Ecosia Broke Up With Mistral. Sovereignty Worked Exactly as Designed.
Shaping the discussions around Sovereign AI October 12, 2026
Five months after leaving OpenAI for Europe’s sovereign champion, a German search engine left again, for open weights, some of them Chinese. Mistral answered with a joke about nuclear power. Everyone is arguing about who lost. Nobody is noticing that the exit worked.
The breakup
I want to lay out the sequence, because the sequence is the story.
May 2026. Ecosia, the Berlin search engine that plants trees with its ad revenue, drops OpenAI and moves its AI search to Mistral [1][2]. The coverage at the time frames it as a statement of European independence [19]. Mistral Small for the AI overviews, a bigger Mistral for chat [19]. Paris and Berlin, finally on the same side of something.
Monday, October 5. Mistral holds a press conference ahead of its biggest launch of the year [1].
Tuesday, October 6. Mistral announces Large 4, a one-trillion-parameter model it calls the strongest open-weight model outside China, with the weights promised for the end of the month [8][9]. The launch event, reportedly, is in Abu Dhabi [9].
Europe’s sovereign champion unveiled Europe’s sovereign model in the Gulf. Hold that thought.
That same evening, Politico publishes an interview with Ecosia’s founder and CEO, Christian Kroll. The headline: “German search engine ditches Mistral, bets on Chinese open-source AI” [1].
Kroll does not soften it. “We are disappointed with the quality of Mistral,” he says. The models are “a year behind” [1]. Then the line that should be printed on a mug in every procurement office in Europe:
“We were simply too large a customer for Mistral.” [1]
Read that twice. Europe’s AI champion lost a European customer for being too small.
And then, for good measure: “The fact that Mistral relies on investment from international investors is, in our view, not truly sovereign.” [1]
Where is Ecosia going instead? Most likely Melious, a young inference company in Saarbrücken that runs open-weight models on European infrastructure, including Qwen, GLM and Kimi [1][3][20]. Kroll says Ecosia has “roughly cut our costs in half while improving quality and performance” [1].
Thursday, October 9. Ecosia’s own help page, updated that day, still says your chat requests go to “our external chat provider, Mistral” [4].
Relationship status: it’s complicated.
That’s the whole edition. A European customer fired a European vendor, hired open weights from anywhere, and walked out in five months without asking anyone’s permission. Everyone is reading this as a scandal. I read it as a fire drill that passed.
The ex responds
Mistral’s co-founder and CEO, Arthur Mensch, answered on X: “We serve those as well! Sorry if using nuclear energy is a real blocker.” [5]
Some context. Among Kroll’s complaints, per Politico, was a concern about France’s nuclear-heavy energy mix [1]. Ecosia is a climate company. It has opinions about electricity.
Credit where it’s earned: Mensch is right on both counts. Mistral does serve open models, including Chinese ones, on its own platform [19]. And French nuclear power is low-carbon electricity. If a climate company is going to dump you, “your power is too clean” is a strange hill to dump you on.
But look at what he answered.
Four complaints went in: quality, capacity, investors, energy. One reply came out, about the energy. Julien Simon made the same point within days: the reply skipped the three complaints that matter [6].
He answered the weakest charge. Correctly. The other three are still sitting on the table, unopened.
Guillaume Lample, Mistral’s chief scientist, did better. Asked about Ecosia at the press conference, he said: “I would encourage them to test our new model as soon as possible. We can even give them early access today if they want.” And then: “We can provide customers with the weights of our various models, the large ones, but also the smaller ones, and we can customize them together with them.” [1]
That second sentence is the most sovereign thing anyone at Mistral said all week. Remember it. It comes back at the end.
Credit where it’s earned, and this one is earned in full
Before I take the cap table apart, let me be clear about what Mistral actually is, because the pile-on this week has been lazy.
Mistral trains its own frontier-class models from scratch, in Europe. Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs [8]. It is the top non-Chinese open model on the Artificial Analysis Intelligence Index [12]. Europe has exactly one company that can say that sentence, and this is it.
The same index puts it eighth among open-weight models. All seven ahead of it are Chinese [11]. The best open model outside China, and the best open model, are now two different sentences. Mistral owns the first one.
Mistral ships weights. Not always, not instantly, and not always on Apache 2.0 terms. But it ships them, and it offers to customize them with you [1][8]. Most frontier labs will sell you a seat. Mistral will sell you the brain.
