The Eindhoven startup’s round is large enough to be notable; the identity of who co-led it is the structural signal worth reading carefully.
What Happened
Bloomberg reports that Euclyd, an Eindhoven-based semiconductor systems company roughly two years old, has closed a Series A of more than €200 million — reported at approximately $231 million. The round was co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund, and Innovation Industries; individual investor amounts were not disclosed. One outlet described it as the largest European AI-inference chip round of 2026, though that ranking has not been independently verified.
Euclyd is building custom compute silicon, memory architecture, and data-centre systems designed together rather than assembled from commodity parts. Its chip, CRAFTWERK, packs 16,384 custom processors that work directly on data held in memory. The company’s argument, as reported, is that GPUs spend most of their energy moving data between memory and compute cores — that characterisation is Euclyd’s own, as reported, with no published percentage attached — and that designing the memory and compute layers together can reduce cost and energy consumed per token.
Peter Wennink, the former president and chief executive of ASML, has joined Euclyd’s board as chairman — in a personal capacity, not as a representative of his former employer. What is missing is not specifications but independent verification. Euclyd has published headline figures since unveiling CRAFTWERK in October 2025 — 8 PB/s of UBM bandwidth, FP4 compute, and roughly 7.68 million tokens per second at about 125 kW for its CWS 32 system — and it has said it is targeting systems rolling out in 2028 and thousands of enterprise customers by 2030. None of those figures has been independently tested, no named customer has been disclosed, and the round’s other participants include EIFO, imec.xpand, BOM and Quadri.
The key insight: The most informative detail in this round is not the size — it is that a memory manufacturer co-led it. Samsung’s position in the syndicate is a structural signal about where the memory industry believes the centre of gravity in AI compute is moving.
The Structural Read
The correct frame for understanding Euclyd is not “another GPU challenger.” It is a company that has identified that inference is a structurally different problem from training — not a smaller version of it — and is betting that difference creates an opening the incumbent’s architecture cannot close by iteration alone.
Training is capacity-bound. It rewards raw arithmetic throughput on dense matrix work with high data reuse. The modern accelerator and the cluster economics built around it were optimised precisely for this workload, and the incumbent has already won that contest on its own terms. GPUs remain extraordinarily good at what they were built to do — that is not in question.
Serving a trained model is a differently shaped problem. Weights have to be streamed past the compute units. Arithmetic intensity is low. The binding constraint becomes memory bandwidth, not floating-point capability. Every challenger that has attacked the GPU on FLOPs has fought on the ground the incumbent selected, and lost accordingly. Attacking on bytes moved — the cost of data movement rather than the count of multiplications — contests a different metric on a workload that is becoming the majority of the industry’s compute.
Map of AI — Capacity-Bound vs. Bandwidth-Bound
The basis of competition has changed layers
The AI stack’s competitive bottleneck is not fixed. Training made FLOPs the scarce input; inference makes bytes moved the scarce input. A company that designs silicon, memory architecture, and data-centre systems as a single integrated stack is competing on the metric that matters for the workload that is now the majority. That is the structural opening — not a marketing claim about being better at the same thing, but a claim about competing on a different dimension entirely. Whether Euclyd has actually solved it is a question its own published figures cannot settle, because nobody outside the company has tested them.
The Samsung signature is the most informative thing in the round. Under the prevailing architecture, memory vendors are suppliers to the compute layer: high-bandwidth memory is sold into accelerator packages, and margin concentrates with whoever integrates the system rather than with whoever supplies the parts. An architecture that computes directly on data held in memory would relocate the centre of gravity toward memory itself.
On that reading, a memory manufacturer co-leading a round for in-memory compute is the memory industry buying an option on becoming a principal rather than remaining a supplier. This is a structural observation about the position Samsung occupies in the AI stack — not a claim about its stated intentions, which have not been reported, and not an assertion that the company has announced any such strategy.
Peter Wennink’s appointment points at a different constraint entirely. His value to a two-year-old company is not technical. European deep-tech hardware rarely fails for want of good engineering; it routinely fails for want of standing with fabs, packaging partners, supply chains, and sovereign funders — all relationship businesses with long memories. Recruiting the former chief executive of ASML as chairman is a statement that the binding constraint is credibility and access, not design talent. His appointment is in a personal capacity; nothing here asserts any involvement, relationship, or endorsement on ASML’s part, and nothing claims he will secure any particular outcome.
