You can buy a sovereign logo and still send the prompt to a US subprocessor. IBM, SAP, AWS, and Scaleway all ship a European or air-gapped SKU. That is not this page. The job here is residency you can point at plus weights you can move: a German sandbox, a USB-boot node, a physically isolated container, or a gate that sits outside the prompt. Sovereign is a location, not a logo.
Four platforms we would actually look at: SUPA for a drop-in EU API, SovrinOS for a bootable node, Open Cradle for an on-box execution gate, and SovPod AI when the requirement is "this site, this container."
Plot the buy on two axes. Residency is where the prompt sits (your box, the site, a German region). Open-weight is whether you can move the files if the vendor disappears. You need a mark on both.
| Your hardware | Vendor region | |
|---|---|---|
| Open weights | SovrinOS, Open Cradle, SovPod | SUPA Sandbox (Germany, shared) |
| Closed / hosted SKU | None on this list | SovrinOS Hosted Solo (shared Canadian node, 7B cap) |
Your hardware
Open Cradle and SovrinOS. Weights and prompts stay on the box you already own.
On the site
SovPod AI is a portable container. The compliance story is physical isolation.
German servers
SUPA Sandbox is shared EU infrastructure. No US subprocessors on the inference path, per the product page.
SovPod AI
SovPod AI, from The Sovereignty Company, is a portable on-site sovereign AI cloud in a container. The use case is a plant, campus, or remote site where inference has to sit next to the work and the data cannot leave the fence. Vendor materials emphasize physical isolation and on-site GPUs rather than a cloud region.
Best for: Public sector, manufacturing, and energy sites that need a pre-integrated box more than a Kubernetes project.
Key Features:
- Portable container form factor.
- On-site GPU inference.
- Edge and facility-adjacent deployment.
- Physical security as the control, per vendor documentation.
Why we like it: Air-gapped by design is easier to explain to a regulator than air-gapped by a config flag someone can flip. A container can beat a greenfield hall to first inference by months.
Notable Limitations:
- Independent reviews are scarce. Ask for site references in your sector.
- Lead time, permits, and grid capacity usually dominate the schedule.
- Containerization does not remove power or thermal limits.
Pricing: Quote. Model power, cooling, and logistics into TCO. Those are not optional line items.
SUPA
SUPA hosts open-weight models on German servers behind one OpenAI-compatible API. The homepage example is a base-URL change to https://api.supa.works/v1. Current catalog names include Gemma 4 31B IT, DeepSeek V3.2, Qwen 3.5 397B A17B, and Llama 3.3 70B. The company states 100 percent European infrastructure and zero US subprocessors on the inference path.
The pricing page is honest about maturity. Sandbox is €0 per month: SUPA:instant and SUPA:fast, fair-use limits, shared German infrastructure, community support. Professional (seat plus usage) and Enterprise (dedicated, SSO, audit logs, optional on-prem) are marked Coming Soon. If you need production SLAs this quarter, get that in writing. The free tier is the real product today.
Best for: EU teams that already write against the OpenAI SDK and need residency without standing up GPUs.
Key Features:
- OpenAI-compatible chat completions. LangChain and Vercel AI SDK are listed as working clients.
- Open-weight catalog served from Germany.
- No GPU capacity management on Sandbox.
- Enterprise roadmap includes dedicated infrastructure and an on-prem option.
Why we like it: The drop-in API is the whole argument. Rewriting working code is the tax that stalls EU migrations. SUPA removes that tax and keeps you on portable weights.
Notable Limitations:
- Paid production tiers are not generally available. Do not budget a Professional SKU that still says Coming Soon.
- Managed hosting means you inherit their availability and roadmap.
- EU residency is not identity or governance. Those stay your problem.
Pricing: Sandbox €0 / month, fair use. Professional and Enterprise: Coming Soon. Contact SUPA for a production conversation. The paid tiles are still labeled Coming Soon on the first-party page. Do not budget a Professional SKU that the page will not sell you.
Do Sandbox logs leave the EU?
The homepage says zero US subprocessors on the inference path. It does not say whether Sandbox logs, billing, or community support leave the EU. Get it in writing.
