On October 3 — German Reunification Day, because subtlety has never been a German virtue — Aleph Alpha released Kolibri, an open-weight model for German and English under the Apache 2.0 license, weights posted to Hugging Face, trained from scratch on infrastructure in Germany and Finland.

The spec sheet: 78.1 billion total parameters, 3.46 billion active per token, a 20-trillion-token pretraining run, and — the detail that deserved more attention than it got — pretraining finished on September 11 and the model shipped three weeks later. Aleph Alpha’s own benchmark table puts Kolibri at 96.9 on AIME 2025 against 84.6 for Qwen3.6-35B-A3B and 91.7 for Nemotron 3 Super 120B-A12B. Hacker News handed it 543 points and 307 comments inside a day, which is roughly a decade of European AI enthusiasm compressed into one thread.

The Benchmark Table Is Not the Product

Both readings of Kolibri are capability readings, and both are wrong in the same direction.

The skeptical reading: a 78-billion-parameter mixture-of-experts with 3.5 billion active parameters is a competent industrial model, not a frontier one, and Apache 2.0 just means Europe gave away the only asset it had. The cheerleading reading: at last, sovereign capability, and the numbers hold up. One camp is buying a horse; the other is selling one. Neither has noticed that the race is not the business.

Benchmark leads at this tier have a shelf life measured in quarters. The gap over Qwen on AIME 2025 will be closed or reversed before most procurement offices finish reading the technical report. If Kolibri’s value is that it scored a 96.9, then Kolibri’s value expires in about eleven weeks, and Europe has spent a continent’s worth of political capital on a rounding error.

The thing that will outlast the benchmark table is duller. It is who can be made to answer questions about the training run.

Compellability Is the Actual Asset

Twenty trillion pretraining tokens went into this model. Where they came from, who licensed them, what was scraped, what was excluded, and who is legally on the hook if the answer is wrong — that is now a question a European regulator can ask and expect to have answered.

A German data protection authority, a Bundestag committee, a state procurement auditor, a civil litigant: ask a San Francisco lab for the corpus provenance and the honest answer is a trade secret plus a jurisdictional shrug. Ask Aleph Alpha and the answer comes from a company inside the reach of German law, trained on machines in Germany and Finland, staffed by people who can be deposed, accountable through a technical report that has to survive contact with auditors holding contractual leverage. That is a legal product. It is not glamorous.

“We’re not buying a model,” a procurement lawyer for a German state authority told me in the hallway outside a Landtag committee room in Düsseldorf. “We’re buying someone we can serve papers on.”

Apache 2.0 is what completes the purchase. It is not the “free” part that matters — the weights could have been cheap or dear and the legal point would stand. It’s that the deployed artifact can be hosted, fine-tuned and audited on European-owned racks, which means the entity running inference is also inside the jurisdiction. Open weights are usually sold as a cost story. Here they are a venue story.

The tell sits in the corner of Aleph Alpha’s own table: German-language AIME 2025, 87.5 for Kolibri against 82.9 for Qwen. A model that reasons natively in administrative German is worth more to a Finanzamt than a model that wins a chemistry quiz nobody assigned. Nobody buys sovereign AI to win AIME. They buy it so the answer arrives in a language their auditors read and from a corporate entity their courts can reach.

Nobody Needed Seven Gigafactories for That

On July 30, the EuroHPC Joint Undertaking opened a call for up to seven AI gigafactories, targeting roughly €30 billion in combined investment — up to €5 billion from the EU, matched by at least €5 billion from national governments, with the remainder solicited from private investors. Seventy-six expressions of interest came back, and ten countries are angling to host, per reporting on the tender.

If the product you’re buying is compellability, provenance and deployment rights, that is not the quantity you buy. You buy it with clauses: source-code escrow, weights custody, mandatory disclosure of pretraining corpus provenance, contractual audit rights, an obligation to publish the results. That costs lawyers. It does not cost seven buildings.

The public money in the gigafactory plan is, by its own terms, an anchor-customer commitment — the EU promising to purchase compute from facilities it is helping finance. That can be defended. But it is a compute purchase defended as a compute purchase, and it has almost nothing to do with whether a Munich court can get a straight answer about where twenty trillion tokens came from.

Here is the uncomfortable part for the people who dismiss sovereign AI as subsidy theater, and who are right often enough to be annoying: the narrow version of the sovereignty argument is not a subsidy argument at all. Governments buy jurisdiction constantly — they call it courts, contracts, subpoena power — and nobody files that under industrial policy. Jurisdiction over a training run is a genuinely new article of commerce.

And here is the uncomfortable part for the sovereignty crowd: they keep buying capability, and what they keep getting is models that are honestly competitive and irrelevant to the thing they claim to want.

Three Weeks of Discipline, One Day of Theater

Aleph Alpha sat on Kolibri Origin — pretraining finished June 11, no public release — and then shipped Kolibri three weeks after its own run completed. That is unusually good discipline in a field that treats announcement calendars as product.

The Reunification Day launch date is theater, and the Hacker News front page is theater’s receipt. Both will be worth less than a single escrow clause.

If Europe’s claim is that it can run AI on its own terms, the claim isn’t proved in an AIME table or by a symbolic release date. It’s proved in an audit a European regulator can compel and a European court can enforce — a small, boring, legally serious product that Aleph Alpha has now partly demonstrated. That is a real argument for a real thing.

It is not an argument for €30 billion.

Sources