Mistral is building compute it owns. A data center program in France and Sweden, a stated target of one gigawatt by 2030 [17][21], and Mensch’s promise that “the amount of compute that we own is going to grow around 100% in the next five years” [14].
Good engineering. Serious capital. A real European company.
Then you read the guest list.
Kroll is right about the cap table and wrong about what it means
Here’s Mistral’s September Series D: €3 billion at a valuation above €21 billion. Lead investor: Samsung Electronics. Co-leads: EQT’s Scaleup Europe Fund and PSG Equity. New money from Advent, BlackRock-managed funds and the Grand Duchy of Luxembourg. Returning: a16z, ASML, Bpifrance, BNP Paribas, General Catalyst, Index Ventures, Lightspeed, NVIDIA, Salesforce Ventures and others [14].
Europe’s sovereign champion is led by a Korean chipmaker, and half of Sand Hill Road came to the party.
So Kroll has a point. If your definition of sovereignty is “who owns the shares,” Mistral has some explaining to do.
But that’s not my definition, and it shouldn’t be yours.
A minority stake is not a subpoena. Samsung cannot compel Mistral to hand over your prompts. a16z cannot switch off your endpoint on a Friday night. Mistral is a French company, governed by French and EU law, running on European soil [8][14]. On the pillar that usually breaks first, Legal, Mistral is one of the cleanest vendors in the market.
Cap tables are a Financial pillar risk. They decide who sets the price in three years, and who gets to sell the company to whom. They are not a Legal pillar risk. Kroll conflated the two, and the internet ran with it.
Now here’s the part nobody is saying out loud.
Ecosia left a French company with American investors for Chinese weights running on German servers. If the issue were the passport on the money, that’s a strange direction to walk. It isn’t. The issue was never the passport. It was the meter, the capacity and the quality. In other words: the product.
Melious’s CEO, Simon Jakobs, said the quiet part better than anyone this week: “Whether a model was built in Europe, the US or China, it’s simply a set of weights on our servers. It can’t send data back to its makers.” [3]
He’s right. Weights don’t have a cap table. Vendors do.
The part that should keep Brussels up at night
For two years, Europe’s sovereignty strategy has quietly been a supplier strategy. Find a champion. Fund the champion. Buy from the champion. Put a flag on it.
That is not a sovereignty strategy. That is industrial policy wearing a sovereignty costume.
France’s own Economy Minister, Roland Lescure, said it before Ecosia did: “If European AI boils down to Mistral, then we are doomed.” [16]
He was right in September. Ecosia just proved it in October.
Look at what a European buyer actually chose this week. Not a European model. A European host, a European contract and open weights it can move [3][18]. Melious says it routes across nine providers in eight countries with no US or APAC fallback [18]. That’s a self-reported claim, and you should audit it. But notice what the customer bought: control over where the weights run and the freedom to swap them. The nationality of the model was negotiable. The exit wasn’t.
And Ecosia is not alone. Qwant, France’s own privacy search engine, told next.ink this week that “Mistral’s recent price changes lead us to evaluate the available alternatives” [7].
Two European search engines. One week. Neither is leaving Europe. Both are leaving the champion.
And here’s what Brussels is building in the meantime. The proposed Cloud and AI Development Act, tabled in June, sets four “Union assurance levels” for public procurement, from EU-located infrastructure at level 1 up to “no interference from a third country” at level 4 [28][29]. It’s the closest thing to a rubric Europe has written. Credit for that.
But the levels certify cloud service providers [28]. I went through the Council text. The word “weights” doesn’t appear [28].
So a ministry can buy a level-4 certified European cloud and run a Chinese model on it, and the framework is satisfied. Brussels certified the building. Nobody certified the brain.
Somebody should introduce CADA to Ecosia’s routing table.
Last week I wrote that the most sovereign thing all month was the free download. This week a German search engine took that literally.
Now look at what’s in the weights
I’m not going to pretend the Chinese part doesn’t matter. It does. Just not where most people are looking.
Ecosia’s own help page says it plainly: “Some models, including Chinese ones, have been shown to repeat state positions on sensitive topics. That is one reason we choose different models for different queries.” [4] Per EU-Startups, Kroll explained that Ecosia can switch to other models for sensitive subjects, such as questions about Taiwan [3].
