The useful way to read this week is not “Mistral had a bad week”, though it did. It is that four things which used to be bought as one are now being bought separately, and a serious buyer just proved it.
Start with the mechanism. For two years “European AI” meant one thing: a European lab trains a model, and you buy the model, the hosting and the political comfort in a single decision. This week each of those came unbundled. Mistral previewed Large 4, a trillion-parameter model it calls the strongest open-weight model outside China, which on Artificial Analysis works out to sixth among open models and 37th overall. Three days later Ecosia, the Berlin search engine, dropped Mistral for open models, some of them Chinese, served by Melious.ai on European green power. Brussels began benchmarking Chinese models itself. And OpenAI became the first frontier lab to watermark output for the AI Act.
So the layers are: who trained the model, where it runs, who paid for the building, and what you have to stamp on the output. Each now has its own supplier and its own price.
The money underneath is public, which is the part I find most telling. Bpifrance went into a Paris company that bootstrapped for twelve years. CDP sits behind a Milan pre-seed fund. EuroHPC is offering €10 billion to whoever can raise €20 billion more to build the compute. The state is not picking the model layer; it is funding the layers below it and letting buyers choose.
Here’s the implication I would hold onto. Once a buyer can mix a Chinese model, a German host and a French compliance stamp, the European lab stops being the default and becomes one option in a stack. Ecosia is small. The decision is not.
There is one layer nobody can route around, and it is the grid. Italy’s numbers, further down, show how far behind it is.
Ecosia dropped Mistral. Ecosia is steward-owned and has never taken venture money, which matters for how to read its choice. Founder Christian Kroll says the point is open weights, not China, and that Mistral’s US investors make him doubt its independence. The timing, three days after the Large 4 preview, is what turns a procurement decision into a market signal: a European buyer chose European infrastructure over a European lab, and the rest of the market will have noticed.
SAP is buying TechWolf. The Ghent company maps the skills employees actually use from workforce data; customers include HSBC, PayPal and Booking.com. Price undisclosed, closing in Q4. TechWolf keeps its name, its Ghent headquarters and its non-SAP customers under CEO Andreas De Neve. The structural point is an eight-year-old Belgian company getting a European exit without leaving Belgium, which people here will tell you does not happen.
Wiremind raised €35 million. The Paris company sells machine-learning pricing to rail and air operators. It ran twelve years without outside money, stayed profitable, and grew revenue 67% last year. The round is IronWave, Bpifrance’s Large Venture fund and Seaya Andromeda, with the founders keeping the majority. I think this is the honest picture of European AI that earns: boring customers, a decade of proof, and the state bank arriving after the proof rather than before it. The Paris company sells machine-learning pricing to rail and air operators, ran twelve years without outside money, stayed profitable and grew revenue 67% last year. The round is IronWave, Bpifrance’s Large Venture fund and Seaya Andromeda, with the founders keeping the majority. It matters because this is what European AI that makes money mostly looks like, and because the state bank came in after the proof, not before.
SimpL raised €1.2 million pre-seed. The Milan startup builds AI that finds B2B sales leads and charges only for meetings that actually happen. Techshop Capital led, with Turin’s Vento Ventures and angels. Techshop’s new fund had a €43 million first close with CDP Venture Capital as anchor. The same public-money-plus-private-fund shape as Wiremind, now at pre-seed scale in Italy.
Also this week: PSV Tech closed Fund II at €56 million in Copenhagen; Monte Carlo Capital raised €13 million for deeptech and AI; seed rounds for Klarent in Zurich, Besso in Bern and TODAY in Copenhagen.
OpenAI will watermark ChatGPT in the EU. On October 5 it published the method it will use to meet Article 50 of the AI Act: an invisible mark on text generated for EU users. The API stays opt-in. Two things follow. It is the first visible compliance move by a frontier lab, so the others now have a reference point. And anyone building on the API for European customers owns the obligation themselves, which is the kind of sentence that moves cost from a lab’s balance sheet to yours.
Brussels is benchmarking Chinese models. Politico found that ENISA and the Commission’s Joint Research Centre have started testing Chinese open models alongside US and European ones, on secure machines, with no plan to use them in daily work. It feeds the joint AI testing platform from the July cybersecurity plan. The EU is measuring the same gap Ecosia acted on; the difference is that Ecosia had to decide and Brussels only has to observe.
The gigafactory call is open. EuroHPC will fund up to seven AI gigafactories, four medium and three large, at roughly €4 to 5 billion each: €10 billion public, at least €20 billion private that the bidders raise themselves. Minimums run from 25,000 to 100,000 H100-equivalent GPUs; JUPITER, Europe’s largest today, has about 24,000. The coordinator must be EU-incorporated and EU-controlled. This is the biggest single public bet on European compute, and the winners decide where the next decade’s models get trained. Note the ratio: for every public euro, two private ones, found by the bidder. That is a test of who in Europe can raise €13 billion for a building.
Italy’s grid is the bottleneck. The Italian Datacenter Association’s study, out October 8, projects €36.9 billion of investment by 2036 and capacity rising from 460 MW to 2.3 GW by 2031. Operators have asked for 60 to more than 100 GW of grid connections; about 1.9 GW of projects are actually advanced. Read those two numbers against each other. The gap between what is requested and what is buildable is where the money and the politics will go, and no model license changes it.
The UK is writing the agent rulebook. The ICO opened a six-week call for evidence on agentic AI and data protection on October 8, aimed at deployers as well as developers. It feeds a statutory code of practice on AI and automated decisions expected in 2027. If you sell agents into the UK, the arithmetic is simple: answering now is cheaper than complying later.
Mistral Large 4 weights, 27 October. Download and license published after three weeks of red-teaming with security partners and state authorities. The license is custom this time, not Apache 2.0 as Large 3 was. The fine print decides who can build a business on the model, which after this week is the whole question.
Web Summit, Lisbon, 9 to 12 November, and Slush, Helsinki, 18 to 19 November. The two biggest founder-investor gatherings in Europe, back to back. This is where the gigafactory consortia and the open-model inference companies will be pitching. Lisbon hotels are already going.
Gigafactory bids close 12 November, Consortia submit via EuroHPC’s eSubmission; 18 member states have signed up to co-fund. The shortlist will tell you which countries and which private backers are serious about compute.
AI Act watermarking deadline, 2 December. Generative systems already on the EU market before August 2 must carry machine-readable marks by then; newer systems already should. It is the first AI Act obligation with a hard date that lands on product teams rather than lawyers.
Global Startup Awards, applications open. Twelfth edition, free, no equity, built for early-stage founders. Worth it if you are pre-seed and want a cheap signal for the next raise.
If you’re building something, reply and tell me what it is, I read everything. If you’re in Turin, say so.
I write two other newsletters, and they come from the same place this one does: twenty-odd years at agencies and scaleups as a creative and later a CMO, mostly where marketing and new technology collide.
AI-Ready CMO is the daily one for marketing leaders. What changed in AI today, what it means for your team, what to do about it.
AI-Ready Ads is daily too, and short: one example a day of AI used well in advertising.
Both are free. If this one is useful to you, those probably are too.