AI chips (2025)

Chips: little miracles that appear on our dinner plates and in our supercomputers. But it's the latter version — sorry french fries — that's currently the the most coveted thing in the world. Specialised chips are powering the boom in chatbots like ChatGPT, Claude and Grok. Everyone’s hungry for them: just look at the AI infrastructure spending plans of the great tech powers. Google will invest $75bn in chips and data centres this year; Microsoft will spend $80bn and Facebook’s parent company Meta plans to spend up to $65bn.

The small wrinkle — for Europe — is that it’s entirely reliant on chips made from a handful of US firms, most often world-leading chip designer Nvidia, and to a lesser extent Advanced Micro Devices (AMD) and Intel. Nvidia’s graphics processing units — or GPUs — are valued for their ability to crunch the massive amount of data required to develop AI models. They sell for eye-popping sums (anywhere from $15k-$40k).

Such is the global demand (and geopolitical sensitivity around) Nvidia chips that the US has capped access to them. Western European countries generally face no restrictions, while most eastern European countries now have limits on the amount of chips they can buy. Good for France, bad for Poland. Still, Paris startups say it's a struggle to acquire the chips. “We have to fill in pages and pages of [export control] documents,” Robert Marino, CEO and cofounder of Qubit Pharmaceuticals, told the Paris AI summit last week.

Can European chipmakers rise up to compete with Nvidia? Europe has semiconductor specialists like ASML, Infineon and STMicroelectronics but they can’t do what the Silicon Valley market leader does. Nor is Europe likely to develop AI chips without more firepower: the region’s chip challengers raised some €192.9m in equity and grants in 2024: small fry when huge sums of money are required to bring chip ventures to market.

Wannabe challengers to Nvidia include Amazon, AMD and several others, which are beginning to offer credible new options for a phase of AI development known as “inferencing” — this is the process that enables AI to serve up answers with chatbots. One of Europe’s inference hopefuls is London-based Fractile (we talk to its founder and CEO below). Another European chip player to watch is UK-headquartered Arm Holdings, which the FT reports is planning to launch its first-ever complete semiconductor after securing Meta as one of its first customers.

Fortunately for European startups, amassing hordes of chips is no longer the only way to get ahead in the AI race. Chinese startup DeepSeek showed the world a different path recently when it unveiled a powerful AI system using far fewer computer chips than many experts thought possible. Today’s AI chatbots require 16k specialised chips or more; DeepSeek said it needed only about 2,000. It’s the sort of tech trickery Europe needs to learn from.