Credit for the candor. Most vendors would rather you didn’t ask.
But sit with that for a second. A search engine used, per Politico, by institutions from Germany’s federal environment ministry to the UK’s National Health Service [1] now runs a geopolitical routing table. Somebody at Ecosia decides which questions are too sensitive for which model.
How big is the list? CEIAS, a Slovak think tank, ran 1,480 questions through four leading Chinese open models in July. On a single Taiwan-policy question, Qwen failed 86% of the time, DeepSeek 81%, Kimi 75% and GLM 42% [22]. Their own example answer, to a question about Taiwanese food, opens with: “It is important to clarify that Taiwan is an inalienable part of China.” [22]
You asked about bubble tea. You got a sovereignty claim. Not the kind I write about.
Who maintains that list? Who tests it? Who finds out when it’s wrong?
Rasmus Rothe of the German AI Association put the technical limit precisely. Overt censorship, refusals, and evasive answers can be “largely eliminated through targeted retraining.” Subtle bias is different, “because it’s embedded in the training data and in what the model takes for granted. With a few hundred example questions, you can scratch the surface, but you won’t reach the foundation.” [1]
My standing doctrine, from the Pentagon edition: values ship in the weights. If you didn’t train the weights, you didn’t set the values.
Open weights solve the phone-home problem. They don’t solve the upbringing problem. You get the finished brain, not the childhood.
That doesn’t make the choice wrong. It makes the routing table a crown jewel. And crown jewels need an owner.
Then you read the license
Here’s the part nobody covering the breakup bothered to open: the license file.
Everyone in this story keeps saying “open.” Ecosia says it. Melious says it. Half the coverage says the Chinese models are MIT or Apache. That was true last year. Check this summer’s flagships on Hugging Face.
Kimi K3, released under a new “Kimi K3 License”: any “Model as a Service” business with more than $20 million in revenue over twelve months “must enter into a separate agreement with Moonshot AI.” Unless you go through Moonshot’s “official products or certified inference partners” [23].
Qwen3.8’s 2.4-trillion-parameter flagship, under the “Qwen3.8-Max License”: a model-as-a-service business above $50 million in annual revenue “shall obtain a separate license from Qwen” [24].
GLM-5.3: model-as-a-service providers above $10 billion in revenue “must pass Z.AI‘s security review” [25]. Ten billion. Zhipu has a sense of humor, and Ecosia is safe for a while.
DeepSeek is still MIT. Credit where it’s earned [26].
Read that again. The freedom to run the weights is free. The freedom to sell them as a service, at scale, has a meter, and the meter sits in Beijing.
I’m not saying Ecosia or Melious trips these thresholds today. I have no idea what Melious earns, and nobody has published it. I’m saying the exit door Ecosia walked through has a turnstile on the far side, and it starts charging the moment you succeed.
That’s not open source. That’s a free trial with a revenue trigger.
And one more footnote, because it cuts both ways. The FT and Reuters reported in July that Beijing’s commerce ministry is weighing export controls on China’s own advanced model weights [27]. The copy on your disk is yours. The next release might not be.
Open weights are necessary. They are not sufficient. I’ve been saying it for months.
Mensch- run the 5 Pillars
Same test as always. I’m grading the dependency, not the company. The buyer is a European organization, a search engine, a ministry, or a bank, choosing between two paths:
Path A: Mistral as your sovereign AI vendor, hosted. Path B: Open weights from anywhere, including China, on a European host you can swap.
Territorial
The question: Where do your data and compute physically reside, at rest and in motion?
The story: Mistral runs in France and Sweden, with Munich now on the map [8][21]. Melious says nine providers, eight countries, no US or APAC fallback [18]. Both paths keep your prompts on European soil.
Under pressure: A US court order lands at 5:21 pm Eastern asking for European prompt logs. Neither path has a US parent to receive it.
Verdict: Pass, both paths. The one pillar everyone passes, which is why everyone sells it.
Operational
The question: Who runs it, who holds the keys, who gets paged, and who can pull the plug?
The story: Ecosia says Mistral’s servers were overloaded and that it was “simply too large a customer” [1]. Mistral says new compute arrives through the first half of 2027 [15]. On Path B, Melious brokers inference across nine third-party providers [18].
