A cluster of European semiconductor startups is securing major funding to build hardware specifically for AI inference, directly challenging Nvidia's GPU-centric approach. The move signals a strategic shift toward designing for the operational demands of running models, not just training them.

Dutch firm Euclyd, founded by former ASML director Bernardo Kastrup, is negotiating a round exceeding €100 million. The company claims its architecture, which processes data in multiple locations within the memory stack, can achieve a 100x power efficiency gain over Nvidia's latest chips for inference workloads. However, this design remains unproven in large-scale commercial deployment.

The engineering trade-off is clear: abandon the generalized, programmable architecture of GPUs for a more specialized, fixed-function approach optimized for inference. This promises lower energy and cost per query but sacrifices flexibility. As Patrick Schneider-Sikorsky of the Nato Innovation Fund notes, 'The existing GPU architecture wasn't built for [inference] in ways that matter most at scale.'

Other startups are exploring more radical physical-layer shifts. UK-based Olix is developing photonic processors, using light to move data, to address the thermal and miniaturization limits of electronic chips. Yet, these companies face the immense systems challenge of moving from lab prototypes to volume production in a region with a less mature foundry ecosystem than Asia or the US.

The funding surge, while notable, underscores a gap. European AI chip firms have raised $800 million this year, compared to $4.7 billion for US counterparts. The path from tape-out to deployed data center rack is long and capital-intensive, testing whether novel architectures can transition from compelling benchmarks to reliable infrastructure.

Source: CNBC