You think you understand your cloud privacy posture until a regulator asks you to map every data flow, retention setting, and subprocessor across borders. The surprises show up in predictable places: keeping inference inside your own VLAN, standardizing on an OpenAI-compatible API without swapping one lock-in for another, and moving agents and vector stores next to your GPUs rather than three network hops away.
The compliance calendar is the reason this became urgent in 2026, and it changed materially this year. The EU's Digital Omnibus on AI, adopted by Parliament on 16 June 2026 and the Council on 29 June 2026, rewrote the AI Act timeline. Article 50 transparency obligations still apply from 2 August 2026 and were not deferred.
But high-risk obligations for stand-alone Annex III systems moved from that same date out to 2 December 2027, and to 2 August 2028 for AI embedded in regulated Annex I products. New prohibited practices and the extension of Article 50(2) to legacy systems land on 2 December 2026.
That reshuffle bought most organizations runway, not relief: the underlying obligations did not change, and there is a grandfathering rule that resets the moment a system is substantially modified.
Meanwhile the sovereign infrastructure market has filled out fast, with IBM, SAP, AWS, and Scaleway all shipping European or air-gapped offerings. This guide compares five smaller, open-weight-focused platforms for teams that want inference under their own control.
Sovereign AI Hosting Platforms at a Glance
| Platform | Best for | Pricing model | Standout |
|---|---|---|---|
| Soverstack | Self-hosters with existing GPUs | Open source plus services | Fast path to running Llama, Mistral, and Qwen on your hardware |
| SovPod AI | Regulated or industrial sites | Custom project | Portable container for physically isolated on-site inference |
| SUPA | EU teams wanting API access without GPUs | Managed service, free tier advertised | OpenAI-compatible drop-in API on German servers |
| Open Cradle | Mid-market on-prem agent workflows | License plus support | Agent runtime and model routing inside your network |
| SovrinOS | Labs wanting a quick offline endpoint | Subscription | USB-bootable OS with an OpenAI-compatible API |
How We Evaluated These Platforms
Sovereignty claims are easy to make and hard to verify, so the assessment focused on what is checkable:
- Where inference actually runs: the jurisdiction, the legal entity, and the named subprocessors, rather than a region label on a marketing page.
- Open-weight support and portability: which model families run, and whether you could take the same weights elsewhere without rewriting your application.
- API compatibility: OpenAI-compatible endpoints mean your existing libraries work and your exit path stays open.
- Deployment control: managed, self-hosted, containerized on-site, or air-gapped, and what each demands operationally.
- Governance readiness: audit logs, model registries, change control, and update attestation, which the AI Act timeline now makes concrete rather than theoretical.
- Verifiable public footprint: independent reviews, documentation, and third-party references, which are unusually thin across this category.
The 5 Sovereign AI Hosting Platforms
1. Soverstack
Soverstack is positioned as an open-source, self-hosted sovereign cloud that runs open-weight LLMs on your own GPUs, with vendor documentation focused on hosting families like Llama, Mistral, and Qwen.
Best for: Engineering teams with in-house GPUs who want an open-source path to self-hosted LLM inference.
Key features:
- Open-source stack with no proprietary runtime dependency
- Support for popular open-weight model families
- Self-hosted cluster deployment
- Optional AI agent layer, per vendor documentation
Why we like it: Open-source control with a pragmatic focus on getting inference live on hardware you already own is the right shape for teams whose blocker is time rather than capability. When the weights and the runtime are both open, your exit path is a file copy rather than a migration project.
Limitations:
- No independent reviews, analyst coverage, or third-party references were verifiable as of July 2026, so request customer references and run a technical proof of concept before committing
- Self-hosting means you own cluster hardening, patching, and production SRE
- Dense GPU nodes routinely surprise teams on power and cooling; Uptime Institute research has repeatedly flagged rising rack density and liquid cooling requirements as a gating constraint on AI deployments
Pricing: Not publicly available. If you self-host, budget GPU capacity and facility costs. For reference, a single H100 80GB runs around $2.50 per hour on specialist clouds in 2026, roughly $10 per hour for a 4-way node, and about $36 per hour for 8x H200 at FP8.
2. SovPod AI
SovPod AI, from The Sovereignty Company, is presented as a portable on-site sovereign AI cloud delivered in a container, aimed at edge and facility-adjacent deployments where inference must sit next to critical operations.
Best for: Public sector, manufacturing, and energy sites needing on-premise or remote-site inference with physical isolation.
Key features:
- Portable container form factor
- On-site GPU inference
- Edge deployment focus
- Physical security emphasis, per vendor documentation
Why we like it: For plants or campuses where data genuinely cannot leave the premises, a pre-integrated container can beat a greenfield build to first inference by months. Air-gapped by design is a much easier compliance story than air-gapped by configuration.
