Cohere absorbs Aleph Alpha: what sovereign AI means for EU SMEs

A Berlin announcement, a Canadian deal

On the morning of 24 April 2026, in a hangar-style venue close to the Brandenburg Gate, two CEOs walked on stage together: Aidan Gomez of Cohere (Toronto) and Jonas Andrulis of Aleph Alpha (Heidelberg). Behind them stood the Canadian and German digital ministers. The framing of the joint statement was explicit: this was, in the speakers’ words, “sovereign AI for Europe”, “an option between the hyperscalers and the hegemon”, “built on the bedrock of shared Canadian and German values”.

The deal numbers, as reported by TechCrunch on 25 April and confirmed by CNBC, PitchBook and Futurum: a combined valuation of approximately $20 billion; a Series E round led by Schwarz Group with €500m (about $600m) of structured financing; intellectual property remaining in Canada; Cohere staying majority Canadian-controlled; Gomez continuing as CEO of the combined entity.

A Canadian firm absorbing the second-largest European LLM player at a $20bn valuation, with German federal endorsement and the implicit backing of an EU-resident retailer-conglomerate. The question this briefing tries to answer, calmly: what does the word “sovereign” do in that sentence, and what should an EU SME picking an LLM in mid-2026 actually require?

The strategic logic on paper

Read in isolation from the rhetoric, the industrial logic of the merger is reasonably clean.

Cohere had built itself around a general-purpose enterprise model line (Command A family), retrieval (Embed, Rerank), and the North agentic platform, selling primarily to Fortune 500 buyers (Oracle, Dell, RBC, SAP, Salesforce, Accenture). It had passed roughly $240m of ARR in 2025 (per Futurum and TechCrunch). It had not, however, built a credible European go-to-market or a small-language-model capability.

Aleph Alpha, by contrast, had pivoted hard from competing on frontier general-purpose models, an effort it could no longer sustain after the gap with US labs widened in 2023-2024, towards the PhariaAI orchestration and deployment layer, plus a tokenizer and small-LM specialism tuned for German and other European languages. It had public-sector and DAX-level contracts (Bundeswehr, the federal government, Bavaria, Siemens, BMW), a 250-person team, and Schwarz Group’s STACKIT sovereign cloud as a privileged distribution channel. Its revenue base was, by industry standards, minimal.

Aidan Gomez’s quote to TechCrunch captures the complementarity precisely: “Their focus on small language models, European languages and tokenizers is a really complementary one to our own, which is more of a general focus on large language models.”

The merger therefore solves real problems for each side. Cohere acquires European public-sector references, a German anchor customer through STACKIT, and small-LM capability in EU languages. Aleph Alpha acquires the frontier model line it could no longer afford to keep building, a global sales footprint, and a capital base. Schwarz Group locks in an anchor tenant for STACKIT and consolidates its position in the European AI sovereignty market, on top of the €11bn data-centre investment near Berlin that Futurum cites.

It is, on the industrial merits, a reasonable deal. The question is what one calls it.

Sovereignty rhetoric versus ownership reality

The phrase used on stage on 24 April was “sovereign AI for Europe”. It was repeated by both governments in attendance. It is now in the press release boilerplate.

Three observations.

First, the buyer is not European. Cohere is incorporated in Ontario, headquartered in Toronto, and remains majority Canadian-controlled post-deal, with intellectual property held in Canada. Canada’s Minister of AI Evan Solomon, quoted by BetaKit, called it “a big moment for Canadian AI” and emphasised that Cohere is “Canadian-founded, Canadian-headquartered, and built on Canadian research excellence”. Both framings, “Canadian AI moment” in Ottawa and “sovereign AI for Europe” in Berlin, were issued the same day, by aligned governments, about the same transaction. They cannot both be the primary fact.

Second, the German federal posture is unusually direct. The German digital ministry described the deal, per Futurum’s reporting, as having “high geostrategic and economic value”, with Berlin “plan[ning] sovereign AI procurement prioritisation”. Read literally, this is industrial policy: Germany is not claiming the buyer is German, it is claiming the buyer is non-hyperscaler, non-American, and operationally embedded in German infrastructure (STACKIT, Schwarz’s €11bn data-centre footprint). That is a coherent position, but it is a redefinition of “sovereignty” from ownership to deployment.

Third, the financial structure makes the ambiguity load-bearing. A Series E led by Schwarz Group with €500m of structured financing is not a controlling stake. It is enough to anchor STACKIT distribution and to give the Schwarz Group a strategic seat at the table. It is not enough to make the resulting company European in the sense an EU procurement officer might assume.

The honest summary: the deal is European in customer base, European in infrastructure footprint, European in regulatory exposure, and non-European in ownership and corporate control. Calling that simply “sovereign AI for Europe” is a political choice, not a factual description.

