Euclyd raises €200M and receives Samsung funding to develop inference accelerators competitive with NVIDIA for AI workloads.

European AI chip sovereignty moved from aspiration to institutional capital on Tuesday as Eindhoven-based Euclyd announced a Series A financing round of more than €200 million (approximately $231 million)—one of the largest semiconductor raises in European history—co-led by Samsung, Somerset Capital Partners, the EU's Scaleup Europe Fund managed by EQT, and Innovation Industries. Peter Wennink, who ran chip-equipment giant ASML for eleven years and oversaw its ascent into the most strategically significant company in global semiconductor manufacturing, joined Euclyd's board as Chairman.

The centerpiece of Euclyd's bet is a custom AI inference chip system called CRAFTWERK, which the company claims delivers roughly 100 times greater power efficiency per generated token than Nvidia's newest Vera Rubin chips on comparable large language model workloads—a figure not yet independently validated at commercial scale. If it survives production, CRAFTWERK would represent a meaningful answer to the memory-bandwidth wall that consumes most of the power in today's GPU-based AI data centers.

**Understanding CRAFTWERK's Architecture**

The challenge CRAFTWERK addresses is fundamental to GPU power consumption. Nvidia's Vera Rubin chip, which entered full production in 2026, packs 288 gigabytes of HBM4 (High Bandwidth Memory, fourth generation) and delivers approximately 22 terabytes per second of memory bandwidth per GPU—2.8 times Blackwell's already-formidable 8 terabytes per second. Yet the NVL72 rack housing 72 Vera Rubin chips draws between 190 and 230 kilowatts of power. The reason is architectural: HBM sits adjacent to the GPU die but not on it; data must travel across physical connections between the memory stack and the compute cores, an energy-expensive process. For inference specifically—running a trained model to answer a user's query—this memory-movement cost can consume 60 to 80 percent of a system's power budget.

CRAFTWERK takes a different path. Instead of pairing a large compute die with adjacent HBM stacks, it integrates 16,384 custom SIMD (Single Instruction, Multiple Data) processors directly with what Euclyd calls Ultra Bandwidth Memory (UBM)—a custom memory architecture built specifically for this system rather than adapted from a standardized module—in a palm-sized single System-in-Package (SiP) measuring roughly 100 by 100 millimeters. The goal is to reduce the physical distance and energy cost of every data movement within the chip by processing information in place across many locations simultaneously.

The company's claimed specifications for the CRAFTWERK STATION CWS 32—a rack-scale system housing 32 of these chips—are: aggregate bandwidth of 8,000 terabytes per second, compute performance exceeding one exaflop (1,000 petaflops), 32 terabytes of on-system memory, and power draw of approximately 125 kilowatts. Running Meta's Llama 4 Maverick language model, Euclyd projects the system can generate 7.68 million tokens per second. The 100× efficiency figure cited against Vera Rubin refers to power efficiency per token on inference workloads, not raw compute.

Every specification comes from Euclyd's own modeling. TechRadar Pro noted when CRAFTWERK was unveiled at the KISACO Infrastructure Summit in Santa Clara in September 2025 that independent testing was absent. RuntimeWire.com stated directly: "a term sheet for a chip company is not the same thing as a shipped device, a qualified manufacturing flow or a customer-validated benchmark."

**Samsung's Dual Role**

The investor list is striking, but one detail warrants attention: Samsung is simultaneously a co-lead investor in Euclyd's Series A and the manufacturer of CRAFTWERK's first silicon in South Korea. In the semiconductor industry, manufacturing partners and financial investors are rarely the same entity, because the manufacturer holds enormous leverage over a startup's IP, production schedule, and unit economics. Samsung's willingness to wear both hats signals more than financial confidence in Euclyd's team. Samsung is one of only three suppliers—alongside SK Hynix and Micron—certified by Nvidia to produce HBM4 for Vera Rubin. By simultaneously backing a non-HBM inference architecture with its own capital and manufacturing capability, Samsung is hedging against the possibility that the AI inference market does not converge on Nvidia's GPU-plus-HBM model. If Euclyd's UBM architecture works at scale, Samsung is positioned to be the sole supplier of both the memory technology and manufacturing process for a serious Nvidia alternative.

Bernardo Kastrup, Euclyd's founder and CEO, acknowledged Samsung's importance beyond capital. "They are one of the biggest memory manufacturers in the world," Kastrup told CNBC. "They do a lot of engineering, they know a lot about systems, they know the supply chain, they have a huge network."

Kastrup's path is unusual even by startup standards. He holds two PhDs—one in computer engineering with specialization in reconfigurable computing and AI, and one in philosophy focused on ontology and philosophy of mind—and is the author of eleven books on analytic idealism, a philosophical position holding consciousness rather than matter as the fundamental substrate of reality. Before founding Euclyd, he co-founded Silicon Hive, a parallel processor company that Intel acquired in 2011, and worked at ASML as a technology strategist. He began developing CRAFTWERK's architecture in his home attic in Eindhoven, working late nights after his day job.

"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," Kastrup said in the Series A announcement.

**The EU's Strategic Bet**

The Scaleup Europe Fund's participation signals the fund's strategic priorities. Formally established by the European Commission on August 4, 2026, with €5 billion in total target capital, the fund made its first major investment the following day in Finnish satellite intelligence company ICEYE. Euclyd's round represents the fund's second major commitment in its opening weeks and its first in AI silicon specifically.

EQT, the Swedish private equity firm managing the fund, was selected in a competitive process that included 27 eligible managers. The fund backs European deep-tech scaleups in AI, quantum, clean energy, space, and biotech, with typical check sizes between €100 million and €500 million. Ted Persson, Partner at EQT and co-head of the Scaleup Europe Fund advisory team, said Euclyd "combines deep semiconductor expertise with a mission to tackle some of the hardest constraints in AI infrastructure."

**Wennink's Signal**

Peter Wennink's arrival as Chairman adds credibility beyond any specific technical claim. During his tenure as ASML's President and CEO from 2013 to 2024, Wennink oversaw the commercial deployment of extreme ultraviolet (EUV) lithography—the technology that makes manufacturing chips below 7 nanometers physically possible. Without EUV, there are no Nvidia H100s, no Apple M-series chips, no frontier AI acceleration. ASML is the world's only supplier of EUV equipment, a position Wennink spent over a decade defending and expanding. He has consistently argued that Europe faces difficult, consequential choices about securing its semiconductor future.

"Europe possesses world-class capabilities across semiconductors, advanced manufacturing, and systems engineering," Wennink said in the Series A announcement. "EUCLYD can transform those strengths into a globally competitive AI infrastructure platform."

That Wennink chose a two-year-old startup with no commercial silicon over a more established player reflects a judgment about where the inference silicon market is heading—and how seriously Europe's engineering community is taking the AI chip sovereignty question that Kastrup has framed as inseparable from control over hardware infrastructure.

---

**Note:** The source text appears to be cut off at "Can It Actually Compet" and lacks a conclusion. Please provide the complete article if you'd like the full body rendered.