Three Implications
IMPLICATION 1 — The FLOPs attack always fails; the bytes attack is different
Every prior GPU challenger competed on arithmetic throughput — the metric the incumbent defined and dominates. Euclyd’s thesis is that inference’s binding constraint is memory bandwidth, not FLOPs. That is a different competitive axis, not a marginal improvement on the same one. Whether the architecture survives contact with real models, real toolchains, and the accumulated software ecosystem that makes a general-purpose accelerator usable by ordinary engineers is a separate and entirely open question — but the framing is structurally distinct from prior attempts.
IMPLICATION 2 — Memory vendors are buying options on moving up the stack
In the current architecture, memory is a component sold into accelerator packages; value capture concentrates with the system integrator. In-memory compute, if it works at scale, would restructure that relationship — memory becomes the compute substrate, not the commodity input. Samsung co-leading this round is best read as a strategic hedge: a memory company investing in an architecture where memory companies become principals. Whether that restructuring materialises is unknown; what is clear is that Samsung is paying to have a seat at the table if it does.
IMPLICATION 3 — Credibility and access are the European deep-tech constraint, and the round addresses them directly
Capital reduces execution risk. Peter Wennink as chairman addresses the constraint that has historically killed European hardware ventures before capital: access to fabs, packaging partners, sovereign funding relationships, and supply chains that require standing built over decades. The round’s structure — a distinguished chairman, pan-European institutional co-investors alongside a global memory giant — is an attempt to solve the credibility problem in advance rather than after first silicon. It does not reduce architectural risk, which remains entirely unaddressed by anything in the public record.
The Bottom Line
Euclyd is a well-funded and unusually well-timed bet on a real structural shift — the migration of the industry’s dominant workload from capacity-bound training to bandwidth-bound inference. The architecture is unverified rather than undisclosed: the company has published headline performance figures and a roadmap pointing at systems in 2028, but none of it has been independently tested, no customer has been named, and in-memory compute has a long and instructive history of arguments that read cleanly on a slide and then did not survive contact with real models, real toolchains, and the software ecosystem that makes a general-purpose accelerator usable at scale. Capital and a distinguished chairman reduce execution risk; they do nothing to reduce architectural risk. What the round establishes is not that the shift has been captured, but that serious capital — including a memory giant with structural reasons to care — believes the shift is real enough to fund at this size. That is the honest read, and it is the only one the available evidence supports.
Sources: Bloomberg — Chip Startup Euclyd, Backed by Ex-ASML CEO, Raises More
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity. Euclyd has published headline performance figures for CRAFTWERK since unveiling it in October 2025, including 8 PB/s of UBM bandwidth, FP4 compute and roughly 7.68 million tokens per second at about 125 kW for its CWS 32 system, and has stated a roadmap targeting systems rolling out in 2028 and thousands of enterprise customers by 2030. None of these figures has been independently verified, no customer has been named, and nothing in this article should be read as evidence that the approach works at scale. The distinction throughout is between claims that are undisclosed and claims that are unverified: these are the latter. The characterisation that GPUs spend most of their energy moving data between memory and compute cores is the company’s, as reported. No percentage has been published, none is asserted here, and it is not presented as a measured finding. Individual investor amounts in the round were not disclosed and none is stated here; participants beyond the co-leads include EIFO, imec.xpand, BOM and Quadri. No quotation is attributed to Peter Wennink, any founder, Samsung or any investor anywhere in this article. Peter Wennink has joined Euclyd’s board as chairman in an individual capacity. Nothing here asserts any involvement, relationship, endorsement or investment on the part of ASML, and nothing claims he will secure any particular outcome for the company. The reading of Samsung’s participation as the memory industry taking a position on in-memory compute is a structural observation about the layer Samsung occupies. Samsung’s intentions have not been reported, and no announced strategy is attributed to it. The compute-mix percentages are an analyst estimate of Anthropic and OpenAI’s combined workloads. Neither company publishes workload splits, so these are modelled rather than disclosed figures, the final two quarters are explicitly estimates, and while the direction is corroborated by Epoch AI’s work on reinforcement learning taking a rising share of frontier training compute, the specific percentages represent a single model and should not be treated as measurement. Nothing here predicts that Euclyd succeeds, that any incumbent loses share, or that GPUs are displaced from any workload; general-purpose accelerators remain extremely capable at the tasks they were designed for. No market-size, revenue, share or share-price claim is made. Euclyd is a private company; Samsung, ASML and NVIDIA are publicly listed. This is business analysis, not investment advice, no view is expressed on any security, and no recommendation is made.More Resources
