns that models beyond about 120 billion parameters may need more than one B300, and that hosting foundation models requires scaling past eight GPUs, across servers. A rough rule shows why: at 8-bit precision (FP8), the weights alone take about one gigabyte per billion parameters, before any memory is left for the context of concurrent users.

• DeepSeek V4.1 Flash, the largest open-weight model in Evanesca's catalogue at 763 billion parameters, needs about 763 GB for its weights: at least 3 B300s.
• Kimi K3, at 2.8 trillion parameters, needs about 2.8 TB — more than the 2.30 TB of a full eight-GPU server, so it cannot run on one machine at all (Evanesca does not serve it: no European operator runs it for Evanesca).

Quantising to 4 bits halves these figures, at a quality cost you have to measure on your own tasks. Either way, today's large open-weight models are sized for clusters. What open-weight AI means applies the same arithmetic to smaller models.

What owning buys, and what it leaves open

What you gain

On an air-gapped site that you run, no outside company holds your data, so none can be ordered to produce it. That is the strongest custody there is, and some defence or classified work requires nothing less. Owning also brings custom or fine-tuned weights, training, capacity that nobody else's traffic can queue in front of, and no per-token bill once the hardware is paid for.

What stays open

Dependency does not disappear. In this design the platform software comes from Palantir, the GPUs and networking from NVIDIA and the reference server configurations from Dell — three US companies — and support runs through their contracts. On a disconnected site that puts your data within nobody's reach, but your licences, security updates and next hardware generation depend on vendors you do not control.

Then come the costs no specification lists: power and cooling for GPU racks, engineers for the platform and the GPU fabric, on-call cover, and a redeployment each time a better open-weight model ships. Utilisation decides whether owning pays: an idle GPU costs as much as a busy one.

The API route: check the chain, link by link

With an API you give up custody — the operator processes your data — but you choose who that operator is. Sovereignty then depends on every company in the chain, so ask four questions:

1. Who runs the servers that run the model, and who controls that company? An EU region of a US cloud is still run by a US company.
2. Who runs the layer in between — the app, the gateway — and does it keep prompts?
3. What does each link keep, for how long, and with which exceptions?
4. Is the answer given model by model? A catalogue that mixes operators has to say which model runs where.

Evanesca's answers, stated plainly. Its models run at Scaleway in Paris (with OVHcloud in Gravelines as a fallback), in Mistral AI's EU region, at AKI.IO in Germany and at Infomaniak in Switzerland. OVHcloud, Scaleway, Mistral AI and Infomaniak are European companies with no US parent, which puts them out of the CLOUD Act's direct reach; OVHcloud and Scaleway have entities in the United States, which is why “direct” matters. AKI.IO is a German company whose shareholders are not published. Evanesca keeps no copy of your prompts or answers: they stay in memory while the answer is written, and your history lives in your browser. What the operators keep is theirs to say: Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days, and Mistral AI keeps its models' inputs and outputs 30 days to monitor abuse. Evanesca's own servers run on Vercel, a US company, in its Paris region, and keep no copy of prompts.

Evanesca serves no model outside Europe, and a model is never rerouted elsewhere: if its operator fails, the request fails. That is less custody than a cluster you own — and the passport under each answer shows the route it took.

Side by side

On Evanesca, every model has its own price per million tokens, VAT included: see pricing and the models. Accounts and keys open with the service; until then, this site shows how it will work.

How to decide

Build when at least one of these holds:

• you need an air gap, or your risk analysis rules out any third party on the data path;
• you run custom or fine-tuned weights, or you train models;
• you have sustained, saturating volume and a team that already runs infrastructure at this scale;
• the model you need is served by no operator you can accept.

Otherwise, start with an API whose chain you have checked, and measure. Real token volumes are the input every build decision needs, and few teams have them at the start. Moving later is cheap when the API follows OpenAI's format: a new base URL and key, not a rewrite. To run the routing layer yourself on top of operators you choose, see self-hosted AI gateways; for the regulatory side, what the AI Act asks.

The blueprint closes on one line: “The future of cloud is on-prem.” For an organisation that can absorb sixteen servers, it may be right. For a ten-person company or a local authority's first AI project, the first step towards sovereign AI is a choice of supplier, not a data centre.

Frequently asked questions

Read next

• European inference providers: who actually runs AI models in EuropeWho runs AI models in Europe: OVHcloud, Scaleway, Mistral AI, IONOS, STACKIT, AKI.IO, Infomaniak, Nebius and more. Company, data centres, prompts.
• Is OpenAI GDPR-compliant? What ChatGPT and the API do with your dataOpenAI offers EU contracts and transfer clauses, but it remains a US company within reach of the CLOUD Act. What that means for ChatGPT, the API and you.
• GDPR and the CLOUD Act: why hosting your AI in the EU is not enoughThe GDPR governs transfers; the CLOUD Act follows the company. Why an EU data centre run by a US group stays within reach, and how to check a provider.