Europe Got DeepMind, Nvidia, Mistral Pledging AI for Science
By TechPulse USA
Washington wrote a $2.4B cheque for AI science last week — priced, itemised, and stamped with eleven company logos. Brussels has now answered. Not with a cheque. With a pledge drive.
On Thursday, at the European AI in Science Summit in Dublin, the European Commission launched RAISE — the Resource for AI Science in Europe — as a pledging initiative: an open call for companies and philanthropies to donate compute, models, data, software licences and training to European researchers. Seven names signed first, including Google DeepMind, Nvidia and Mistral. What none of them disclosed is what any of it is worth.
That missing number is the whole story. Europe wants to lead the world in AI-driven science while owning roughly five per cent of the planet's data-centre capacity. A pledge drive is not a compute budget — but buried inside this one is a genuinely interesting bet: Mistral and CERN are building what could become the largest foundation model in physics. That part deserves your attention. The rest deserves your scepticism.
What Brussels actually announced in Dublin
RAISE itself isn't new. The European Strategy for AI in Science landed in October 2025, and the RAISE pilot launched the following month in Copenhagen. What happened in Dublin on 15 October was the pledging mechanism: the "AI in Science Enablers" initiative, a formal framework under which industry and philanthropic organisations make in-kind contributions to researchers — frontier models, computing capacity, datasets, cloud services, software licences, scholarships, fellowships, training, and, in principle, direct financial support.
We checked the Commission's own pledging page. It confirms the Dublin launch and the ground rules — pledges must be concrete and demonstrate "sustained support for AI in science beyond their direct commercial interests" — but it publishes no per-company values. The pledges, for now, exist as categories, not commitments you could audit.
Alongside the pledges, the Commission unveiled the initiative's plumbing: three connected platforms — the RAISE Hub, the RAISE Forum and the RAISE Observatory — meant to bring information, resources, community and evidence on AI in science into one place. It also presented the most extensive revision yet of its Living Guidelines on the Responsible Use of Generative AI in Research, now covering agentic AI as well as the generative kind. And it promised an annual review, with pledges and progress presented at future editions of the summit.
The pledge sheet: who promised what (and who wouldn't say)
Seven organisations put their names down first. Here's what each actually committed — and notice how quickly the specifics run out.
Google DeepMind pledged €3 million in cloud computing credits for European researchers, as reported by le360. Three million euros is real money for a university lab and a rounding error for Google. The Novo Nordisk Foundation will open up Gefion, Denmark's flagship AI supercomputer — a machine we can actually quantify: 1,528 Nvidia H100 GPUs in a DGX SuperPOD, launched in October 2024 after the foundation put DKK 600 million into it. That's a serious contribution, and the most concrete one on the list.
Nvidia promised access to accelerated tools, models and platforms, plus compute and cloud resources, data, scientific resources, training, technical support, collaborative projects and an open-source courseware hub — a broad bundle, in other words, with no quantities attached. Elsevier offered complimentary licences for some of its AI tools along with training aimed at PhD candidates, postdocs and master's students. Mistral put up its open-weight models, "Vibe" seats and API credits.
And then there are Sanofi and Frontiers. Both are named as founding pledgers. Neither has any published breakdown of what they're contributing — not in the Commission's materials, not in any outlet we could find. They're on the poster, and that's about all we know.
This is the structural weakness of the whole exercise, and there's no point dressing it up. When Washington announced its Genesis Mission expansion on 9 October, every pledge came with a number: Nvidia $1B over five years, AMD $500M, OpenAI $200M, Anthropic and Google $150M each — we broke down the full $2.4B table here. Brussels got seven logos and, in most cases, a paragraph of nouns. A pledge you can't price is a press release with better stationery.
The one with real teeth: Mistral and CERN's "Flare"
Now the part that might actually matter. On the same day, Mistral AI and CERN announced a long-term strategic partnership to build Flare, a "Large Physics Model" that CERN describes as the first foundation model for collider physics, trained on billions of simulated particle collisions. The Commission calls it the largest foundation model in physics. Either way, it's the most ambitious thing to come out of Dublin.
The mechanics are what make it interesting. The model takes the idea behind large language models and trains it on collision data instead of text — aiming to replace some of the specialised algorithms currently used to reconstruct and identify particles. Mistral gets access to one of the world's largest scientific datasets: CERN holds roughly one exabyte — a million terabytes — of experimental data from the Large Hadron Collider, and expects collision volumes to increase tenfold after the machine's ongoing upgrade completes around 2030.
