Nvidia turned a graphics processor into the engine of artificial intelligence. Its chips run the overwhelming majority of AI inference — the stage where a trained model answers a query instead of learning from data. Now a 19-person company in Eindhoven wants to do that job without a GPU at all.
On September 15, it raised more than €200 million. The company is Euclyd. Samsung co-led the round with Somerset Capital Partners, EQT's Scaleup Europe Fund and Innovation Industries. EIFO, imec.xpand, the Brabant Development Agency and Quadri also participated. At roughly $231 million, this is the largest European AI-inference chip round of 2026, and the biggest deeptech cheque the Scaleup Europe Fund has written.
The bet is co-design: build the processor and the memory as one system, rather than sourcing the parts separately.
Nothing is proven yet. The first commercial silicon is scheduled for 2028, and the company books no product revenue today.
Euclyd Series A, September 2026
Co-led by Samsung, EQT's Scaleup Europe Fund, Somerset Capital Partners and Innovation Industries. · Euclyd press release, 2026
What Euclyd is actually building
The company calls itself a semiconductor systems company, and the wording is deliberate. It is not selling a faster chip the way a challenger sells a faster car. It is rebuilding the vehicle around a different engine.
Two products sit on the roadmap. craftwerk is a programmable accelerator that the company describes as the first silicon designed for agentic AI — workloads where a model runs many steps, calls tools and holds state, rather than answering one prompt. CWS, or craftwerk station, is a rack-scale machine the company bills as the lowest-power exascale AI factory. Exascale means a computer capable of roughly a quintillion operations per second. The claim is less about peak speed than about how little electricity it takes to reach it.
The interesting engineering lies in the pairing. Most inference systems drop an off-the-shelf memory stack next to a compute die and hope the interconnect keeps up. The company designs both sides together, an approach known as processor-memory co-design. Data then moves less often, and each move costs less energy. That is where the efficiency argument comes from.
What does "agentic AI silicon" actually mean?
Why Samsung wrote the cheque
Samsung is the world's largest memory manufacturer, and for an inference startup that matters more than the cash. Memory bandwidth, not raw arithmetic, now sets the ceiling on how fast and how cheaply a model can serve requests. A partner who controls memory supply — and can also fabricate logic — removes two of the hardest bottlenecks a young chip company faces.
CEO Bernardo Kastrup told CNBC that the investor knows the supply chain as well as the balance sheet. The subtext is allocation. In a market where the best memory is committed years in advance, a strategic backer is worth more than a larger cheque from a passive fund.
The company also recruited a chairman with unusual weight. Peter Wennink, the former president and chief executive of ASML, joined the board. ASML builds the lithography machines that make advanced chips possible, and Wennink spent a decade at the centre of Europe's semiconductor supply chain. His arrival signals that the company is pitching industrial execution, not a research paper.
The bottleneck is memory, not arithmetic
Every serious inference company tells a version of the same story. Models got large. Serving them got expensive. The expense is no longer dominated by the maths. It is dominated by shuttling weights and activations between memory and compute, and by the energy that shuttling consumes.
AI is becoming a foundation of economic growth, scientific discovery and national competitiveness, but its potential will remain constrained unless we fundamentally change the infrastructure beneath it.— Bernardo Kastrup, Founder and CEO, Euclyd
As we wrote in September, the inference bottleneck is memory, not compute. The company is selling that thesis as silicon: a system where memory and processor are designed as one. A second piece in our archive traced how inference spending overtook training across 2026. It is the latest company to raise against that thesis.
The moat still to cross
The case against is straightforward and well rehearsed. Nvidia's real advantage is not the chip. It is CUDA, the software layer a generation of engineers learned first, plus a supply chain that took twenty years to assemble. Hyperscalers — Google, Amazon, Meta — are building their own inference silicon with unlimited internal demand and no need to win external customers. The startup has neither a software moat nor captive demand. It has an architecture, a balance sheet and a theory.
| Parameter | Nvidia | Hyperscaler in-house | Euclyd |
|---|---|---|---|
| Software ecosystem | ✔ CUDA, mature | ◐ internal only | ✗ unproven |
| Production scale | ✔ shipping now | ◐ ramping | ✗ first silicon 2028 |
| Memory strategy | ◐ sourced | ◐ sourced or co-packaged | ✔ co-designed |
| External customers | ✔ broad | ✗ captive only | ✗ none yet |
Assessment based on company statements and public reporting, September 2026
Read the table one way and the company looks overmatched. Read it another and the picture narrows. On memory co-design — the single axis it chose — no listed competitor is optimised the same way. That is a beachhead, not a fortress. Beachheads can grow. Most do not.
The other variable is capital intensity. A startup that designs both chips and memory must fund two expensive disciplines at once before it has a customer. The €200 million buys roughly two years of runway at that burn, which is one product cycle in silicon. Miss the 2028 window and the money is gone. This is a bet that the architecture is not merely better but cheap enough to reach the market before the incumbents close the gap.
What would it take by 2028?
Probability: 40% — the technical premise is credible and the backers are real, but silicon schedules slip and an unproven architecture must earn its place against software that already works.
✅ Arguments for
Sovereign AI programmes in Europe want an alternative to US supply, and the startup is European by design.
Inference, not training, is the faster-growing workload, which rewards hardware built for serving.
Confirmation criteria: a named design win with a hyperscaler, sovereign fund or large enterprise before end-2028.
❌ Arguments against
Two years of runway is one silicon cycle. A single delay consumes it.
Hyperscalers can absorb years of losses on in-house chips that a venture-backed startup cannot.
Disconfirmation criteria: a slipped 2028 ship date, or a design win that never converts to volume.
Development scenarios
🟢 Optimistic scenario (25%)
Implications: the company becomes Europe's credible inference challenger and raises again at a steep step-up.
🟡 Base-case scenario (55%)
Implications: a viable niche business, but not yet a threat to Nvidia's core.
🔴 Pessimistic scenario (20%)
Implications: the architecture survives as licensed intellectual property; the company does not survive as an independent systems vendor.
A named tape-out or foundry partner announcement before mid-2027.
The first independent benchmark of cost per token on craftwerk hardware.
Whether Samsung's memory allocation is contractual or merely strategic.
Any hyperscaler licensing its intellectual property instead of buying its systems.