Europe can still win the AI race: here's how

If despondent LinkedIn posts and press articles are to be believed, our insistence on putting citizens rights before growth has lost us the AI race. But as AI’s momentum stalls in the US, Europe has a unique opportunity to catch up, if it plays its cards right.

A year ago, you’d be forgiven for thinking the AI race was over: the two biggest AI companies, OpenAI and Anthropic, were well ahead of anyone else on the market: their strategy of throwing more data and compute at the problem was extraordinarily costly but it was resulting in incremental but undeniable performance gains.

Then came Deepseek: a Chinese model that was trained and could run at a significantly lower cost, that got AI companies and developers alike asking: “Is this really the best approach?” From then on, the cracks started to show: performance gains stalled, US AI companies rolled back their promises of revolutionary “Artificial General Intelligence”, and investors started to ask themselves if these companies — who are yet to find a path to profitability, and are losing money even on their paid subscriptions — were actually the solid investments they once believed them to be.

OpenAI's projected 2024 costs. Image Credit - The Information

The US has doubled down on the approach of its AI companies, building full-blown ecosystems around the two giants, and partaking in unsustainable multi-billion dollar “circular financing”, putting growth before profitability and sustainability, and further inflating the bubble ever closer to a popping point, and funnelling enormous proportions of available investment capital into generative AI, a technology that is costly, stagnating, and unprofitable.

A chart demonstrating the circular nature of investments in AI companies. Image Credit - Bloomsberg News

Meanwhile, in Europe, things have been quiet: with the exception of Mistral, Europe has no major commercial Large Language Models, but we were never going to be first out of the gate: Europe’s investors and private equity firms are dramatically fewer in number, and significantly more risk averse: for them, a clear path to profitability is key, something these US firms lack.

Herein lies the opportunity for Europe: by doubling down on generative AI, and ever-larger Large Language Models, the US has left little interest and capital for investment in other types of AI: be it smaller LLMs or non-generative, purpose-specific AI models, types of AI Europe might be extremely well-placed to build.

A first option for Europe would be to fund building a “good enough” Open Source LLM, focussed on efficiency and smaller datasets, leveraging existing Open Source efforts such as EleutherAI’s Common Pile dataset and Comma model, and augmenting them with additional, ethically sourced European data. Such a model could serve as a foundation for European AI companies to build on, as Open Source models can be reused freely. This could serve as an engine for economic growth.

But a second, much more interesting option, would be for Europe to focus on doing what it has always done well: specialised, business to business AI systems: weather, agriculture, health, and industrial AI might not be sexy and consumer facing, but it’s the kind of AI businesses actually need, and its cheaper to build, train and run. It would also allow Europe to leverage something Silicon Valley start-ups simply can’t: decades of experience, data and expertise our companies have in these specialised areas.

To achieve this, Europe has to be strategic: the EU has already committed to spending significantly on compute and funding for AI projects, but EU funding in these areas has a reputation for being scattered, unopinionated, and ineffective. If we want to propel Europe forward, we should focus investments strategically to achieve the aforementioned goals: funding should be heavily focussed on specialised AI with proven use-cases, built in collaboration with industry. The proposed Apply AI strategy should be strengthened with this collaboration in mind, and the creation of European Data Spaces should be aggressively accelerated.

European Dataspaces are meant to facilitate free flow of data across sectors and countries, and enable more sharing and collaboration around data.

Funding of LLMs and Generative AI should be restricted to a very limited number of projects, to ensure compute and investment is focussed on specialised AI, with a requirement that the outcomes of EU-funded generative AI projects be Open Source so that they can serve as a foundation for European companies to build on.

Finally, one of the biggest barriers to the success of such a plan is talent. Again, here, the United States has offered Europe a once-in-a-lifetime opportunity: the collapse of the rule of law, increasing political instability, and demonisation of immigrants and other minorities has meant the US has become a less attractive place to build a career.

In this context, Europe needs to make its case to the world: Europe is a continent where innovation, democracy and citizens well-being can co-exist, and where talent is welcome. There is a vacancy for a land of opportunity: it is vital that Europe fills it, and becomes the place to be for researchers and experts in the AI space. Although that doesn’t mean deregulating: our laws are what balance our continents success with our citizens well-being, so while we should focus on streamlining compliance, we don’t want to upset the balance between innovation and rights, and we don’t need to: what we really need is a vision.

The AI race is a classic hare-and-tortoise scenario, and Europe has always been the tortoise, but Europe won’t be starting from scratch: there is a wealth of research and Open Source code and models we can work from, and although we’re the underdog, we do have a lot of assets: our slow and steady approach to investment and innovation means we haven’t sunken billions into generative AI, while our extensive industrial experience gives us data and insights to build on. If Europe builds a clear, strategic vision on how to grasp this opportunity, we can prove that slow and steady really can win the race.

This article reflects my personal views alone. Thanks to j4p4n on OpenClipArt for the features picture.