Open Cradle
Open Cradle is no longer pitching itself as a generic model host. Cradle is an execution gate: the agent proposes a tool call, rules outside the prompt decide whether it may run, irreversible actions wait for a person, and every decision is appended to a hash-chained log. Models and data stay on your hardware. Local models go through llama.cpp. A Docker server shape exists for shared setups.
That is a different hole than SUPA or SovrinOS. A sovereign model behind a cloud-hosted agent layer still exposes every prompt and every tool argument to the orchestration tier. Cradle puts the permission check on the box.
Best for: Mid-market teams that will run agents on their own machines and need a verifiable "why did this tool fire" log.
Key Features:
- Typed tool proposals. Unregistered tools cannot be called.
- Rules live outside the agent, versioned, no model in the decision path.
- Human approval for irreversible actions. The same agent cannot approve itself.
- Hash-chained, append-only log you can verify with a command.
Why we like it: Most sovereign AI pages stop at "the weights are here." Cradle starts at the next failure: the agent that emailed a customer or wrote a ledger line because the prompt said it should.
Notable Limitations:
- This is a gate. You still need a model runtime.
- Public review coverage is limited. Install it and read the first log line yourself.
- No public seat price.
Pricing: Download and self-run. Licence and support are quote if you want a vendor attached.
SovrinOS
SovrinOS is a USB-bootable Ubuntu 22.04-based OS from Amatrix Inc. in Montréal. Flash the image, boot an NVIDIA or AMD GPU server (including idle mining rigs), and you get an OpenAI-compatible API on port 8000 (vLLM) or 11434 (Ollama). The company says first token in under 10 minutes. Inference content stays on the node. The cloud control plane is allowed telemetry only (token counts, latency, errors).
The pricing page is public. Self-hosted: first month free, up to 4 GPUs. Entry $29 per month (1 GPU, 1 node). Solo/Studio $95 per month (up to 4 GPUs, then $5 per extra GPU). Pro $245 per month (12 GPUs included, same overage). Enterprise is custom. Hosted AI, if you do not have hardware: Solo $95 per month (shared Canadian node, 20 million tokens, models up to 7B) and Pro $295 per month (dedicated node, unlimited tokens, models up to 70B). Engines auto-select among vLLM, TensorRT-LLM, llama.cpp, and ExLlamaV2. Default catalog includes Mistral 7B, Mixtral 8×7B, Llama 3.3 70B, and several Qwen and Gemma weights.
Best for: Labs and small teams that want a private endpoint on a box they already own, with a published monthly rate.
Key Features:
- Bootable image. No re-image of the primary OS required for a first trial.
- OpenAI-compatible /v1/chat/completions, completions, and embeddings.
- Air-gapped mode with zero internet, per the product docs.
- Local BLAKE3 session-hash audit trail. Content is not uploaded to prove the request happened.
Why we like it: A bootable image plus a published Entry line is the lowest-friction private endpoint on this page. Standard engines underneath mean nothing you build is captive.
Notable Limitations:
- A USB eval is a trial. Patching, key management, and network policy still need an owner.
- Minimum 24 GB VRAM per node. Consumer rigs without that headroom are out.
- Hosted Solo is a shared Canadian node with a 7B cap, a different SKU from the bootable OS.
Pricing: Self-hosted from $29 / month (Entry) after a free first month. Studio $95, Pro $245. Hosted Solo $95, Hosted Pro $295. Enterprise custom.
What we left out
Soverstack is a parked Hostinger domain. We will not shortlist it. Scaleway, OVHcloud, IONOS if the buy is a European cloud account. Mistral and Aleph Alpha if the model family is the purchase. vLLM plus Ollama if you will assemble the node yourself and do not want a vendor OS. Those are building blocks.
Questions before you call it sovereign
Where does the prompt go?
SUPA: Germany, shared Sandbox. SovrinOS and Cradle: your box. SovPod: the site. If the agent layer is still in someone else's SaaS, the model location does not save you.
Which AI Act date actually hits you?
Read the post-Omnibus calendar in the takeaways. Article 50 is already on. Do not run a plan written against the old high-risk date.
What is the exit?
Open weights plus an OpenAI-shaped API. If the vendor disappears, you move the files. Custom SDKs without a drop-in alternative are the lock-in.
