EUCLYD, a 2024 startup based at High Tech Campus Eindhoven, said on Sept. 15 that it had signed a Series A financing round of more than €200 million to build ultra-efficient infrastructure for foundation AI models. The round was co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund and Innovation Industries, with EIFO, imec.xpand, Brabant Development Agency and Quadri also participating. Former ASML president and CEO Peter Wennink is joining as chairman.
That is a large early-stage vote of confidence in a part of the AI stack that is starting to matter more than training bragging rights: inference, the repeated work of serving models to users and software agents at scale. The practical question for customers, though, is not whether EUCLYD can raise money. It is whether a well-funded processor-memory design can become a deployable system with measurable advantages over the GPU platforms that already dominate data centers.
Why this round matters now
AI infrastructure economics are shifting. Training still commands attention, but inference is where operators can end up paying the bill every day. Each prompt, tool call and generated token keeps moving model weights and activations between compute and memory. That movement is often the real bottleneck. Chips can sit underused while memory bandwidth, interconnects, rack power or cooling become the limiting factor.
That is the opening EUCLYD is trying to exploit. In its Sept. 15 press release, the company positioned itself not as a cloud provider but as an inference-infrastructure company building around custom silicon and a rack-scale system. The pitch is straightforward: if AI inference is increasingly constrained by moving data rather than doing math, a purpose-built architecture that redesigns compute and memory together may deliver more useful work per watt and per euro.
The investor list gives that idea strategic weight. Samsung’s presence is especially notable because memory, packaging and semiconductor manufacturing know-how are central to whether this kind of architecture works in practice. That does not, by itself, establish a manufacturing deal or preferential memory access. It does suggest that major industry players see the inference bottleneck as important enough to back aggressively.
Wennink’s appointment matters for a different reason. Deep-tech chip companies do not succeed on architecture alone. They need capital discipline, manufacturing realism, ecosystem partnerships and customer trust. A chairman with long experience in the semiconductor equipment industry does not remove execution risk, but it does signal that EUCLYD intends to operate like a serious infrastructure company, not just a lab project.
What EUCLYD says it is building
EUCLYD’s platform centers on craftwerk, which the company describes as “agentic-AI silicon,” and craftwerk station CWS, a rack-scale system. Its core technical idea is processor-memory co-design: treat compute, memory and the surrounding system as one optimization problem rather than bolt a faster accelerator onto a conventional server design.
In a September 2025 technical announcement, EUCLYD described craftwerk as a system-in-package with 16,384 custom SIMD processors, up to 8 petaflops of FP16 or 32 petaflops of FP4 compute, and 1 terabyte of custom ultra-bandwidth memory. It also outlined a CWS 32 system built from 32 such modules, with 1.024 exaflops of modeled FP4 performance and 32 terabytes of that memory. The same announcement projected 7.68 million tokens per second at 125 kilowatts on a multi-user Llama 4 Maverick workload, along with a 100x improvement in power efficiency and cost per token versus leading alternatives.
Those figures are why the company stands out. They describe an attempt to attack the memory wall at the package and rack levels, not merely to build another accelerator card. If the architecture works as intended, EUCLYD could appeal to hyperscalers, AI service providers, enterprise operators and even other semiconductor companies looking to license inference IP.
But the same numbers also define the current gap. They are modeled design-stage figures, not public production measurements. The supplied sources do not show taped-out silicon, benchmarked hardware in customer hands or a disclosed manufacturing plan. SiliconANGLE reported that EUCLYD aims to launch silicon in 2028, sell hardware to data-center operators and license the technology to other chipmakers. Those details fit the broader story, but they go beyond what the short primary release confirms.
What buyers need before this becomes a GPU alternative
A Series A of more than €200 million can finance the expensive middle stretch between architecture and product: engineering hires, tape-out preparation, packaging work, software tooling, validation, system integration and the long process of turning a chip into something a procurement team can actually buy and support. What it cannot do is shortcut proof.
For buyers, the relevant comparison is not a modeled headline number against an abstract “leading alternative.” It is a full rack-level result against a real GPU-based system running a specified workload. That means silicon-based benchmarks with the model named, context length disclosed, batch and concurrency documented, and latency distribution shown alongside throughput. It means power measured at the wall, not only at the chip. It means including the host CPU, networking and storage costs that agentic systems need to orchestrate tool use and workflows.
Software will be just as important as silicon. A custom ASIC can look compelling for a stable workload and still fail commercially if model coverage is narrow, compiler support lags or teams cannot port existing frameworks without painful rewrites. Customers will want to know how much of the stack is ready: kernels, compilers, model support, observability, virtualization, failure handling and the operational support needed for production data centers.
They will also be comparing EUCLYD against a moving target. GPU vendors are not standing still on memory architecture, packaging or system design. The competitive bar is not today’s accelerator; it is whatever mature GPU stacks can deliver by the time EUCLYD has shipping hardware. A custom design can lose its edge if model architectures change, if utilization turns out to be lower than expected or if supply and support prove harder than the raw math suggested.
That is the right way to read this financing. It is a strong capital-markets wager that inference has become one of AI infrastructure’s hardest and most valuable bottlenecks. It is not yet proof that EUCLYD has solved it. The company now has the backing to make its case at the level that matters most: measured, deployable economics in a real data-center environment.
By
By
By
By
By
