A quiet but consequential shift is underway in the global semiconductor race, as a cluster of European startups moves to challenge the dominance of Nvidia by targeting what many see as the next decisive frontier in artificial intelligence: inference.
While Nvidia’s graphics processing units have become the backbone of the AI boom, powering the training of large language models and other advanced systems, the industry is increasingly turning its attention to how those models are deployed at scale. That transition is not merely technical. It carries profound economic implications, particularly as the cost of running AI systems begins to eclipse the cost of building them.
This is the opening that European startups are attempting to exploit.
Across the Netherlands, the United Kingdom, and France, companies are lining up substantial funding rounds to develop alternative chip architectures designed to run AI models more efficiently, according to CNBC.
The scale of ambition is evident in the capital being sought. Dutch startup Euclyd is in discussions to raise at least €100 million, while Britain’s Optalysys is preparing a similarly sized round. Firms such as Fractile and Arago are also seeking nine-figure investments, underscoring the growing appetite for hardware bets in a sector that had, until recently, been dominated by software narratives.
Early capital flows suggest that investors are beginning to treat AI infrastructure as a strategic asset class rather than a speculative niche. More than $200 million has already been deployed this year into companies such as Axelera and Olix, reflecting a broader recalibration of how value in AI is expected to be created and captured.
This shift is being buoyed by a growing recognition that inference, not training, will ultimately determine the economics of artificial intelligence.
Training large models remains capital-intensive, but it is episodic. Inference, by contrast, is continuous, embedded in applications, and highly sensitive to efficiency gains. Even marginal improvements in power consumption or latency can translate into significant cost savings when scaled across millions or billions of queries. This is where incumbents may be vulnerable, not because their technology is inadequate, but because it was not originally optimized for this phase of the AI lifecycle.
“Inference is dominant now, and the existing GPU architecture wasn’t built for it in ways that matter most at scale,” said Patrick Schneider-Sikorsky of the Nato Innovation Fund, capturing a view that is gaining traction among investors and engineers alike.
The startups emerging in Europe are not attempting incremental improvements. They are pursuing architectural departures.
Euclyd, for instance, is developing systems that process data across multiple points rather than shuttling it continuously through memory, a design choice that founder Bernardo Kastrup says could deliver orders-of-magnitude improvements in energy efficiency. If such claims are validated in real-world deployments, the implications would extend beyond cost reduction to the physical footprint of AI infrastructure, potentially easing the growing strain on data center capacity and power grids.
The company’s lineage reflects Europe’s deep, if often underappreciated, semiconductor expertise. Founded by a former director at ASML and supported by its former chief executive, Euclyd sits within a broader ecosystem that has historically excelled in chip equipment and design, even as it lagged in manufacturing scale.
Elsewhere, startups are exploring even more radical alternatives. Photonics-based computing, which uses light instead of electrons to move and process data, is being positioned as a potential successor to traditional semiconductor architectures. Companies like Olix are betting that the physical limitations of electronic chips, particularly heat generation and energy inefficiency, will accelerate the shift toward optical systems.
The argument is grounded in the realities of scaling. As transistor miniaturization approaches physical limits, the gains that once came from shrinking chip features are becoming harder to achieve. At the same time, AI workloads are pushing systems to their thermal and energy boundaries, forcing the industry to confront constraints that cannot be solved through incremental engineering alone.
Yet for all the momentum, the competitive gap remains formidable. Nvidia is not a static target. The company has aggressively expanded into inference optimization, while maintaining a dominant position in training. Its research and development spending, which exceeded $18 billion in its latest financial year, gives it the capacity to adapt to emerging architectures, including photonics. Its acquisition of assets from inference-focused startup Groq and investments in photonics technologies signal a clear intent to remain ahead of potential disruptors.
This dynamic complicates the narrative of disruption.
European startups are not competing against a complacent incumbent. They are confronting a company that has already begun to internalize the very shifts that challengers are betting on. Beyond technology, structural constraints continue to weigh on Europe’s ambitions.
The region’s semiconductor ecosystem remains fragmented, particularly in manufacturing, where reliance on external foundries such as TSMC exposes startups to supply chain risks and limits control over production timelines. Development cycles for advanced chips are long and capital-intensive, with the path from design to large-scale deployment often stretching over several years.
Funding disparities further highlight the challenge. European AI chip startups have raised around $800 million so far this year, a fraction of the $4.7 billion secured by their U.S. counterparts. The absence of a coordinated funding mechanism comparable to the United States’ defense-backed innovation ecosystem continues to constrain early-stage deep-tech development in Europe.
There are also institutional frictions. Industry executives point to conservative procurement practices among European governments, a lack of incentives to adopt domestically developed technologies, and regulatory fragmentation that complicates cross-border hiring. These factors, while less visible than technical hurdles, play a critical role in determining whether startups can scale beyond the laboratory.
And yet, the geopolitical context is beginning to shift the equation. Export controls, supply chain vulnerabilities, and concerns over technological dependence are driving a growing emphasis on “sovereign compute” across Europe. Governments and investors are increasingly aligned in their desire to build domestic capacity in critical technologies, including AI infrastructure.
This alignment may prove decisive because, for the first time in years, Europe’s semiconductor ambitions are being framed not just in economic terms, but as a matter of strategic autonomy. That framing has the potential to unlock policy support, capital, and market access in ways that were previously unavailable.
“It’s no longer a niche bet. It’s becoming a core part of how people think about AI infrastructure,” said Carlos Espinal of Seedcamp.