Limitations:
- No independent reviews or third-party coverage were verifiable as of July 2026; treat vendor claims as unconfirmed and ask for site references
- Lead times and site readiness usually dominate the schedule rather than the technology
- Local permits and grid capacity are common gating items, and containerization does not remove power or thermal limits
Pricing: Not publicly available. Contact the vendor for a quote, and model power, cooling, and logistics into TCO rather than treating them as line items.
3. SUPA
SUPA is a GDPR-compliant sovereign AI cloud hosting open-weight models on German servers, exposed through a single OpenAI-compatible API. It targets European agencies and companies that need EU residency without operating GPUs.
Best for: EU-based teams that require European hosting and straightforward API access to open-weight models.
Key features:
- Models including Llama 3 and Mistral served from German infrastructure
- OpenAI-compatible API positioned as a drop-in replacement that works with existing libraries
- SDK with a familiar chat completions interface, so migration is a base URL and key change rather than a rewrite
- No GPU configuration or capacity management required
- Fair-use free tier for prototyping, with a path to dedicated GPUs
Why we like it: The drop-in API is the whole argument, and it is a good one. Teams under EU residency pressure usually have working code built against OpenAI's interface, and rewriting it is the tax that stalls migrations. SUPA removes that tax while keeping you on open weights, which means you can graduate to self-hosting the same models later without another rewrite. The free tier makes the evaluation cheap.
Limitations:
- Few third-party benchmarks and limited independent uptime data as of July 2026
- Managed hosting means you inherit the provider's availability and roadmap decisions
- EU hosting addresses residency but not identity and governance risk, which needs separate evaluation regardless of where the GPUs sit
- Detailed pricing beyond the free tier is not publicly itemized
Pricing: A fair-use free tier is advertised for prototyping, scaling to dedicated GPUs. Detailed rates are not published; contact the vendor.
4. Open Cradle
Open Cradle is a self-hosted sovereign AI agent platform aimed at mid-market companies, running Mistral-grade open-weight models inside your own infrastructure without requiring an in-house AI team to deploy them.
Best for: Teams standardizing on on-premise agents with connectors to enterprise chat and a REST API footprint.
Key features:
- On-premise agent runtime with model routing kept inside your network
- Open-weight model support
- REST API with chat integrations
- LoRA-style fine-tuning pipeline on your own hardware, per vendor documentation
Why we like it: The mid-market framing is the useful part. Most sovereign AI tooling assumes either a hyperscaler budget or a platform team that can operate Kubernetes and vLLM competently, and a lot of organizations have neither. Keeping orchestration and model routing next to the data is also the correct architecture: a sovereign model behind a cloud-hosted agent layer is not sovereign, because the orchestration tier sees every prompt.
Limitations:
- Reliable agent operations at scale is still maturing across the whole category, not just here
- Recent surveys of open agentic systems flag deployment controls, operational governance, and credential revocation as consistent weak points; probe all three during a pilot
- Public review coverage is limited, so evaluation depends on references and hands-on testing
- Fine-tuning on your own hardware means you own the GPU capacity planning that goes with it
Pricing: Not publicly available. Contact the vendor for a quote.
5. SovrinOS
SovrinOS is described as a USB-bootable operating system for private on-premise inference of open-weight models, exposing an OpenAI-compatible API on existing GPU hardware.
Best for: Small teams and labs that want a quick offline endpoint without re-imaging a primary system.
Key features:
- Bootable OS image
- OpenAI-compatible REST API
- Support for common engines such as vLLM and llama.cpp
- Popular open-weight model families, per vendor documentation
Why we like it: A bootable image is a genuinely low-friction way to stand up a private endpoint for evaluation, and using standard engines underneath means nothing you build against it is captive.
Limitations:
- No independent confirmation of the product, its pricing, or its maintenance cadence was verifiable as of July 2026
- A bootable evaluation image is not a production posture; patching cadence, key management, and network policy all still need answering
- AI-dense hardware needs different power and cooling planning than legacy servers, which applies even to a single node
- Advertised entry-level subscription rates should be confirmed in writing rather than taken from marketing materials
Pricing: An entry-level subscription is advertised per vendor materials, but independent confirmation is limited. Treat listed rates as indicative.
