On September 15, 2026, Dutch semiconductor startup EUCLYD announced a €200 million (approximately $230 million) Series A, co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund, and Innovation Industries. Former ASML President and CEO Peter Wennink joins EUCLYD’s board as Chairman. The raise is one of the largest early-stage rounds for a European AI infrastructure company and signals that the race to build purpose-built inference hardware has become a global competition, not a Silicon Valley monopoly.

The core thesis behind the funding is that running large language models at production scale costs far more than it should, and that the chip architecture driving current inference economics was never designed for the job.

What EUCLYD Is Building

EUCLYD describes its flagship product, craftwerk, as the world’s first agentic AI silicon: a programmable ASIC chip built from the ground up for the memory-access profile of LLM token generation, not retrofitted from graphics processing.

Each craftwerk SiP (system-in-package) integrates 16,384 custom SIMD processors capable of delivering up to 8 PFLOPS (FP16) or 32 PFLOPS (FP4). More importantly, each SiP pairs those processors with 1 TB of custom ultra-bandwidth memory (UBM) at 8,000 TB/s bandwidth. That memory bandwidth figure is the core design bet: EUCLYD is optimizing for the bottleneck that limits GPU-based LLM inference, where most compute cycles are spent waiting for model weights to transfer from high-bandwidth memory to compute units rather than performing arithmetic.

The rack-scale system, craftwerk station CWS 32, integrates 32 SiPs to deliver 1.024 exaflops of FP4 compute and 32 TB of UBM across the full rack. In multi-user mode, EUCLYD projects CWS 32 will achieve 7.68 million tokens per second at 125 kW of power draw. The company claims this represents a 100x improvement in power efficiency and cost per token versus leading GPU alternatives for Llama 4 Maverick inference workloads.

Those figures are modeled performance targets based on the chip design, not results from commercially deployed hardware. EUCLYD is targeting commercial availability starting in 2028. But the design philosophy they represent is the key signal: a chip purpose-built for inference, optimized for memory bandwidth per watt rather than raw FLOPS per watt, organized around the actual bottleneck of LLM serving rather than the bottleneck of graphics rendering.

Metriccraftwerk station CWS 32 (Projected)
Total compute (FP4)1.024 exaflops
Memory (UBM)32 TB
Memory bandwidth per SiP8,000 TB/s
Multi-user throughput (Llama 4 Maverick)7.68M tokens/second
Power consumption125 kW
Efficiency vs GPU alternatives (modeled)100x (power and cost per token)
Commercial deployment target2028

Sources: EUCLYD craftwerk announcement, September 2025; EUCLYD Series A press release, September 2026.

The Efficiency Wall Problem

EUCLYD’s framing of the problem as an “efficiency wall” captures something concrete. Each successive generation of foundation models is more capable than the last, and each successive generation costs more per inference token to deploy. The incremental cost of running GPT-6 Astra or Claude Fable 5.1 in a production agentic workflow is meaningfully higher than running the models from two years ago, even accounting for hardware improvements.

For enterprises, that math creates a ceiling. Not every use case can absorb per-token costs at frontier model prices. A team that wants to run a continuous sales intelligence agent, an always-on document review pipeline, or a real-time GTM data synthesis layer runs into economics that force trade-offs between capability and cost before it runs into model capability limits. The infrastructure is the constraint, not the model.

Nvidia’s GPU architecture was designed for training, where raw FLOP counts determine time-to-result and memory access patterns are structured differently than during inference. The CUDA ecosystem and the installed base of H100 and B200 clusters mean Nvidia dominates inference today as an extension of its training dominance, not because GPUs are the optimal inference architecture. Several companies, including Etched, Cerebras, and now EUCLYD, are making the bet that a purpose-built inference architecture will outperform repurposed training hardware once it reaches production scale.

CNBC reported that EUCLYD CEO Bernardo Kastrup sees two paths to revenue: selling physical hardware to enterprises that want secure, on-premise inference, and licensing the company’s IP to organizations that want to build their own chips using EUCLYD’s architecture. The dual model is designed to let EUCLYD scale both direct enterprise revenue and indirect platform leverage, similar to how Arm licenses its instruction set architecture while also selling optimized core designs.