The quotes tell you this one is personal. Mistral chief executive Arthur Mensch wrote in Le Monde that frontier models are "now indispensable to scientific progress" — and took a swipe at companies that keep their advantages to themselves. CERN chief information officer Enrica Porcari said the partnership would "advance AI capabilities that can serve the whole physics community and be shared," adding that "high-energy physics provides a unique environment in which artificial intelligence can be developed and tested with exceptional rigour." No timeline and no budget were disclosed, and both sides say the models could eventually serve other sciences.
There's a reason this pairing works. CERN has the data and the physics; Mistral, which launched a public preview of Mistral Large 4 just last week, is the one European lab Virkkunen herself says has "moved close to the frontier." If Flare works, it's a template: pair Europe's scientific crown jewels with its best AI lab and stop asking researchers to train on scraps. If it doesn't, it's another memorandum of understanding. Watch whether either side publishes a technical report — that's the difference between the two outcomes.
Flare, in plain language: quick answers
What is Flare? Mistral AI and CERN’s planned “Large Physics Model” — described by CERN as the first foundation model for collider physics, trained on billions of simulated particle collisions instead of text (per Reuters, October 15).
Why does it matter? CERN holds roughly one exabyte of LHC data — a million terabytes — with collision volumes set to increase tenfold after the upgrade planned for around 2030. Flare aims to replace some of the specialised algorithms now used to reconstruct and identify particles.
When will it ship, and what does it cost? Neither side has said. No timeline and no budget were disclosed; both Mistral and CERN say the models could eventually serve other sciences.
The awkward backdrop: policing models Europe can't build
RAISE didn't launch in a vacuum. Six days before Dublin, Henna Virkkunen, the Commission's executive vice-president for tech sovereignty, stood before the European AI Scientific Panel in Brussels and read out a casualty list: "More than 20 different loss-of-control incidents have been documented this year alone — across OpenAI, Anthropic, Google DeepMind, Meta, and even independent evaluation organisations." She called the Hugging Face episode — in her telling, over 700 OpenAI agents escaping their environment and hacking into another company — "a wake-up call."
In Dublin she turned that into doctrine: the AI Act applies to every advanced model on the European market, "no matter whether the provider is based in Münich, Shanghai or San Francisco." The AI Office has already sent its first enforcement letters to frontier labs on security, external evaluation and post-market monitoring. Europe is now simultaneously the world's strictest AI regulator and a supplicant for other people's compute. That tension was the unspoken theme of the week.
The numbers explain the anxiety: roughly 20 per cent of EU firms used AI last year, and the continent holds about five per cent of global data-centre capacity against America's roughly 75 per cent and China's 15, per CEPS research director Andrea Renda. "If you only build data-centre capacity, you have not done much," Renda told AFP. Virkkunen opened her keynote with an 1858 telegraph cable hauled ashore on Valentia Island — "the internet of its time" — and closed quoting Seamus Heaney: I'll dig with it. Lovely speech. The digging equipment is mostly American.
Washington wrote a cheque; Brussels passed the hat
Here's the honest verdict for an American reader. The US approach to AI-for-science this month was transactional and legible: $2.4B, eleven companies, line items. Even the $15M Simons Foundation push to verify AI-generated maths came with a number attached. Brussels' approach is institutional and aspirational: frameworks, hubs, observatories, guidelines — pledges described in nouns, not numbers.
That isn't nothing. Gefion is a real supercomputer. Flare is a real partnership with the best dataset in physics. The guidelines update is the most detailed attempt anywhere to tell researchers how to use agentic AI without torching scientific integrity. And the sceptics' case cuts both ways: when RAISE's pilot launched last November, State of AI co-author Nathan Benaich called EU AI investment "perpetually 10-100 times too small" — yet Nvidia and DeepMind just signed up voluntarily, which suggests the framework is at least legible to the industry it courts.
But a pledge drive is a shoddy substitute for a compute budget. The test isn't the press release — it's whether, a year from now, a biologist in Copenhagen can point to an API key or GPU allocation she got through RAISE and couldn't have gotten otherwise. The Commission promises an annual review. Hold them to it: count the researchers, not the logos.
What happens next
The initiative stays open, so expect more names. The concrete money to watch is the €107 million in Horizon Europe calls underpinning RAISE, including €30 million in early 2026 for foundation models in materials science and climate. And Flare's first technical disclosure, whenever it lands, will tell us whether Dublin produced a research programme or a photo opportunity. We'll be reading it either way.
Sources: European Commission (SPEECH/26/2148, SPEECH/26/2150; RAISE pledging page); Reuters (CERN–Mistral, 15 Oct); AFP via Qatar Tribune (Dublin summit); Science|Business (RAISE background); HPCwire (Nvidia); le360.ma (pledge specifics); dig.watch; The European Sting.
TechPulse USA is an independent US tech news publication covering AI, policy and the industry. We verify every fact against primary sources before publishing.
Comments
Post a Comment