Bloomberg: Dutch chip startup Euclyd is seeking about 200 million euros for a Series A, not yet closed

Why it matters

A 200 million euro Series A would put substantial private capital behind a Dutch semiconductor startup trying to turn company-claimed AI inference efficiency projections into silicon. But Bloomberg says the round is not closed, the amount could change, and Euclyd still has to prove its hardware in the market.

Bloomberg reported on June 24, 2026 that Dutch chip startup Euclyd is seeking about 200 million euros ($229 million) in Series A financing. The company has received several term sheets, according to the report by Yazhou Sun and Sarah Jacob, but the round is not finalized and the amount could change.

That distinction is the story. Euclyd has not announced that it raised 200 million euros, Bloomberg did not name a lead investor, and no valuation was disclosed. What Bloomberg did verify is narrower and still material: Euclyd is a Dutch chip startup backed by former ASML Holding NV Chief Executive Officer Peter Wennink, and it is in the market for a very large Series A by early-stage hardware standards.

Euclyd's own materials fill in the company's self-description. On its homepage, Euclyd says it is building "ultra-efficient silicon systems for foundation AI models, including large language models," and lists the company as headquartered in Eindhoven with offices in San Jose, California. The same page identifies Bernardo Kastrup as founder and CEO, Gerard Egelmeers as co-founder and VP of processor firmware co-design, Ingolf Held as co-founder and VP of product, and Harm Peters as co-founder and VP of silicon design engineering. Those are company-disclosed roles, not Bloomberg's reporting on the financing.

Euclyd's public positioning is not that it is selling another software layer around frontier models. The company is presenting itself as a semiconductor startup aimed at the expensive part of AI deployment: inference hardware, memory movement, packaging and system-level efficiency. That is why the proposed financing matters. Custom silicon consumes capital before revenue, and a term sheet for a chip company is not the same thing as a shipped device, a qualified manufacturing flow or a customer-validated benchmark.

The Wennink signal

Wennink's name is the establishment signal attached to the financing report. Bloomberg described Euclyd as backed by the former ASML CEO. Euclyd's homepage goes further, listing Peter Wennink, Federico Faggin and Steven Schuurman as "Investor and Advisor," while Euclyd's CRAFTWERK announcement says the company is "mentored and backed" by those names, according to Euclyd's own product release. The company does not disclose on those pages how much any of them invested, whether they are participating in the Series A Bloomberg described, or whether any holds a board seat.

ASML's relevance is not that Euclyd is making lithography equipment. It is that ASML sits at the center of the global semiconductor supply chain, and Wennink's post-ASML backing gives the startup a shorthand credibility marker in a market where credibility is usually earned slowly. ASML said its April 24, 2024 annual general meeting marked the end of Peter Wennink's and Martin van den Brink's terms as presidents of ASML and the beginning of Christophe Fouquet's term as president and CEO, according to ASML's AGM results.

Euclyd has also put Wennink into its own product narrative. In Euclyd's CRAFTWERK announcement, the company quotes Wennink as saying: "I believe AI inference will dominate datacenter silicon." That sentence should be read as part of Euclyd's announcement, not as an independently reported interview.

The financing logic follows from that inference thesis. Training captured the first AI infrastructure cycle because foundation models required enormous clusters to be built. Inference is the repeated-use market: every query, agent workflow and enterprise deployment has to run again and again. If the cost per token and watts per token remain high, AI adoption becomes a margin problem for cloud providers, model companies and enterprises. Euclyd is positioning its hardware directly at that pressure point.

The product claim is large, and still a claim

Euclyd's named product is CRAFTWERK. In its product release, Euclyd says CRAFTWERK is "currently in advanced design" and describes it as an inference architecture for agentic AI. The company says the system-in-package includes 16,384 custom SIMD processors, up to 8 PFLOPS at FP16 or 32 PFLOPS at FP4, 1 TB of custom ultra-bandwidth memory and 8,000 TB/s of bandwidth, according to Euclyd's CRAFTWERK announcement.

Euclyd also projects that its rack-scale CRAFTWERK STATION CWS 32 would deliver 7.68 million tokens per second at 125 kW in a multi-user scenario. The same release says the company's 100x improvement claim in power efficiency and cost per token is based on modeled performance for Llama 4 Maverick, according to Euclyd.

Those numbers are company claims and projections, not independent benchmarks. Euclyd has not disclosed customer-validated performance data in the materials cited here. The right way to read Bloomberg's 200 million euro Series A target is as a bid to finance the conversion of an ambitious architecture into silicon and systems, not as proof that the claimed performance has already been achieved in the field.

The manufacturing path Euclyd has disclosed

Euclyd has announced at least one implementation partner. On Nov. 3, 2025, the company said ADTechnology had entered a development partnership with Euclyd to bring CRAFTWERK to silicon. Euclyd's announcement says ADTechnology will act as a "key implementation partner" responsible for "backend design and coordination of the manufacturing process" under "Samsung Foundry's FinFET process technologies."

That partnership matters because chip startups do not simply write code, push it to production and iterate overnight. They need design implementation, verification, packaging decisions, manufacturing coordination and eventually hardware that customers can test. Euclyd's ADTechnology announcement is therefore a concrete step in the product story, but it is not the same as volume production or commercial traction.

The same distinction applies to the fundraising. Several term sheets, if Bloomberg's report holds, are evidence of investor interest. A closed round would be a different fact. A closed round plus working silicon, validated performance and paying customers would be different again.

The missing terms are still material

Bloomberg's report leaves the central financing terms undisclosed. It does not name the lead investor, participating funds, valuation, expected close date, existing capital raised, customer commitments or manufacturing schedule. Euclyd's own pages identify executives, advisors, a product name and an implementation partnership, but they do not disclose pricing, revenue, signed customers or how much capital its named backers have committed.

For a software startup, some of those omissions would be ordinary. For a semiconductor company seeking about 200 million euros at Series A, they define the risk. Silicon programs can absorb large sums before first revenue, and design misses may appear late, after expensive work has already been committed. The company's challenge is not merely to persuade investors that inference is a large market. It has to prove that its architecture can be manufactured, packaged, powered, cooled, programmed and sold into real AI workloads.

Euclyd's timing is rational. Inference cost has become one of the central constraints in AI deployment, and Euclyd's own pitch is aimed squarely at that constraint. The company offers investors an easy narrative: an Eindhoven-based chip startup, a founder-led team disclosed on its homepage, backing from a former ASML chief, a named AI inference architecture, and a Samsung Foundry implementation path through ADTechnology.

But the burden is higher than the narrative. If Euclyd closes a 200 million euro Series A, the money will not simply validate its pitch. It will buy the company time to prove whether CRAFTWERK can move from company-modeled performance and partnership announcements into shipped hardware that customers can measure. In chips, that is the benchmark that matters.