With the AI race at full speed, European countries have been looking for their ticket to the latest technology wave. I’m spending the majority of my time in Europe these days and I’ve been restlessly watching Europeans contort themselves in futile attempts to gain a solid footing. As things stand now, the continent is destined to miss this wave the same way it did with the web, social media, cloud computing (hyperscalers), and mobile. The reasons lie mostly in a misguided mindset and wrong priorities.
The largest confusion I see in Europe is misunderstanding what you need to do when you’re catching up after decades of falling behind, and why “sovereign AI” might sound good, but is a losing idea. The end result is that the continent, instead of sitting on the frontier of value creation, chooses high-school version of the AI race.
The cry of sovereign AI
The idea behind sovereign AI is that since AI development is dominated by American industrialists, for Europe to participate we should work on our own versions of AI, e.g. foundational models to have a say in how they work.
For example, there are calls for national LLMs that would support local languages better (German, Polish, French, etc.). As a byproduct local communities would develop competency in the technology, and have a say in development priorities of the AI models.
This sounds good on the surface, and is followed by a call to local governments for funding these efforts. Almost every country has something like that going on: Spain has Salamandra/Alia, Germany has NeoLM, Poland has Bielik, etc.
The problem? Language support and reflection of the local culture are side-quests. The leading LLMs still struggle with reasoning, latency, and instruction-following, which limits deployment to the real world tasks. Fixing these problems is where all the steam - and billions of dollars - are going into. The European versions of local LLMs, having smaller budgets and engineering resources, cripple their abilities and limit even further the tasks they can take on.
The idea of training LLMs on specialized corpus of data has been around for a while, but ever since BloombergGPT experiment, we know it’s a losing strategy. Generalist models trained on a wide variety of tasks beat specialist models in their specialist domain (e.g. finance for BloombergGPT). The power of generalist models applies to languages and cultures too.
The sovereign AI strategy relies on a group of engineers in California making a choice to not focus on gathering training data for Polish or German. The moment that becomes a priority, the support for these languages and cultures are added, any temporary advantage sovereign AIs had evaporates.
In the meantime, the leading LLMs will get the tooling, best practices, and infrastructure (chips, security, orchestration, etc.) built for them. Over time, the rational choice will be to switch to leading LLMs, on cost, quality and time-to-market basis. The “this technology is more valuable for me because of others using it” phenomenon is called network effect.
Europeans consistently ignore network effects while it is being one of the strongest forces in the modern technology development. Ben Horowitz - the legendary entrepreneur and investor - explains the importance of the network effect.
Anyone mulling over the development of “sovereign AI” should ask themselves: why shouldn’t we be funding a European version of Microsoft Word? And if we funded it, would anyone want to use it over the real Word?
Lastly, anyone thinking seriously about “sovereign AI” should study the history of Quaero - “European Google” created in the name of sovereign over access to the internet. If you’ve never heard of it, you get the point.
While sovereign AI initiatives consume precious attention and engineering talent, they distract from where Europe could actually win. The real opportunity isn’t in recreating foundational models - it’s in being first to transform industries with AI applications. And here’s why.
AI application layer is the king
We’re early in this technology development cycle, and the end result of it will be that for the first time computers will be producing units of cognition. Increasingly, software will be directly taking on knowledge intensive and messy human labor. The vast majority of value creation and industry creation will happen not at the foundational model layer, but at the application layer. Similarly to how much of the value in cloud lies in tens of thousands of vertical software products built on top of it, for all sort of real-world problems and industries.
Europe has lost the foundational race before it even started due to little investment in AI for the last decade, not having big tech companies with vast amounts of readily available data, underdeveloped risk capital, a meager infrastructure, and high energy costs.
However, the real AI race is at the application-layer products and that competition is still at day 1. Instead of trying to run a high-school AI race of small budgets awkwardly copying what American labs are doing, the continent should focus on making gobs of money with deploying vertically-integrated AI, transforming both European and world industries. At this stage of technology development, making great AI products requires top-notch talent that Europe has. It also requires a mindset of winning, which Europe forgot.
The real sovereign AI will emerge from having a thriving set of industries deploying world-winning AI products. Once you have reach that milestone, you’ll have resources, deep expertise and specific demands informing whether you need to build your own foundational technology. For now, we should ride the wave of humongous investment and risk taken by American tech giants, and benefit from the tailwind of technology improvements that happen every couple of months.
A call to my European peers: modern technology is driven by network effects and not wishes or virtues. In every major technology category there’s room for only 1-2 winners. We should remember that, and intensely focus on building these winners most likely found in vertical AI applications. With the grind, effort, ingenuity, grit and customer obsession it requires. With understanding that those who succeed in this game deserve to be paid tens of millions of dollars. Bamboozling about sovereign AI won’t get you anywhere.
Europe often thinks “innovation” precedes “commercialization”. If you study the actual history of technology, you’ll see it’s the other way around. Cloud in the form of AWS came out of the largest bookstore, Amazon, not the other way around.
Since I’m writing from Poland, here’s a word for my country. Poland should focus on vertically-integrating AI into areas it already understands at global level: back-office work it insources (tens of billions of dollars to be made there), logistics, talent outsourcing, and any other we have deep domain expertise. Not in training a Polish LLM that nobody will remember in a couple of years (similar to Grono and Nasza Klasa, the social networks of the day).
The continent should play offense where it can, instead of defense where it cannot.
P.S. A notable, very recent example of a winning product mindset in Europe is the Lovable team (https://lovable.dev).
Thanks to Jacek Migdał, Adriaan Moors, Jim Lee and Sebastian Kondracki for reading drafts of this post.