Buying Sovereignty: What European AI Actually Delivers in September 2026

On 8 September 2026, Mistral raised three billion euros. Its own announcement states the purpose unusually plainly: to push sovereign, open-weight AI to the technology frontier. For a debate that has been carried in Germany for years with the line that Europe has nothing to offer anyway, that is the hardest counter-figure in a long time.

More interesting than the sum, though, is something that happened four weeks earlier and made barely any headlines. On 11 August, the processing region became a product you book — rather than an architecture you build. That is precisely what shifts the question I hear most often in advisory conversations.

At a glance

• What applies: EU/US data residency has been generally available at Mistral since 11.08.2026, and three billion euros in funding followed on 08.09.2026. Apertus has been available since 24.07.2026 with open weights, open data and open training code.
• What follows: The data region is no longer an architectural decision but a configuration parameter — in the Claude API, settable per session as inference_geo since 07.08.2026.
• The catch: Europe’s large data-centre projects were only put out to tender on 30.07.2026 and will not run before mid-2028. You procure sovereignty from providers, not from programmes.

What has become bookable since the summer

Three movements point the same way, each pinned to a date. The overview separates by maturity — between “generally available” and “preview” lies a great deal in a procurement decision:

Date	What became available	Maturity	What it does not solve
24.07.2026	Apertus 1.5 (8B/70B): open weights, data, training code; 262,144 tokens	openly available	operations stay with you
07.08.2026	inference_geo in the Claude API — region overridable per session	available	the provider remains a US company
11.08.2026	Regional Endpoints for EU/US data residency	generally available	service quality is separate
11.08.2026	Priority Tier with service-level commitments	public preview	no production SLA
08.09.2026	Funding round of EUR 3 billion	closed	capital is not a product

One point from 11 August went unnoticed: Mistral now also runs open third-party models on its own platform, starting with Zhipu AI’s GLM-5.2 — which separates the provider you contract with from the model that computes. Apertus, in turn, delivers something no one else delivers: alongside the open weights come training data and training code — the difference between “we can operate the model” and “we can explain what it came from”. It is carried by ETH Zurich, EPFL and the CSCS.

The advisory question has shifted

I wrote my guide comparing EU-hosted and US models in December 2025; it describes sovereignty as a ladder of four rungs that you climb and then stay on — that ladder has turned into a switch in the middle.

The decision is therefore no longer “on premise or cloud”, once for the whole company, but: which processing in which region, under which commitment, verifiable against what. The third question is routinely forgotten, because a setting is not a control — a region parameter says nothing about which law the provider is subject to. In practice: the value you set belongs in the record of processing activities, the commitment in the contract, and the proof that the two match in your review routine.

Three objections that belong here

First: the large EU projects will not help you this financial year. On 30 July 2026, the Commission opened the call for up to seven “AI Gigafactories” with at least 100,000 AI chips each — roughly 30 billion euros in total investment. Submission deadline 12 November 2026, awards early 2027, operation from mid-2028. That is compute capacity for future European foundation models, not an option for a decision you have to take in the fourth quarter.

Second: announced is not available. In early July 2026, Mistral’s leadership announced an open frontier model under the Apache 2.0 licence, with a broader release promised “this summer”. The changelogs published since contain no general availability under any recognisable name. No reproach, but a procurement rule: only what you can obtain today belongs in a selection decision.

Third: the fine print decides. Data residency is generally available, while the service-level commitments sit next to it as a public preview — anyone who merges the two is selling a preview as a commitment. And with open weights, the hardware decides: Alibaba’s Qwen3.8, opened up on 12 August 2026, has 2.4 trillion parameters — open does not mean operable, certainly not in a mid-sized company’s server room. The realistic sizes are those of Apertus, and you check the licence beforehand.

What stands today, stands: a European provider with a bookable data region and an open model. The third shape is European compute capacity — tendered and funded, an outline until 2028.

What this means for your company

Sort by data class, not by provider. If the region is a parameter, the choice of provider is not the first question. The first is which data flows into which use case — from public content through to trade secrets and special categories. Only then do you assign region and contract.

Treat open models as a second pillar, not an ideology. The lesson of the summer is not “open is better” but that an open model is the exit option that makes switching providers negotiable. Anyone who has settled the open versus proprietary trade-off can now implement it more cheaply.

Put the ability to switch into the contract, not into your intentions. Separating the contracting party from the computing model is the lever at the heart of vendor lock-in — usable only once export formats, notice periods and data return are settled.

Conclusion

Europe’s position is better than the discourse describes it — and worse than the headlines suggest. Better, because a provider with three billion in capital, bookable EU data residency and open weights is no longer a promise but a supplier. Worse, because the infrastructure projects have not even been awarded yet and the actual work still sits with you: classify data, check commitments, prepare the exit. Sovereignty has thus become a procurement matter, not an architecture project — cheaper, but less comfortable, because contractual diligence is hard to put on display.

If you want to know which region holds up for which of your use cases, and how to verify the commitment, let’s talk. I read contracts like these as a business lawyer and build the systems in question myself.

FAQ

Can I simply buy data sovereignty in September 2026?

For many use cases, yes: EU data residency has been generally available at Mistral since 11 August 2026. What you buy is a commitment about where processing happens — not immunity from foreign law. Who the provider is, and which law it is subject to, remains the prior question.

What is inference_geo, and why does it change the advisory question?

A parameter in the Claude API that, since 7 August 2026, lets you set per session where a model computes. The processing region is therefore no longer an architectural decision but a configuration per use case. The question is no longer “cloud or self-hosted” but: which region, under which commitment, verifiable against what.

Is on-premise still worth it when EU regions exist?

Yes, but for fewer cases than before. An EU region lowers the transfer risk; self-hosting remains the instrument of choice when data must not leave the building. Apertus 1.5 has been available since 24 July 2026 with open weights, open data and open training code in 8 and 70 billion parameters — hardware and maintenance stay your job.

When will the European AI Gigafactories arrive, and does that help me?

Not soon enough to wait for. The Commission opened the call for up to seven sites on 30 July 2026; the submission deadline is 12 November 2026, awards are planned for early 2027 and operation from mid-2028. That is compute capacity for future European models, not a procurement route for this financial year.

---

Sources — as of 12/09/2026

• Mistral, news (funding round 08.09.2026, Regional Endpoints 11.08.2026) — https://mistral.ai/news
• Mistral, changelog (Regional Endpoints, Priority Tier) — https://docs.mistral.ai/resources/changelogs
• Apertus 1.5, Swiss AI Initiative (ETH Zurich, EPFL, CSCS) — https://www.apertus-ai.org/articles/2026-07-apertus-1-5/
• European Commission on the AI Gigafactories — https://digital-strategy.ec.europa.eu/en/news/eu-launches-ai-gigafactories-call-boost-europes-computing-capacity-and-unlock-more-eu30-billion
• Claude API changelog (inference_geo, 07.08.2026) — https://platform.claude.com/docs/en/release-notes/api.md
• Techtimes on the announced open frontier model (06.07.2026) — https://www.techtimes.com/articles/319798/20260706/mistral-ai-targets-frontier-gap-open-weight-model-entering-july-early-access.htm
• Alibaba Cloud Model Studio, Qwen3.8-2.4T-A95B (12.08.2026) — https://www.alibabacloud.com/help/en/model-studio/newly-released-models

This article is general information and not legal advice in an individual case. As of 12 September 2026; offerings and maturity levels move quickly here — please check the current state before procurement decisions.