The map of EU LLM providers in May 2026

The Cohere-Aleph Alpha acquisition removes one of the two pillars of the EU-owned LLM landscape. The picture that remains, as of mid-May 2026, is sparser than the policy rhetoric would suggest.

Provider	HQ	Ownership status May 2026	Notes
Mistral AI	Paris (FR)	Privately held, mostly European cap table	Series C €1.7bn raised September 2025 at €11.7bn valuation. Paris-area data centre at Bruyères-le-Châtel (44 MW, 13,000+ GPUs) announced 2025.
Aleph Alpha	Heidelberg (DE)	Acquired by Cohere (CA), closing 2026	Part of $20bn combined entity. IP held in Canada. STACKIT remains the anchor distribution channel.
Cohere	Toronto (CA)	Independent, post-acquisition	Canadian-controlled. European base in Germany. Not an EU-owned vendor.
Silo AI	Helsinki (FI)	AMD subsidiary since 12 August 2024	Acquired by AMD (US) for $665m all-cash. Poro and Viking model families continue. Not an EU-owned vendor.
LightOn	Paris (FR)	Privately held, European	Enterprise-focused, smaller scale. Part of OpenEuroLLM consortium.
Ellamind	Germany	Privately held, European	Smaller scale. Part of OpenEuroLLM consortium.
OpenEuroLLM consortium	Multi-country	Public-research-led	First model releases due mid-2026, final by 2028. Mistral not part of the consortium.

Source: company filings and announcements, AMD press releases (10 July, 12 August 2024), Mistral AI press releases (September 2025), TechCrunch and PitchBook reporting on the Cohere-Aleph Alpha deal (April 2026), European Commission AI Factories pages · Last verified 2026-05-19

Two observations on this table. The first is that, after April 2026, the only EU-owned general-purpose LLM provider at meaningful commercial scale is Mistral. The OpenEuroLLM consortium is a substantive public-research effort, but its first model releases are still months away and its enterprise distribution is unproven. The second is that this is not a unique European weakness, it is a function of the capital intensity of frontier model training. The Cohere-Aleph Alpha consolidation, as Futurum notes, “leaves unresolved the fundamental question of capital intensity” for frontier training. The same question applies to Mistral.

Three definitions of sovereignty, kept separate

If “sovereign AI” is to be useful as a procurement criterion rather than a slogan, it has to be unbundled. There are at least three distinct, partially independent attributes hiding inside that one phrase.

Sovereignty 1: where inference runs

This is the easiest to verify and the cheapest to obtain. It asks: in which physical region, in which legal jurisdiction, on which legal entity’s books, does the inference traffic actually run? Microsoft’s EU Data Boundary, Azure OpenAI’s eu-west regions, AWS Bedrock’s Frankfurt and Paris regions, Mistral’s Paris infrastructure, STACKIT’s German data centres, all qualify, in the narrow sense, as “EU inference”. A Canadian-owned vendor running on STACKIT delivers EU-hosted inference. So does a US-owned hyperscaler running on EU regions. The fact that “EU inference” is achievable from a non-EU vendor is what lets the marketing phrase “sovereign AI” stretch as far as it does.

Sovereignty 2: who owns the vendor

This is the criterion that the Cohere-Aleph Alpha deal does not satisfy. It asks: which legal entity, controlled by whom, sells the service, holds the contract, owns the intellectual property, and is subject to which jurisdiction’s extra-territorial laws (CLOUD Act for US owners, comparable Canadian instruments for Canadian owners, GDPR plus AI Act for EU owners)? Under this definition the post-deal Cohere-Aleph Alpha is non-EU, and Silo AI under AMD is non-EU. Mistral is, today, the principal EU answer at scale.

Sovereignty 3: who can legally use the weights

This is the strongest form of sovereignty, and the one most often quietly abandoned in vendor demos. It asks: can our organisation, today, take the model weights, run them on infrastructure we control, fine-tune them, and continue operating if the vendor disappears tomorrow? Mistral’s open-weight releases (Mistral Large 2, the Ministral family, Codestral, Mistral Large 3) satisfy this in part. Aleph Alpha’s Pharia-1 open-weight release in late 2024 satisfied this in part. Cohere’s Command A family does not. The OpenEuroLLM consortium models, when released, are designed to satisfy this fully.

These three criteria are independent. A buyer can have any combination: EU inference and non-EU vendor (most hyperscalers), EU vendor and closed weights (some Mistral enterprise tiers), open weights and US vendor (Meta’s Llama family), EU inference and open weights and non-EU vendor (Llama on Azure EU regions). Every credible EU procurement decision in 2026 must specify which of the three matter, in which order, and to what threshold.

What an EU SME should actually require

For SMEs, who do not have the scale to run a private STACKIT-style infrastructure, the sovereignty conversation has to be pulled down to a small number of contractual questions. The framework we suggest at this desk is five lines on the procurement form.

