Samsung Backs Dutch AI Chipmaker Euclyd in $231 Million Round to Build Non-GPU Inference Silicon
Euclyd, a Dutch AI chipmaker founded in 2024, has closed a 200 million euro Series A — approximately $231 million — with Samsung participating alongside co-leads Somerset Capital Partners, the EQT-managed Scaleup Europe Fund, and Innovation Industries. CEO Bernardo Kastrup announced the raise exclusively to CNBC on September 14.
The company is building an AI chip system for inference that uses a different architecture from GPUs. Unlike most inference accelerator startups, Euclyd is designing both the processor and the memory architecture together, targeting the paired bottleneck of compute throughput and memory bandwidth that limits current inference at scale. The company says its silicon will reduce energy consumption and cost for data centers running foundational models.
What Euclyd Is Not
Euclyd is not building a training chip. The distinction matters. The GPU alternative landscape has split into two camps: companies targeting training (where Nvidia’s H100/H200/B100 series dominate) and companies targeting inference (where the incumbent advantage is softer and the cost sensitivity of hyperscalers is highest). Euclyd is firmly in the second camp.
Samsung’s participation signals both strategic interest and a stake in diversifying the AI accelerator supply chain beyond Nvidia. Samsung manufactures memory and has its own semiconductor operations; backing a company that is co-designing processor and memory architecture from the ground up represents a defensive hedge as well as a commercial bet.
Market Context
Nvidia commands near-monopoly pricing on high-end AI compute. Hyperscalers — Amazon, Google, Meta, Microsoft — are all running internal chip programs to reduce that dependency. The gap they leave is in the mid-tier: companies that want inference capacity without building their own silicon and without paying Nvidia’s full rack pricing.
Cornelis Networks raised $205 million the same week on a networking fabric approach to the same problem: reduce the time GPUs spend waiting for data. Euclyd’s approach is more fundamental, targeting the chip itself rather than the interconnect. Both bets are wagers that there is a large addressable market in the space between Nvidia’s stack and hyperscaler custom silicon.
Euclyd’s chips are not yet in production. The Series A proceeds presumably fund tape-out, lab testing, and the first commercial customer engagements. At 2024-founded, the company is still years from displacing anyone’s infrastructure — but at this funding level, it has enough runway to prove the architecture.