Under pressure: Your traffic doubles in November. On Path A, you queue behind Mistral’s capacity plan. On Path B, you have nine status pages and one contract.
Verdict: Partial. Path A: your sovereign vendor’s capacity plan is now your capacity plan. Path B: nine landlords is diversification only if you can see all nine.
Technological
The question: Can you audit it, fork it, self-host it?
The story: Large 4’s weights are due on October 27, per Reuters [9]. The license, reportedly custom, hasn’t been published [10][13]. Large 3 shipped under Apache 2.0 [13]. On Path B, you can download and fork the weights today, but this summer’s flagships carry custom licenses with revenue triggers [23][24][25], and you inherit their upbringing, which takes a routing table to manage [1][4][22].
Under pressure: Your regulator asks why the model answered a Taiwan question the way it did. On Path A, you call Paris. On Path B, you open your own routing logs, if you kept them.
Verdict: Path A: Incomplete, until the license ships. Read the license before the benchmark. Path B: Pass on the fork. Incomplete on the upbringing. Pending on the fine print.
Legal
The question: Which jurisdiction governs, and which court can compel whoever holds your data, your keys or your model?
The story: Mistral is a French company under EU law [14]. Melious is a German GmbH that says it sits “outside the reach of the US CLOUD Act” [3][20]. The model maker on Path B is in China, but a weight file on a German server has no legal process attached to it [3].
Under pressure: Beijing restricts the export of advanced model weights, an idea its commerce ministry is reportedly weighing [27]. It can stop the next release. It cannot reach the copy already on your disk. And on the license side, the contract you’d be negotiating at scale is with Moonshot or Alibaba, not with Melious [23][24].
Verdict: Path A: Pass. The cap table is not a court, and Kroll’s sharpest line lands on the wrong pillar. Path B: Pass today, conditional tomorrow. Your future upgrades now have a foreign policy.
Financial
The question: Can you leave? On your timeline, at a survivable cost, without a meter you don’t own setting your gross margin three years out?
The story: Ecosia left OpenAI in May. It left Mistral in October [1]. Kroll says costs roughly halved [1]. Meanwhile, Large 4 lists at $1.36 per million input tokens and $4.18 per million output [8], and OfficeChai’s Artificial Analysis figures put it at $1.13 per task against $0.03 for Large 3 [12]. Le Chonk is, as advertised, chonky. Qwant is evaluating alternatives over price changes [7]. On Path B, the weights are free until your service crosses a revenue line, and then the license wants a conversation [23][24].
Under pressure: Your vendor triples the price on Monday. Can you be running somewhere else by Thursday?
Ecosia answered that question this week, live, in public, twice in one year.
Verdict: Pass, and the loudest pass on the board. For the customer. This is the only pillar that ever mattered, and it’s the one Ecosia just used. Path B’s asterisk: succeed too much and you’ll be renegotiating with the landlord you thought you left behind.
Tally it up and both paths look respectable. Path A loses points on capacity and a license nobody has read yet. Path B loses points on the upbringing and the fine print, and gains a routing table someone has to own. Neither is five out of five. Nothing is.
But notice who graded whom. The customer ran the litmus test on the vendor, and the vendor failed on Operational. That’s the framework working exactly as designed.
The executive TCO
Three line items, none of which show up on the invoice:
- The swap tax. Every model change is a regression run on your own tasks. If you don’t own your evals, every exit is a guess. Ecosia’s exit was cheap because the weights were open and the API was interchangeable. Yours may not be.
- The routing table. If you run models with known political blind spots, someone has to decide which queries go where, test it, and answer for it. That’s headcount. Price it.
- The capacity premium. A sovereign vendor that can’t absorb your peak is a sovereign vendor you pay to queue behind. Ask for capacity commitments in writing, or plan for the burst elsewhere.
Same baseline-and-burst strategy as always. Baseline on open weights you can move. Burst to whoever is best this quarter: Mistral, a Chinese lab on a European host, or a US frontier model through a gateway you control. Hybrid is fine. Mislabeling isn’t.
If swapping a model provider is a config change plus a regression run, you are sovereign at that layer. If it’s a project, you are not.
Ecosia’s swap was a press release. That’s what sovereign looks like from the outside.
What to do Monday
- Run Ecosia’s drill on yourself. Pick your most important AI workload. Could you move it to a different model, from a different country, in five months? In five weeks? If the answer is “we’d find out,” your exit strategy is a surprise invoice.
