EUCLYD raises €200M+ Series A to build AI chips that make intelligence cheaper

EUCLYD (Eindhoven) has raised over €200 million in a Series A co-led by Samsung, Somerset Capital Partners, the Scaleup Europe Fund (managed by EQT), and Innovation Industries, to build AI silicon, memory architecture, and data-centre systems designed to reduce the cost per token of running foundation models at scale. Additional backing came from EIFO (Denmark's export and investment fund), imec.xpand, the Brabant Development Agency, and Quadri. Peter Wennink — President and CEO of ASML for seven years until 2024 — has joined the board as Chairman.

Founded in 2024 by Bernardo Kastrup (CEO) and Atul Sinha at High Tech Campus Eindhoven, the company makes the case that AI's practical ceiling is not model architecture but infrastructure cost. The round is among the largest early-stage raises in European deep tech in 2026.

The efficiency thesis

The argument EUCLYD is making is structurally simple: the silicon underneath foundation models was designed for general-purpose compute, not for the specific memory-access patterns that transformer inference requires. Memory bandwidth is the dominant bottleneck — every forward pass fetches model weights from memory at every layer, and the energy and latency of those transfers dominate the cost curve. The result is that most of the compute budget in a large inference cluster goes to data movement, not to arithmetic.

EUCLYD is building two things in parallel. craftwerk, its custom ASIC, is designed from the ground up around transformer inference patterns. CWS (craftwerk station) packages that silicon with EUCLYD's own memory architecture and data-centre systems layer into a full-stack solution. The claimed output is a meaningful improvement in energy-per-token and cost-per-token relative to general-purpose accelerators — numbers the company has not publicly disclosed but which are the core of every enterprise and government conversation it is having. Target customers span hyperscalers, government agencies, and enterprise AI operations that cannot or will not use US-controlled platforms.

CEO Bernardo Kastrup: "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."

The Eindhoven advantage and what Samsung signals

The High Tech Campus location is not incidental. The Netherlands has an unusual concentration of semiconductor-adjacent talent and supply-chain depth — ASML, NXP, and imec's Dutch operations are all within proximity — and EUCLYD benefits from all three. Wennink's chairmanship converts that geographic adjacency into direct access to the ecosystem he helped build at ASML over two decades.

Samsung's co-lead position carries independent meaning. Beyond the capital, it signals potential procurement and integration pathways into Samsung's own AI infrastructure stack, including its data-centre memory products (HBM and related). For a hardware company at pre-revenue stage, a strategic co-lead with manufacturing scale and a live AI infrastructure business is a materially different signal from a pure financial investor.

EIFO's participation is notable: the Danish fund typically backs companies with credible export potential and technology differentiation, not early hardware bets. Its presence suggests EUCLYD has made the technical case beyond the Eindhoven ecosystem.

The round is large enough to be credible for the task — a full custom ASIC tape-out, qualification, and ramp cycle typically runs into hundreds of millions of dollars — but not so large that it solves the problem by itself. EUCLYD's 18-month question is whether purpose-built European AI silicon can reach production readiness before Nvidia, Google, AWS, and Groq narrow the efficiency gap further from their own angles. The company's sovereign architecture thesis — European design, European data governance — is a genuine differentiator for government and regulated-enterprise customers that US hyperscalers structurally cannot serve. Whether that differentiator is large enough, and arrives early enough, is the open bet.