Line 1: legal entity and jurisdiction. Which legal entity signs the contract, where is it incorporated, and which third-country laws can compel disclosure (CLOUD Act, Canadian production orders, mainland Chinese national-security laws)? This is a factual question with a factual answer.

Line 2: data residency and processing region. Is the inference, the training, the logging, and any human review performed in a named EU region with a contractually defined boundary? Microsoft’s EU Data Boundary, AWS Bedrock EU regions, Mistral’s France-resident inference and STACKIT all qualify here, with different boundary scopes. Read the boundary definition. Especially the human-review carve-out.

Line 3: model weight licensing. Are the weights available to us under a sufficiently permissive licence (Apache 2.0, MIT, Mistral’s research and commercial licences for Mistral Large 2, Aleph Alpha’s Pharia-1 sovereignty licence) for us to keep running inference on our own or third-party infrastructure if the vendor goes away? Or are we entirely dependent on the vendor’s API? This is the single most undervalued criterion.

Line 4: AI Act qualification. Is the provider explicit, in writing, about its status under the AI Act (general-purpose AI model with or without systemic risk, downstream provider obligations, transparency under Article 50)? A vendor that cannot answer this in May 2026 is not credible. The framework collapse we covered in Why most EU AI strategy frameworks did not survive 2026 is partly about exactly this gap.

Line 5: continuity. What happens to our deployment if the vendor is acquired, goes insolvent, or is forced out of the EU market by regulation? Aleph Alpha’s customers in early 2024 lived through a near-death funding scare. Their successor entity is now Canadian-owned. The continuity question is not theoretical.

These five lines do not contain the word “sovereign”. That is the point. The word does too much work and resolves too little.

A post-mortem framework: what failed, what worked, what is uncertain

This deal is not the failure of EU AI policy. It is, at most, a clarifying event. A short framework, in the spirit of the post-mortem we ran on consulting frameworks.

What failed.

• The implicit 2022-2024 European bet that two or three EU-owned generalist LLM champions would emerge to balance the US labs. By May 2026, Aleph Alpha is folded into a Canadian acquirer, Silo AI is folded into AMD, and Mistral is the single remaining EU-owned generalist provider at scale. The arithmetic on champions did not work.
• The conflation of “European hosting” with “European vendor” in EU industrial communication. The Cohere-Aleph Alpha announcement weaponises that conflation. It is not a hostile move, it is a rational corporate move that uses an ambiguity governments have been content to leave in place.
• The framework-led approach to EU AI strategy (see post-mortem), which produced maturity ladders and value-creation wheels but very few procurement-grade contractual clauses an SME could lift.

What succeeded.

• The infrastructure layer. Schwarz Group’s STACKIT, the EuroHPC AI Factories network (19 operational AI Factories and 13 antennas across 15 Member States as of early 2026), and the InvestAI facility’s planned €20bn for up to five AI Gigafactories represent real EU-controlled compute. Even if the model layer above it is partly non-European, the physical layer is increasingly European.
• The open-weight option. Mistral’s strategy of releasing open weights for substantial parts of its line, and the OpenEuroLLM consortium’s planned mid-2026 release of multilingual EU-language models under a sovereignty-friendly licence, give EU SMEs a real third path, separate from “buy from a US hyperscaler” and “buy from an EU-owned vendor at frontier prices”.
• The regulatory layer. The AI Act, despite ongoing criticism, has succeeded in establishing a shared procurement vocabulary that lets EU buyers ask precise questions of any vendor, EU-owned or not.

What remains uncertain.

• Whether Mistral can sustain a generalist frontier line at European cap-table-friendly capital intensity, or whether its eventual choice will resemble Cohere’s, raise strategic capital outside the EU, accept partial loss of European ownership, gain global scale. The €830m debt round in March 2026 to fund the Paris data centre is the early signal here.
• Whether the OpenEuroLLM consortium models, when released, will achieve enough enterprise distribution to matter, or whether they will remain a public-research artefact.
• Whether the next consolidation event will go further. A buyer for a hypothetical Mistral exit in 2027 would be the test case, and there is no domestic European acquirer of credible scale.

The Cohere Aleph Alpha acquisition is therefore best read as a stress test of EU AI sovereignty rhetoric, not as a refutation of it. The fix is not to retire the term but to unbundle it, into the three definitions above, and to demand that procurement, policy and press releases keep them separate.

That is the dull, useful version of “sovereign AI Europe 2026”. It does not fit on a stage in Berlin. It does fit on five lines of an SME procurement form.

---

Marcus Heller covers Market and DACH for flintbrief. Briefings are written from public sources, dated, and republished when material facts change. This piece will be revisited when the Cohere-Aleph Alpha closing documents are filed and when Mistral’s next funding event is announced.