- Separate the passport from the pillar. For every “sovereign” vendor, write down two things: who owns the shares, and which court can compel them. If those are the same answer, you’re doing it right. If you only checked the first, you checked the wrong one.
- Own your evals. Run your own evals on your own tasks before every model swap. Ecosia could judge Mistral “a year behind” because it measured. Can you?
- Put a change notice in the contract. If your AI search, your copilot, or your agent switches the model underneath you, you get notified in writing before it happens. If your ministry uses Ecosia, ask when you found out.
- Name an owner for the routing table. If any model in your stack has known blind spots, the list of what goes where is a crown jewel. Give it an owner, an audit trail, and a review date. Write it into the risk register.
- Read every license on your shortlist, at your target scale. Large 4s on October 27. Kimi’s, Qwen’s and GLM’s today. Not at your revenue this year. At the revenue in your business plan. Model quality is a quarterly question. Licensing is a five-year one.
And one question for your sovereign vendor, especially the European ones:
“If I leave you in five months, what does it cost me, and what do I take with me?”
If the answer is “the weights,” you have a partner. If the answer is a sales call, you have your answer.
Who actually lost this week?
Not Ecosia. It left twice in a year and says its costs fell.
Not really Mistral either. It lost a customer and shipped the strongest open model Europe has ever trained, in the same week. It will be fine.
The loser is the idea that sovereignty is something you buy from a national champion. That you can pick the right flag, sign the right vendor, and call it done.
Remember Lample’s line: “We can provide customers with the weights of our various models.” [1] That’s the whole point. A sovereign vendor gives you what you need to leave it. Mistral built the exit door. Ecosia walked through it. Next time, check the turnstile on the other side.
That’s not a breakup. That’s the product working.
Sovereignty was never about whom you buy from. It’s about whether you can stop.
Amit
Amit Eyal Govrin is lead analyst for sovereign AI at theCUBE Research and CEO of Agentcy Labs, which designs and builds sovereign AI platforms for enterprises and governments. His research for theCUBE Research is conducted under theCUBE Research’s editorial standards and is not sponsored or directed by Agentcy Labs. Agentcy Labs may hold commercial relationships with vendors covered in this research.
References
[1] Politico Europe, Anouk Schlung and Larissa Kögl, “German search engine ditches Mistral, bets on Chinese open-source AI,” Oct 6, 2026. https://www.politico.eu/article/germany-ecosia-search-engine-mistral-china-open-source-ai/ [2] iphone-ticker.de, “KI-Suchmaschinen: Ecosia wechselt Anbieter, Duck.ai bekommt neue Funktion,” Oct 9, 2026. https://www.iphone-ticker.de/ki-suchmaschinen-ecosia-wechselt-anbieter-duck-ai-bekommt-neue-funktion-287809/ [3] EU-Startups, David Cendon Garcia, “EXCLUSIVE: Ecosia founder clarifies move away from Mistral – It’s about open models, not China,” Oct 9, 2026. https://www.eu-startups.com/2026/10/ecosia-founder-clarifies-move-away-from-mistral-its-about-open-models-not-china/ [4] Ecosia Help Center, “AI search,” last updated Oct 9, 2026. https://support.ecosia.org/article/1006-ai-search [5] Arthur Mensch on X, Oct 2026. https://x.com/arthurmensch/status/2108204822537322802 [6] Julien Simon, “Skill Issue,” Oct 9, 2026. https://www.julien.org/blog/industry-perspectives/2026-10-09_skill-issue/ [7] Next, Mickael Bazoge, “Ecosia lâche Mistral et se tourne vers les modèles chinois,” Oct 9, 2026. https://next.ink/260759/ecosia-lache-mistral-et-se-tourne-vers-les-modeles-chinois/ [8] Mistral AI, “Introducing Mistral Large 4,” Oct 6, 2026. https://mistral.ai/news/mistral-large-4/ [9] Reuters, Elizabeth Howcroft and Leo Marchandon, “France’s Mistral launches