Dutch semiconductor startup Euclyd has closed a $230M+ Series A round to fund craftwerk, an inference-focused ASIC designed to power AI agent workloads. Backed by Samsung and a former ASML CEO, the company aims to break through the « memory wall » that bottlenecks today’s GPUs.
🔑 Key takeaways
- Euclyd raised over $230M in Series A funding, led by Samsung Electronics, Somerset Capital Partners, and the Scaleup Europe Fund
- The startup is developing craftwerk, an ASIC chip optimized for AI agent inference
- Peter Wennink, former ASML CEO, has joined the board of directors
- Co-designed processor-memory architecture aims to bypass the limits of standard HBM
- Silicon launch targeted for 2028, with thousands of enterprise customers envisioned by 2030
A round led by industry heavyweights
Euclyd announced on September 15, 2026 a Series A funding round exceeding €200 million, or roughly $230 million. The deal was led by three top-tier investors: Samsung Electronics, Somerset Capital Partners, and the Scaleup Europe Fund. Other institutional players also participated, including imec.xpand, the corporate venture arm of Belgium-based nanotechnology research center Imec.
The addition of Peter Wennink, former CEO of ASML Holdings NV, to the board of directors sends a strong signal. Euclyd is headquartered at the High Tech Campus Eindhoven, just a few kilometers from ASML’s headquarters, reinforcing the startup’s roots in Europe’s semiconductor ecosystem. This proximity is no coincidence: Euclyd’s memory strategy sits in the same value chain that made ASML the world leader in lithography equipment.
| Investor | Type | Role in the round |
|---|---|---|
| Samsung Electronics | Strategic | Lead, HBM memory expertise |
| Somerset Capital Partners | Financial | Lead |
| Scaleup Europe Fund | Public/Institutional | Lead |
| imec.xpand | Corporate VC | Participant |
craftwerk, a chip purpose-built for AI agent inference
Euclyd’s flagship product is an AI chip called craftwerk. It will ship inside a system named CWS, designed to power AI clusters exceeding the exaflop threshold (more than one quintillion operations per second). Unlike most general-purpose GPUs on the market, craftwerk is not built for model training, but for inference, the execution phase once a model is deployed.
This focus on AI agents means the chip will likely include CPU cores alongside inference accelerators. Agents need to orchestrate tools, run conditional logic, and handle network calls, tasks that a pure GPU handles poorly. The compute module of craftwerk will be an application-specific integrated circuit (ASIC), meaning silicon custom-designed for a precise workload, rather than a general-purpose programmable GPU.
« We are designing craftwerk for a future where advanced AI is no longer constrained by infrastructure cost, energy availability, or geography. »
Euclyd, official statement
The « memory wall, » the bottleneck Euclyd wants to break
The « memory wall » refers to the bottleneck that throttles AI chip performance by limiting data transfer throughput between compute units and HBM (High Bandwidth Memory) modules. A standard HBM module stacks multiple layers of memory chips on a base die that interfaces with the host GPU. As models grow, this memory corridor becomes the limiting factor.
Euclyd plans to tackle this challenge through « processor-memory co-design »: the memory architecture will be optimized to work in concert with the in-house ASIC, rather than relying on off-the-shelf standardized HBM. The company has not disclosed the exact nature of this memory, but the indication points to a proprietary design drawing on recent industry work.
| Architecture | Vendor | Type | Key advantage |
|---|---|---|---|
| Standard HBM | Multiple (Samsung, SK Hynix, Micron) | Stacked memory | GPU industry baseline |
| NVHBM | Nvidia | Modified HBM, relocated circuits | Up to 30% more GPU die area freed |
| On-chip SRAM | d-Matrix | Integrated high-speed memory | Energy-efficient inference |
| ASIC+memory co-design | Euclyd (craftwerk) | Proprietary architecture | AI-agent optimized, exaflop+ scale |
Samsung, a strategic industrial partner
Samsung’s presence as a lead investor is far from symbolic. The Korean group is one of the world’s largest HBM memory suppliers, alongside SK Hynix and Micron. In August 2026, Samsung unveiled a technology that places HBM memory directly on top of a GPU’s compute circuits, while current designs place memory dies and compute modules side by side. This 3D approach could complement Euclyd’s memory co-design and offer a unique industrial lever.
Nvidia, for its part, had already unveiled its NVHBM architecture in August 2026, which relocates certain circuits from the base die to the overlying HBM stack and reduces the physical channels routing data through the base die. Nvidia says these changes free up to 30% of GPU surface area, allowing more compute circuits to be packed in. Startup d-Matrix relies on SRAM to perform some inference operations directly in memory. Euclyd is positioned in the same race for energy efficiency, but with an explicit focus on agents rather than general-purpose LLMs.
Go-to-market and roadmap
According to information reported by CNBC, Euclyd plans to launch its silicon in 2028. The company intends to sell hardware directly to data center operators and license its technology to other chipmakers. The stated goal is to attract thousands of enterprise customers by 2030, a tight timeline in an industry where ASIC design typically takes three to five years from architecture freeze to volume production.
The commercial bet is twofold: prove that a proprietary memory architecture can compete with the mature HBM ecosystem, and convince cloud operators that « agent » workloads justify dedicated silicon rather than a general-purpose Nvidia GPU. Peter Wennink’s arrival on the board reinforces the industrial credibility of the pitch, at a time when Europe is seeking to densify its sovereignty in advanced semiconductors.
Conclusion
With over $230M raised, a Samsung partnership, and a former ASML CEO on its board, Euclyd has serious firepower to attack the inference chip market. If craftwerk delivers on its energy-efficiency and memory-throughput promises, the startup could emerge as a credible European challenger to Nvidia and other specialized players. The main risk remains timing: hitting a 2028 silicon launch in a sector where every delay is paid in lost market share.
Sources
- SiliconANGLE — Euclyd raises $230M+ to develop chips for AI agents
- Aventure VC — Euclyd raises $230M+ to develop chips for AI agents
This article is published for informational and educational purposes only. It does not constitute investment advice. Do your own research (DYOR) before making any decision.
