Samsung Leads $231M Bet on Dutch Inference Chip as Supply Chain Begins Structural Hedging Against Nvidia

On September 15, 2026, Euclyd, an AI chip company based in Eindhoven, the Netherlands, announced the completion of a Series A round of more than €200 million (approximately $231 million, according to CNBC), co-led by Samsung Electronics, EQT’s Scaleup Europe Fund, Somerset Capital Partners and Innovation Industries, with participation from the Danish Export Investment Fund EIFO, imec.xpand, the Brabant Development Agency and Quadri. Disclosed alongside the financing was a personnel appointment: former ASML President and CEO Peter Wennink joined Euclyd’s board as chairman. The company, founded in 2024, has secured a rare-sized Series A for the chip startup sector, with the goal of launching an alternative AI inference infrastructure to GPUs within two years.

GPU’s Structural Baggage: Inference Was Never Its Main Battlefield

To understand the significance of this financing, one must first clarify a technical contradiction at the industry level: training large models and running large models place fundamentally different demands on chips. Training requires extremely high floating-point compute density, which is precisely the core advantage Nvidia’s GPUs have accumulated over decades. Inference, however, is more concerned with tokens produced per unit of energy, memory bandwidth utilization and response latency—and these three dimensions happen to be the areas where GPU architectures have long carried historical baggage. Nvidia’s GPUs originated in gaming graphics rendering; their massively parallel computing structure was designed to drive pixel fill, not to complete LLM token-by-token autoregressive generation at extremely low power. Using GPUs for large-scale commercial inference provides ample power, but efficiency involves structural compromises.

Euclyd’s proposition targets precisely this gap. The company’s Craftwerk chip uses a programmable ASIC architecture, combined with a proprietary memory architecture and processor-memory co-design, physically coupling compute units and memory more tightly to fundamentally cut the energy consumption and latency of data movement. According to the company, Craftwerk is equipped with 16,384 SIMD processors, reaches peak compute of 8 PFLOPS (FP16), and is delivered to enterprise customers as a complete rack system, Craftwerk Station CWS 32. The company positions it as “the world’s first dedicated silicon for agentic AI.”

Samsung’s Real Calculation Goes Beyond Financial Return

In this deal, the logic behind Samsung’s involvement deserves separate scrutiny. Euclyd CEO Bernardo Kastrup said in an exclusive interview with CNBC: “Samsung can offer us far more than capital. They are one of the world’s largest memory manufacturers, with substantial engineering capabilities, supply chain networks and a vast partner ecosystem.” This statement reveals the real value axis of the investment: one of the hardest technical bottlenecks for inference-specific chips is precisely the supply and system integration of high-bandwidth memory (HBM). Samsung is not only a leading global HBM supplier, but also one of Nvidia’s largest memory suppliers. By entering Euclyd’s equity structure as a lead investor, Samsung is in effect opening a parallel outlet for its memory business that does not depend on the Nvidia ecosystem—if the non-GPU inference direction grows in the future, Samsung’s technology and supply chain capabilities can monetize on two tracks at once, rather than relying solely on orders from a single customer.

According to EU Startups, Dedi Goldschmidt, Senior Vice President at Samsung Semiconductor Innovation Center, said: “The next phase of AI will be defined not only by model innovation, but equally by the efficiency and scalability of infrastructure. Euclyd combines an excellent team with a differentiated vision to directly address the core constraints of AI data centers.” This statement clearly characterizes Samsung’s involvement as proactive positioning rather than passive financial allocation.

The addition of former ASML President Peter Wennink adds credibility on another dimension. ASML is the sole supplier of extreme ultraviolet (EUV) lithography machines globally, dominating the infrastructure for the most advanced chip manufacturing. During his tenure at ASML, Wennink grew it from a mid-sized European equipment maker into a company with structural bargaining power across the semiconductor industry. His endorsement provides implicit support for Euclyd’s access to supply chain resources and industrial credibility within Europe’s semiconductor ecosystem. Combined with follow-on investment from EIFO (Danish sovereign capital) and the Brabant Development Agency, the structure of this funding round shows a policy intent for Europe to actively cultivate a domestic AI chip route—not merely a commercial bet.

Two Paths in the Inference Chip Competition

The closing of Euclyd’s financing contrasts with progress on another non-GPU route. In August 2026, OpenAI announced that “Jalapeño,” an inference-specific ASIC developed in collaboration with Broadcom, has “industry-leading speed and efficiency.” According to multiple technology media reports, the chip showed a 1.5x to 1.9x improvement in AI work per watt in benchmarks and is scheduled for small-volume shipments by the end of 2026. But Jalapeño and Euclyd are pursuing completely different commercial paths: Jalapeño is a fully internalized OpenAI chip that outside companies cannot purchase; in essence, a hyperscaler using its own silicon to carve away its own GPU purchasing demand. Euclyd explicitly targets two externalized commercial paths: selling hardware and rack systems to enterprises (that is, on-premises deployments supporting self-hosted inference), while licensing its intellectual property to companies that want to develop their own chips.

Together, these two paths reveal the structure of competitive pressure Nvidia currently faces: not from a single head-on challenger, but from simultaneous nibbling across multiple ecological niches. Hyperscaler internalization carves away high-end cloud training and inference demand, while independent chip suppliers target enterprise and government markets with strong data sovereignty needs. Neither aims to fight head-on in areas where Nvidia excels; instead, they seek to establish footholds in scenarios Nvidia cannot cover or where its efficiency is clearly low.

For engineering teams actually building LLM applications, Euclyd will not change any practical toolchain choices before 2028. But the signal sent by this financing is that competition to compress inference costs has formally entered a multi-player phase, and the main battlefield is beginning to shift from “whose model is stronger” to “whose inference infrastructure is cheaper.” If Craftwerk’s energy efficiency claims are validated in mass production in 2028, cost per token of inference will enter a new compression cycle, directly affecting the cost logic of building the AI application layer—especially for agent-style high-frequency inference demand, where this variable is more decisive than model capability itself.

The Key Milestone in 2028

Euclyd’s most critical uncertainty is its ability to deliver on mass production, not its technological vision itself. “Shipping in 2028” means the company needs to complete the entire chain from prototype to deliverable commercial system in about two years—design freeze, tape-out iterations, foundry scheduling, system integration and yield ramp—and every step could slip. Samsung’s involvement reduces memory supply chain risk, but cannot replace mass-production validation of the compute architecture itself. Most chip startups do not fail on design; they fail on the distance from design to reliable shipment.

There are three signals for whether Euclyd is on track: first, whether independent third-party technical evaluation data (rather than the company’s own) appears in 2027; second, whether large enterprise or government customers disclose early purchase intentions or proof-of-concept collaborations; and third, whether Samsung makes a formal technical integration announcement linking its HBM product line with the Craftwerk platform. If two or more of these signals materialize by the end of 2027, the credibility of the 2028 shipment timetable will rise significantly. Conversely, if the company’s future external communications focus on financing rather than technical progress, that is usually an early signal of pressure on the mass production timetable.

The true test of this wave of non-GPU inference chip investment is not the day of the funding announcement, but the day in 2028 when the first commercial customer actually deploys Craftwerk. Until then, all claims about energy efficiency are architectural assumptions not yet validated at scale—that is both where the startup opportunity lies and where the risk lies.