Aleph Alpha after the Cohere deal: what you are actually buying now

Reviewing Aleph Alpha in mid-2026 means reviewing two things: a product, which is largely unchanged, and a purchasing argument, which is not.

Roughly 90 percent of the merged company belongs to Cohere shareholders. Aleph Alpha shareholders hold about ten.

Merger announced 24 April 2026, valuing the combination at around $20 billion

This review covers what that does and does not change for a company using it or considering it.

What the product is

Aleph Alpha stopped competing on frontier models before the deal and repositioned around PhariaAI, a platform for sovereign generative AI built on its own tokenizer-free architecture. The focus is enterprise and public sector: regulated environments, German-jurisdiction hosting, domain-specific deployments.

That repositioning was the right call and it is worth saying plainly, training frontier models against the spend of the US labs was not a race a European company was going to win, and pivoting to the layer above is a more defensible position than continuing.

The technology is real, the Heidelberg operation continues, and the compliance posture for German regulated sectors remains one of the strongest available.

What changed

The ownership. That is the entire list, and it is not a small entry.

The deal was announced as a merger and is structurally an acquisition: a stock-for-stock combination with a fresh primary round, roughly 90/10 in Cohere’s favour. The Schwarz Group, previously a central backer of Aleph Alpha, is putting $600 million into Cohere’s Series E. The announcement took place in Berlin with the German digital minister and the Canadian AI minister present, which tells you the deal is as geopolitical as it is commercial.

For a customer, three consequences follow, in ascending order of importance.

Continuity is not in question. Products under active enterprise contracts do not disappear because a cap table changed. Plan for continuity and read your contract’s change-of-control clause once, which you should have done anyway.

Roadmap overlap is a real question. Cohere serves the same regulated enterprise segment. Two overlapping platforms under one owner converge eventually, and the question worth asking your account team is which parts of PhariaAI survive the convergence and on what timeline. Ask now, while the answer is still being decided.

The internal justification needs updating. This is the consequential one. If you bought Aleph Alpha and told your board, your works council or your customers that you chose a German provider, that sentence is now inaccurate. Better to correct it yourself than to have it corrected in a tender or an audit eighteen months from now.

What this says about buying on the sovereignty argument

“The vendor is European” is the least durable of the properties people buy under the heading of sovereignty. It can change in a week, without your involvement, and in April it did.

The properties that survive are the ones written into your architecture rather than into someone’s cap table:

Processing location, contractually. Where the model actually runs, in the DPA, per endpoint, and not confused with storage location.

Retention and training exclusion, contractually. Separate promises from residency, separate clauses.

Substitutability, architecturally. A model behind your own thin interface, with prompts and an evaluation harness you own, so a provider change is a configuration change. That is the only sovereignty property entirely under your control, and it is the one the April deal did not touch for anyone who had it.

We went through the whole picture in European AI sovereignty after the Cohere dealDE (auf Deutsch), including the alternatives.

Who it still fits

Fits: regulated German organisations (public sector, finance, defence, energy, healthcare) with a hard requirement for German-jurisdiction hosting and a procurement process that can evaluate a platform rather than an API. The capability for that profile is genuinely strong and the combined company will have more resources behind it, not fewer.

Does not fit: a mid-sized company looking for a general-purpose model. The platform is built for a scale and a procurement style that a 120-person firm does not have, and the effort of the evaluation exceeds the benefit. Mistral is the more practical European answer at that size, with open weights as a fallback.

And does not fit as a symbolic purchase. If the reason on the slide is “German AI”, the reason has expired, and buying it anyway means defending a position that is no longer true.

Frequently asked questions