Two weeks ago I was typing away on Spain’s southern coast. Then the wifi went out. And then the phone signal.
I assumed it was a local problem, that the lights would begin to flicker at any moment. But it wasn’t to be. As morning rolled into afternoon, it became clear the whole country was in trouble.
When contact with the planet finally resumed, the result was underwhelming. The emerging consensus is there was no foul play, just a creaky Spanish electric system that couldn’t handle a bit of unusual atmospheric movement. How underwhelming. Worse, how embarrassing.
Our American friends couldn’t resist rubbing it in. This was classic Europe, wasn’t it? The place where nothing really works and everyone clocks off at 2pm to have a beer at the local chiringuito.
On the “digital town square” of X dot com, the episode was fodder for the discourse that doesn’t die: the one about idle “Europoors” versus American wage-slaves. On one side sits GDP, car worship and hard power. On the other we’ve got life expectancy, al fresco dining and bureaucracy.
You can’t train models on vibes
It’s a joke, obviously, but one not completely divorced from reality.
Take Big Tech. A few years ago Emmanuel Macron said: “In the United States, they have GAFA. But in Europe, we have GDPR.” It’s a good, albeit outdated, line.
No-one talks about GAFA (Google, Amazon, Facebook, Apple) anymore. Instead we have the Magnificent Seven: Alphabet, Amazon, Apple, Meta, Microsoft, NVIDIA and Tesla. The combined market cap of these firms alone would make them the second-largest country stock exchange in the world — and makes it almost impossible for other companies to compete in the AI race.
AI is a phenomenally expensive enterprise, one that is as much about raw materials and energy as it is clever algorithms. You need to get hold of NVIDIA’s chips, stick them in the datacenter somewhere, and ask them nicely to multiply matrices until the sand starts to think.
This is the basic reality pretty much every national AI strategy outside of the US tends to ignore. You need cash; you need energy; you need know-how — in that order.
While it is possible for countries other than the United States and China to compete, the deck is stacked against them.
This is the context in which Europe eyes American talent, and the researchers know it. They read the news and scroll through X. They see what everyone else can: Europe doesn’t have the support structures, it doesn’t have the track record, and it might not have the stomach to compete.
AI researchers, like most of us, want their work to matter. Because AI is more-or-less now a question of scale, they are naturally hesitant to pack up their bags and move somewhere without a single globally competitive general purpose AI model.
The problem isn’t that the EU doesn’t have money to spend (it does) but that it is being forced to spend it in the first place. In the US, private companies can use their cash reserves to finance the infrastructure required to build bigger and bigger models.
In Europe, startups like Mistral have built solid clusters, but they tend to be the exception that proves the rule.
The EU’s “AI factories” initiative, which aims to repurpose existing supercomputers for training AI models, shows the state can move when it needs to. But the fact so few private alternatives exist points to a deeper sickness.
A question I often hear is: ‘“What about ASML?” The argument goes that if AI is an infrastructure game, and Europe holds a key part of the production stack through the Dutch lithography firm, then no matter what the future is secure, at least in part.
But hold on: Yes, it’s nice to have a piece of the pie rather than none, but ASML’s presence serves to remind us that global chip production is a massive, complex affair with lots of points that Europe should be competing on. Given all that, perhaps we might hope for a bigger slice of the ecosystem rather than a single bottleneck.
Defying gravity
AI progress is cumulative and collaborative.
Labs hire the right people, spend the cash, train the models and smooth out the wrinkles. When I worked at Google DeepMind, I was amazed by how quickly a bunch of smart people could create something extremely cool.
In part that’s because success creates success. Great work happens when people are faced with a critical mass of good ideas, feedback loops from product teams used to shipping and the freedom to experiment with different types of models at different stages of development.
The US is a black hole of elite labs, generous funders, fast-moving startups, and zealous talent. Partly that’s because American AI is magnetic. Not in the sense that it has some unique je ne sais quoi, but because the inputs to AI – compute, money, people and data — obey the laws of attraction.
Each one pulls the others closer to create a virtuous circle. Talent draws investment, investment builds infrastructure, infrastructure enables breakthroughs and breakthroughs bring more talent. Once that loop starts spinning, it’s very hard to stop it.
But even ignoring all of these, European AI firms wonder whether if they can prise talent away from America then they might be able to build a flywheel of their own. They’re right to try, but should remember they’re in for a tug-of-war with a giant.
The first hurdle to clear is the one we’re not supposed to talk about: compensation. If you have ever seen the American AI labs post jobs with salary details, you might be forgiven for thinking someone made a mistake. A top US researcher is probably not too far away from making a million dollars a year. Sometimes more when you factor in equity.
Talented people aren’t greedy, but they’re not stupid either. If your work is going to make billions for someone else, then why wouldn’t you want to make sure you’re as fairly compensated as possible. But for Europe, the money isn’t just smaller, it also comes with extra baggage.
Just look at the flack taken by the UK AI Security Institute for paying more than an average civil service wage. First you get those who think, whatever the increase on the baseline, these types of organisations can’t ever compete with American labs. Then, even worse, you get Treasury-brained penny pinchers who say a public organisation simply can’t afford to pay over the odds in any context.
Finally, there’s culture. I don’t mean that Europeans are ‘less entrepreneurial’ or whatever. I mean that, in the United States, a not-so-small portion of the employees of the frontier labs think there is no place more important they could spend their lives. At the top end of the spectrum you’ll find people who want to capture the light cone of all future value in the universe, and at the bottom end you have people who think AI is the frontline in a new great power competition.
In other words, people believe in AI. You might call it messianic, but it’s nothing if not motivating. Compare that to Europe, where AI is treated as a governance problem before it’s even a real product. For researchers who think it’s time to build, Europe feels like it’s holding the screwdriver upside down.
So yes, Europe is beautiful. The food’s better, the holidays are longer, and nobody asks you to code through Thanksgiving. But if you’re an AI researcher deciding where to create the future, good vibes aren’t enough. You need capital, you need culture — and you need conviction. Right now, the US has all three.
Europe might catch up, but we’ve got a long way to go.
Harry Law is a researcher at the University of Cambridge. He writes about AI, history, and culture at www.learningfromexamples.com.
Read the orginal article: https://sifted.eu/articles/google-deepmind-ai-eu-us-talent-tech/
