Analysis
Euclyd, a Dutch startup building chips for AI inference, has raised more than €200 million (about $231 million) in a Series A led by Samsung, with Somerset Capital Partners, EQT's Scaleup Europe Fund and Innovation Industries co-leading, Bloomberg reported. Former ASML chief executive Peter Wennink is joining as chairman.
Founded in 2024 and based in Eindhoven -- the same city that hosts ASML's headquarters -- Euclyd employs roughly 81 people as of January 2026, per CNBC's earlier reporting on the company's fundraising ambitions. Founders Anuj Gupta and Raghvendra Kushwah, who serves as chief executive, previously raised a seed round of under €10 million; the jump to $231 million in roughly a year is one of the largest Series A rounds in European deep tech this year.
The Pitch: Memory, Not More Transistors
Euclyd's argument is architectural rather than incremental. Rather than chasing Nvidia on raw transistor density, the company has built what it describes as a new memory-centric compute architecture for inference -- moving data less, since data movement, not computation, is what burns most of the power in a large model's forward pass. Euclyd claims up to 100x higher power efficiency for inference compared with Nvidia's upcoming Vera Rubin generation, a figure that has not been independently benchmarked and should be read as a vendor claim rather than a verified result.
The competitive field is crowded and well-funded. Groq, founded in 2016 and based in Mountain View, has raised hundreds of millions building Language Processing Units purpose-built for inference. Cerebras, with more than 400 employees, builds wafer-scale chips and priced its long-delayed IPO earlier this year. d-Matrix is pursuing a similar in-memory-compute thesis with its own inference silicon. Every one of them is chasing the same gap: Nvidia's GPUs were designed for training, and inference -- now the larger and faster-growing workload as AI moves from research labs into products -- rewards different tradeoffs. Pulse has tracked Nvidia's pricing power as the central fact every inference challenger has to underwrite around.
Samsung's participation is the detail worth reading closest. As one of the world's two largest memory manufacturers, Samsung is not a passive financial investor in a memory-centric chip architecture -- it is a potential manufacturing and supply partner, and its presence in the round is a stronger signal than the dollar amount alone. EIFO (Denmark's export-investment fund), imec.xpand and Brabant Development Agency rounded out the syndicate, giving Euclyd a mix of strategic, sovereign and regional capital uncommon even by European deep-tech standards.
The Bear Case
The risk is the one every fabless inference startup carries: claims made in a press release are not silicon shipped to customers. Euclyd has not disclosed a tape-out date, a named anchor customer, or independent benchmark results, and 100x efficiency claims have a poor track record of surviving contact with production workloads once real customers start comparing total cost of ownership rather than a single power metric. Wennink's chairmanship buys credibility and manufacturing relationships; it does not buy silicon that works at scale.
For anyone diligencing the inference-chip category, the number to ask for is not the funding total but the signed pilot commitment -- how many FLOPs of real customer workload has Euclyd actually contracted to run, and on what timeline. Every well-funded Nvidia challenger of the last three years has cleared the funding bar. Few have cleared the deployment one.
