On the morning of June 19th I opened the European Commission newsroom expecting the usual. Another strategy paper, another action plan, another PDF stuffed with the word "ecosystem." Instead there was a real announcement. Europe had picked a team to build its own open-source frontier AI model, trained in all 24 official EU languages, led by an Italian startup called Domyn with Germany's Fraunhofer-Gesellschaft alongside it.
I got genuinely excited, which does not happen often when I read anything with "Grand Challenge" in the title.
Then I kept reading.
What EUROPA actually is, and why it's the right call
The consortium is called EUROPA, and it won the Frontier AI Grand Challenge, a competition EuroHPC and the Commission launched back in February. The brief was a model with more than 400 billion parameters, open weights, capable across domains. Real frontier scale, not a toy.
And it runs on European iron. The plan puts the training on Europe's own supercomputers, machines like JUPITER in Germany, Leonardo in Italy, LUMI in Finland, and MareNostrum 5 in Spain. Domyn is already standing up a 6,000-chip Nvidia Blackwell cluster for it.
This is the thing the sovereignty crowd has been asking for out loud for two years. Open weights, so anyone can inspect and run them. All 24 languages, so it's not just another English model with French bolted on. Built and hosted here, so no US warrant can reach into the training run. If you had asked me in January what a serious European answer to the frontier looked like, I would have described something close to this.
So the concept is right. I want that on the record before I get to the part that deflated me.
The prize was compute, not a check
Here is where my mood turned.
Winning the Frontier AI Grand Challenge does not come with a war chest. What EUROPA gets is up to 2.5 percent of EuroHPC's total computing capacity for one year. That is the prize. A slice of shared, publicly funded supercomputers, for twelve months, and then the clock runs out.
Now, that compute is real and it is not nothing. Time on JUPITER is worth a lot, and handing it to a serious team beats letting it sit idle. I don't want to sneer at it.
But read the shape of the deal again. No multi-year budget sized to the ambition. No cash to hire the fifty best researchers away from the labs currently paying them American salaries. No fund to keep the lights on the day the free supercomputer time expires. Europe is trying to build a frontier model the way you'd fund a hackathon. Great venue, borrowed hardware, see you in a year.
OpenAI loses more than this in a month
To feel how small the bet is, you have to hold it next to the people Europe is trying to catch.
OpenAI's own forecast is a $14 billion loss in 2026, and that is the polite number. Some of the leaked financials put the cash burn closer to $25 billion this year, with the losses stretching out through 2028 before any promised turn to profit.
Do the division. Even the conservative figure is over a billion dollars a month, every month, gone. Whatever you decide a year of borrowed supercomputer time is worth, one American lab sets fire to more money than that before most of us have finished filing our taxes. Two months of OpenAI simply operating at a loss is larger than the entire financial commitment behind Europe's flagship frontier model.
That is the whole gap in one sentence. People love to blame regulation, and the LinkedIn crowd likes to blame a shortage of ambition, but neither survives contact with the numbers. The gap is money, and the willingness to keep spending it long past the point where a European board would have quietly pulled the plug.
Mistral, and the thing I got wrong
I have been telling people Mistral never shipped a proper reasoning model. That is not true, and I should correct it, because someone will fact-check me and they'll be right.
Mistral did ship one. Magistral landed in June 2025, Reuters called it Europe's first real reasoning model, and at launch it benchmarked roughly on par with DeepSeek-R1. So the honest version is not "they failed to deliver." It's that they delivered once, at a decent level, and then the frontier walked away from them.
The bigger dedicated reasoning variant everyone expected has been slow to arrive, and the models defining the reasoning frontier today are mostly not European. I don't think that is because Mistral's researchers got worse. Mistral has raised something like $4 billion in total, and is reportedly raising more at a 20 billion euro valuation. That sounds enormous right up until you set it beside the tens of billions OpenAI and Anthropic have swallowed. Europe's best-funded lab is running on what a US lab loses in a bad quarter.
Same story as EUROPA, wearing a different jacket. The people are here. The runway is not, so the frontier keeps slipping half a generation ahead and staying there.
The one bit of light
I've written before that Europe's real problem is never the tech, it's the on-ramp, the managed, hosted, sign-up-and-go thing you can actually use in ten minutes instead of a GitHub repo you're told to host yourself. Same argument I made about the missing free tier. Open source is not a strategy if using it requires a platform team.
For models specifically, that gap is finally starting to close, and it's the reason I'm not writing this whole thing in a minor key.
Take Melious. German company, a GmbH, and what it sells is boring in exactly the right way. One inference API, OpenAI- and Anthropic-compatible, pointed at a network of European infrastructure across eight countries with no US fallback in the chain. It already serves 60-plus open-weight models, European ones like Mistral's and Voxtral among them. You change a base URL and your existing code runs on sovereign infrastructure.
I tried it last week on a side project. Pointed the client at api.melious.ai, changed one line, and it just worked. (I worked at DeepL, so I've watched Europe build genuinely world-class model tech before. The tech was never the hard part. Getting it in front of people who don't run their own GPUs was.)
That's why this matters for EUROPA. When the model actually ships, someone has to serve it to the rest of us, and a sovereign frontier model that nobody outside a supercomputer can call is a research paper, not a product. Melious is the shape of the company that closes that last mile.
So where does that leave it
The compute-not-cash model is a bet that you can build a frontier model on borrowed time, open weights, and goodwill, then hope the money and the market show up before the free supercomputer hours run out.
I want that bet to pay off. I'm not sure it does.
The encouraging part is that the piece we've been missing for years, somewhere European to actually run these things, is finally getting built by companies like Melious. The unsolved part is the one this whole post keeps circling back to. Who pays to keep EUROPA alive after the twelve months are up, at the scale the Americans spend in eight weeks?
Ask me again in a year. I'll either be pleasantly surprised or writing the sequel.