The Samsung Connection

Samsung’s co-leadership of the round reflects more than financial interest. Samsung is the world’s largest DRAM manufacturer and one of the few companies with the fabrication capacity, memory expertise, and system-level integration experience to make EUCLYD’s processor-memory co-design architecture manufacturable at commercial volume.

The ultra-bandwidth memory at the center of craftwerk requires tight physical integration between compute and memory layers, a design constraint that is fundamentally a semiconductor manufacturing problem. Samsung’s involvement signals that at least one major memory manufacturer has reviewed EUCLYD’s architecture and found it credible enough to back with capital and implied supply-chain access.

According to EU-Startups, the round also includes EIFO (Denmark’s Export and Investment Fund), imec.xpand (the venture arm of the European semiconductor research consortium imec), the Brabant Development Agency (BOM), and Quadri. The participation of imec.xpand is particularly notable: imec sits at the intersection of advanced semiconductor research and European industrial policy, and its endorsement carries technical credibility beyond standard financial backing.

Peter Wennink’s chairmanship connects EUCLYD to the supply chain expertise that made ASML the world’s most strategically important semiconductor equipment company. Wennink oversaw ASML during the period when extreme-ultraviolet lithography transitioned from research concept to the production technology that now manufactures every advanced chip in the world. That background is specifically relevant to a company whose differentiation depends on manufacturing innovations in processor-memory co-design.

What This Means for Enterprise AI Teams

The enterprise AI story of 2026 has been dominated by model launches and agent platform announcements. The infrastructure layer underneath those models receives less attention but determines which agentic use cases survive contact with a finance team.

For enterprises building AI-native architectures at scale, inference economics determine which workflows are viable and which remain too expensive to run continuously. A 100x efficiency improvement, if EUCLYD’s projections hold in production, would shift the economic threshold for agentic AI dramatically. Workflows that currently cost too much to run at enterprise scale (continuous research agents, high-volume document intelligence pipelines, real-time GTM synthesis) could become routine.

Even for teams that will never purchase a craftwerk system directly, EUCLYD’s funding matters. Competition in the inference hardware market creates downward pressure on GPU pricing that flows through to cloud provider per-token costs. Every credible alternative to Nvidia’s inference stack reduces the pricing leverage Nvidia holds over enterprise AI infrastructure budgets. The broader the field of inference hardware competitors, the faster per-token costs fall across the entire market.

The sovereign AI dimension is also significant. European enterprises and government agencies often face legal or operational constraints on data residency that make US-based cloud inference complicated to deploy. A European inference hardware company targeting on-premise deployment, backed by European institutional funds and positioned explicitly for sovereign and enterprise markets, addresses that constraint at the hardware layer rather than through contractual data processing agreements.

Infrastructure Competition Ahead

EUCLYD is early. The hardware has not shipped at commercial scale, the competitive landscape will shift considerably before 2028, and Nvidia is not standing still. Blackwell successors are already in development, and Nvidia’s software ecosystem and enterprise relationships create structural advantages that a chip architecture alone cannot overcome quickly.

What EUCLYD represents is a specific structural thesis: the inference workload of autoregressive LLMs is distinct enough from the training and graphics workloads that shaped GPU architecture to justify a purpose-built compute design. The companies best positioned to evaluate that thesis, including Samsung and imec, have concluded it is worth $230 million in early-stage capital to find out.

For enterprise AI leaders, the relevant question is not whether to buy EUCLYD hardware today. It is whether the efficiency wall the company is targeting is real, and whether the combination of EUCLYD, Etched, Cerebras, and others accelerates a shift in inference economics fast enough to change the deployment calculus for agentic AI in the next two to three years. Based on the capital flowing into this space and the technical credentials of the teams pursuing it, the answer is increasingly yes.

This is the infrastructure layer that will determine how broadly AI agents move from adoption to operational deployment across the enterprise. The model capability is largely there. The economic infrastructure to deploy it continuously and at scale is the next frontier.

If you are designing an AI-native enterprise system and want to think through how inference economics affect your agent deployment strategy, book a conversation with Enera.