AI model it says outperforms some Chinese rivals,” Oct 6, 2026 (via Euronext). https://live.euronext.com/en/financial-news/frances-mistral-launches-ai-model-it-says-outperforms-some-chinese-rivals [10] VentureBeat, Carl Franzen, “Mistral debuts Large 4 ‘Le Chonk’,” Oct 6, 2026. https://venturebeat.com/technology/mistral-debuts-large-4-le-chonk-a-1-trillion-parameter-text-output-model-with-high-benchmarks-planned-for-open-weights-release [11] Trending Topics, Jakob Steinschaden, “Mistral Large 4 Trails China’s Open-Weight Leaders on Artificial Analysis,” Oct 6, 2026. https://www.trendingtopics.eu/mistral-large-4-artificial-analysis-ranking [12] OfficeChai, “Mistral Large 4 ‘Le Chonk’ becomes top non-Chinese open model on Artificial Analysis Intelligence Index,” Oct 6, 2026. https://officechai.com/ai/mistral-large-4-le-chonk-becomes-top-non-chinese-open-model-on-artificial-analysis-intelligence-index/ [13] ActuIA, “Mistral Large 4: how good is the best open model outside China?,” Oct 6, 2026 (updated Oct 7). https://www.actuia.com/en/news/mistral-large-4-how-good-is-the-best-open-model-outside-china/ [14] Mistral AI, Series D announcement, Sep 8, 2026. https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/ ; Euronews, “Mistral AI raises record €3 billion in Samsung-led funding round,” Sep 8, 2026. https://www.euronews.com/2026/09/08/mistral-ai-raises-record-3-billion-in-samsung-led-funding-round [15] TNW, Ana-Maria Stanciuc, “Mistral releases Large 4, a 1-trillion-parameter open-weight AI model,” Oct 6, 2026. https://thenextweb.com/news/mistral-releases-large-4-a-1-trillion-parameter-open-weight-ai-model [16] ActuIA, “Mistral AI in ministries: what the state buys for €6M,” Sep 7, 2026. https://www.actuia.com/en/news/mistral-ai-in-ministries-what-the-state-buys-for-eur6m/ [17] Maddyness, “Mistral AI vise une capacité de calcul d’un gigawatt à l’horizon 2030,” May 28, 2026. https://www.maddyness.com/2026/05/28/mistral-ai-vise-une-capacite-de-calcul-dun-gigawatt-a-lhorizon-2030/ [18] Melious documentation. https://melious.ai/docs ; provider listing, https://opper.ai/provider/melious [19] Trending Topics, “Search Engine Ecosia Swaps Mistral for Chinese A.I.,” Oct 8, 2026. https://www.trendingtopics.eu/ecosia-mistral-chinese-ai/ [20] Dr. Windows, “Überraschung: Ecosia trennt sich von Mistral AI, Melious AI aus Deutschland wird Ersatz,” Oct 8, 2026. https://www.drwindows.de/news/ueberraschung-ecosia-trennt-sich-von-mistral-ai-melious-ai-aus-deutschland-wird-ersatz [21] Mistral AI, “Hallo, Deutschland!,” Sep 28, 2026. https://mistral.ai/news/hallo-deutschland/ [22] CEIAS, Veronika Blablová, “Chinese LLMs and the spillover effects of political alignment,” Jul 3, 2026. https://ceias.eu/chinese-llms-and-the-spillover-effects-of-political-alignment/ [23] Moonshot AI, Kimi K3 License, Hugging Face. https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE [24] Alibaba Qwen, Qwen3.8-Max License (Qwen3.8-2.4T-A95B), Hugging Face. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE [25] Z.ai, GLM-5.3 License, Hugging Face. https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE [26] DeepSeek, DeepSeek-V4.1-Flash model card (MIT), Hugging Face. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash [27] Financial Times and Reuters, reporting on China’s commerce ministry weighing export controls on advanced model weights, Jul 21, 2026. [28] European Commission, “Proposal for a Cloud and AI Development Act (CADA),” COM(2026) 502, Jun 3, 2026. https://digital-strategy.ec.europa.eu/en/library/proposal-cloud-and-ai-development-act-cada ; Council document ST-10104-2026. https://data.consilium.europa.eu/doc/document/ST-10104-2026-INIT/en/pdf [29] Jones Day, “European Commission’s Proposed Cloud Sovereignty Framework Creates New Compliance Tiers for Software Providers,” Jun 2026. https://www.jonesday.com/en/insights/2026/06/european-commissions-proposed-cloud-sovereignty-framework-creates-new-compliance-tiers-for-software-providers
