The hardware behind the future


Euclyd Raises 200 Million Euros to Attack the AI Efficiency Wall, With Samsung on Board

The Dutch startup has no product yet, an ex-ASML chief as chairman, and a plan built around memory rather than raw compute.

Euclyd Raises 200 Million Euros to Attack the AI Efficiency Wall, With Samsung on Board

Euclyd, a chip startup in Eindhoven, has closed a Series A of more than 200 million euros, a sum CNBC puts at about $231 million, and the investor list is worth a look. Samsung co-led the round with Somerset Capital Partners, the EQT-managed Scaleup Europe Fund and Innovation Industries, and Peter Wennink, the former chief executive of ASML, joins as chairman of the board.

If you have never heard of the company, that is fair. CNBC reports it was founded in 2024, and its own release says it started at High Tech Campus Eindhoven under founders Bernardo Kastrup and Atul Sinha. Nothing is shipping yet, but the plan is specific enough to be worth explaining.

The target is inference, which is the part of AI where a trained model actually answers a question, as opposed to training, which is the long cram session that comes first. Euclyd told CNBC it is designing a chip system for inference with an architecture unlike a GPU, covering both the processor and the way memory is arranged around it.

Engineers talk about a memory wall, meaning the gap between how fast a processor can compute and how fast memory can feed it. Picture a chef who can chop faster than anyone can carry vegetables in from the fridge. Euclyd's release says its platform pairs programmable ASIC compute with processor-memory co-design and system-level optimization to get past that wall. An ASIC, pronounced "ay-sick", is a chip built for one job rather than every job.

Two products are named in the release: a chip called craftwerk, which Euclyd describes as the world's first agentic AI silicon, and a rack-scale system called craftwerk station, which it calls the world's lowest-power exascale AI factory. Exascale means a billion billion operations a second. Both superlatives are the company's own, with no benchmark, methodology or third party offered for either, and Kastrup told CNBC the first physical systems would roll out in 2028, with thousands of enterprise customers served by 2030.

"AI is becoming a foundation of economic growth, scientific discovery, and national competitiveness, but its potential will remain constrained unless we fundamentally change the infrastructure beneath it," Kastrup said in the release.

The business has two lanes. Euclyd told CNBC it plans to sell hardware and full rack systems to enterprises that want to run inference on their own premises, and to license its intellectual property to companies that want to build chips on Euclyd's designs. The release says the money goes to engineers, the silicon roadmap and readiness for enterprise, sovereign and hyperscale customers.

Samsung's role deserves a second look, because this is a startup whose pitch rests on memory. "Samsung can help us in more ways than money," Kastrup told CNBC. "They are one of the biggest memory manufacturers in the world. They do a lot of engineering, they know a lot about systems, they know the supply chain, they have a huge network."

The field is crowded: CNBC notes that Google, AWS and Meta are all developing their own AI chips, and that Nvidia holds a near monopoly on the highest-end parts. Euclyd's bet is that a chip drawn from scratch for inference, with memory as a first-class citizen, could beat general-purpose accelerators on cost per token and on watts. CNBC's caveat is the one I would underline: the systems have yet to be proven at scale in commercial deployments. Until 2028, this is 200 million euros of conviction, an ex-ASML chairman, and a very good set of drawings.

Leave a Reply