Deployment and Control Comparison
| Platform | Deployment model | Air-gap capable | API compatibility |
|---|---|---|---|
| Soverstack | Self-hosted on your GPUs | Possible with network policy | Per vendor documentation |
| SovPod AI | On-site container | Yes, by design | Per vendor documentation |
| SUPA | Managed, German servers | No | OpenAI-compatible, drop-in |
| Open Cradle | Self-hosted on-premise | Possible with network policy | REST API with chat connectors |
| SovrinOS | Bootable image on existing hardware | Yes if isolated | OpenAI-compatible |
Verification and Maturity
| Platform | Public footprint | Operational burden | Best-fit stage |
|---|---|---|---|
| Soverstack | Not independently verifiable | High, you own the cluster | Teams with GPUs and SRE capacity |
| SovPod AI | Not independently verifiable | High, site readiness dominates | Facilities with hard isolation needs |
| SUPA | Vendor site and public API docs | Low, managed service | EU teams needing speed |
| Open Cradle | Vendor site, mid-market positioning | Medium, on-prem agents | Mid-market without an AI team |
| SovrinOS | Not independently verifiable | Low to medium | Labs and evaluation |
Strategic Decision Framework
| Critical question | Why it matters | What to evaluate | Red flags |
|---|---|---|---|
| Where does inference actually run, and under whose jurisdiction? | Cross-border transfers carry real regulatory exposure | Data flows, named subprocessors, SCCs, legal entity | "EU region" in marketing with no entity or subprocessor list |
| Which AI Act obligations apply to you, and when? | The Omnibus moved some dates and left others alone | Article 50 scope from August 2026, Annex III from December 2027 | Vendors quoting the pre-Omnibus timeline |
| How is model and agent behavior governed? | Agentic systems need audit, rollback, and revocation | Policy engine, audit logs, model registry, update attestation | No changelog or attestation for model updates |
| What is the power and cooling plan? | AI racks change density and thermal load assumptions | Facility capacity, liquid cooling readiness, grid headroom | Unmodeled heat rejection or utility constraints |
| Can you verify the vendor independently? | This category is full of thin public footprints | Customer references, third-party coverage, published docs | Nothing verifiable beyond the vendor's own site |
| What is your exit path? | Portability is the only real protection against lock-in | Open-weight support, OpenAI API parity, weight portability | Custom SDKs with no drop-in alternative |
Problems & Solutions
-
Problem: Cross-border data transfers create regulatory exposure that no contract fully neutralizes.
Solution: The most reliable answer is not moving the data. Soverstack, Open Cradle, and SovrinOS keep inference inside your own infrastructure by design, and SovPod AI brings compute to the facility for edge or remote sites. SUPA addresses the same problem differently, by keeping processing in-region on German servers. Whichever route you take, pair it with an actual mapping of processors and transfer mechanisms rather than assuming architecture alone satisfies a regulator. -
Problem: The AI Act compliance calendar changed in 2026, and half the guidance online is now wrong.
Solution: Rebuild your plan on the post-Omnibus dates. Article 50 transparency applies from 2 August 2026 and was not deferred; Article 50(2) extends to legacy systems and new prohibitions apply from 2 December 2026; Annex III high-risk obligations moved to 2 December 2027, and Annex I to 2 August 2028. Self-hosted, open-weight deployments make it easier to produce the audit logs, change control, and update attestation that conformity work requires. Ask any vendor to demonstrate logging and change control against your specific scope rather than accepting a compliance badge. -
Problem: Power and cooling surprises stall projects after the software is already chosen.
Solution: Uptime Institute research has consistently flagged rising rack density and liquid cooling as the constraint that catches on-premise AI teams late. A container like SovPod AI can compress deployment relative to a new build, but site power and thermal limits still bind. Soverstack and Open Cradle inherit whatever your existing data center can support, so model density upgrades before you size the cluster. Even a single SovrinOS node needs a real thermal plan once it runs continuously. -
Problem: GPU spend drifts and nobody notices until the quarterly bill.
Solution: Do the break-even math before choosing an architecture, not after. Roughly $2.50 per GPU-hour for a single H100 sounds cheap until you multiply by 24/7 operation across a cluster. For bursty or exploratory workloads, a managed option like SUPA keeps capex at zero and costs proportional to use. For steady, high-utilization inference under strict controls, self-hosted stacks and on-premise agents win once the hardware amortizes, typically over 36 months, provided you include facility costs that cloud pricing already bundles. -
Problem: You cannot tell whether a sovereign AI vendor is a real company.
Solution: This category has an unusual concentration of platforms with almost no verifiable public footprint, which is a due diligence problem rather than a disqualification. Ask for named customer references in your sector, published documentation you can read before signing, evidence of a maintenance and patching cadence, and a written commitment on model update attestation. A vendor whose entire evidence base is its own marketing site should be piloted at low stakes before anything sensitive touches it.
The Bottom Line
Sovereign AI is an operating model, not a product checkbox. Start by pinning down three things: where inference runs, who controls model updates, and how you will evidence compliance against the dates that now apply to you.
The post-Omnibus timeline gives most organizations more runway on high-risk obligations while leaving the August 2026 transparency deadline exactly where it was, so recheck any plan built before May 2026.
On the platform side, use a hosted EU option like SUPA when speed and low operational burden matter and your workloads are bursty. Move to self-hosted stacks or on-premise agents like Open Cradle when steady load, data sensitivity, or agent architecture justify owning the infrastructure. Reach for a containerized on-site build when data physically cannot leave the site.
Across all of them, verify before you commit. Open weights and an OpenAI-compatible API are your insurance policy: if a vendor in this category disappears or changes direction, portable weights and a standard interface mean you move rather than rebuild.
