Product2 publishers2 min readPublished
Euclyd raises $230M for a 2028 agent chip built around its memory architecture
The Dutch startup says craftwerk is designed to run AI agents, and the design details it has published so far are about ASICs and processor-memory co-design. Samsung, one of the biggest HBM suppliers, co-led the round.
The Product Desk · Product desk
What happened
- Euclyd BV, a chipmaker based at High Tech Campus Eindhoven, said it raised more than 200 million euros, about $230 million, in a Series A led by Samsung Electronics, Somerset Capital Partners and the Scaleup Europe Fund.
- Peter Wennink, the former chief executive of ASML Holdings, joined Euclyd's board as part of the raise, and imec.xpand was among the other institutional backers.
- The chip, called craftwerk, will ship inside a system called CWS that Euclyd expects to power AI clusters with more than an exaflop of performance.
- Euclyd says craftwerk's computing module will be an ASIC with an "innovative memory architecture," and that it will pursue processor-memory co-design to attack the memory wall.
- CNBC reported that Euclyd plans to launch the silicon in 2028 and hopes to gain thousands of enterprise customers by 2030.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- contradiction Euclyd's public claim is the workload label; the inference-not-training reading and the CPU cores come from SiliconANGLE. Anyone drawing up a shortlist of agent silicon would be setting a publisher's inference against other vendors' spec sheets.
- exposure Samsung sells HBM and has now helped fund a company pursuing processor-memory co-design, part of a group of custom designs whose stated aim is to replace the standard HBM shipping with GPUs today.
- decision Euclyd intends to license the underlying technologies to other chipmakers. A buyer tracking this would be evaluating a supplier's supplier, not a vendor it purchases from directly.
- capability If the co-design delivers on memory bandwidth, the benefit lands on any inference workload running on the part, so the agent label narrows only the marketing.
Euclyd's own description of the workload is one phrase: craftwerk is "designed to run AI agents" [8]. The rest of the compute story in SiliconANGLE's account is the publisher's own reading, and a fair one. The site wrote that the phrasing "suggests the chip will include circuits optimized for inference rather than training." It added that either craftwerk or the CWS system around it is likely to include CPU cores, because agents need cores to run their tools as well as accelerators for inference [9][10].
The design work Euclyd did describe is memory work, and nothing in it is agent-specific. SiliconANGLE sets it beside two other custom memory efforts. Nvidia's NVHBM, shown last month, moves circuits off the HBM base die into the stack above it and narrows the channels carrying data through the base die; Nvidia says that frees up 30% of a GPU's surface area for more computing circuits [15][16]. d-Matrix does some inference calculations inside SRAM and passes the rest to its own cores, and says its chips are more power-efficient than traditional GPUs [17][18]. Neither is described as an agent chip; both target the memory wall [13][14].
Samsung's own hardware points the same direction. It is one of the world's top HBM suppliers [20]. Last month it debuted a way to put HBM directly on top of a GPU's computing circuits, where designers currently place memory dies and logic side by side [21]. The money went to a company promising processor-memory co-design [12].
Euclyd frames the goal in power and cost terms, saying craftwerk is designed for a "future in which advanced AI is no longer constrained by infrastructure cost, power availability or geography" [19]. SiliconANGLE reported the round on September 15, 2026, which leaves roughly two years between the money and first silicon and about four to the enterprise customer target [1][25].
For the person picking infrastructure for agents now, what separates the pitch from the work is which part of the agent loop the silicon touches: the model's forward pass, or the tool calls and the orchestration between them. The other is whether that answer sits in a spec or in a reporter's inference. On craftwerk, the first answer is not public yet, and the second answer is a reporter's inference [9][10].
What to watch
- Whether Euclyd publishes a spec naming general-purpose cores on craftwerk or inside the CWS system.
- Whether Samsung's HBM-on-logic stacking technology turns up in CWS, or Euclyd's co-design goes to another memory supplier.
- Whether any chipmaker is named as a licensee of the underlying technology before the 2028 launch date.