Euclyd's $231M AI Inference Chip Funding, Explained

Euclyd, a Dutch chip startup, closed a $231 million ($200 million euro-denominated) Series A on September 14, 2026, led by Samsung alongside Somerset Capital Partners, EQT's Scaleup Europe Fund, and Innovation Industries, according to CNBC. Former ASML chief executive Peter Wennink is joining as board chairman. The round is the largest European AI-inference chip raise of 2026, and it matters to anyone building AI products for a reason that has nothing to do with the funding number itself: it is a bet that inference, not training, is where the real hardware bottleneck sits next.

What Euclyd is actually building

Euclyd, founded in 2024 by Bernardo Kastrup and Atul Sinha, is not building a GPU competitor for training frontier models. Its roadmap targets inference specifically: the step where an already-trained model answers a live query, which is the step every AI product you ship actually runs in production. The company's pitch is custom chips paired with a processor-memory co-design, aimed at cutting per-token cost and energy use compared to running inference on general-purpose GPUs.

Kastrup told CNBC the company expects to ship physical chip systems in 2028, targeting thousands of enterprise customers by 2030. That timeline is the detail worth sitting with.

Why builders should read the timeline, not the headline

Coverage of chip funding rounds tends to imply "AI is about to get much cheaper." Euclyd's own numbers say otherwise, at least for now. 2028 is the ship date for hardware that has not yet been built at scale, and adoption curves for new chip architectures run years past first shipment, since anyone with existing GPU infrastructure needs a real reason to migrate. If you are pricing an AI feature into your product roadmap this quarter, a 2028 hardware bet is not a variable in that decision.

What is a near-term variable: this is money, and a serious executive (Wennink ran ASML, the company that makes the machines that make every advanced chip), betting that GPU-based inference has a cost ceiling worth building an entire company around. That is a signal about where the industry expects the pressure to keep building, even if the specific fix is years out.

Euclyd's plan has two revenue lines: selling physical rack systems to enterprises that want to self-host inference, and licensing its underlying IP to other chipmakers. The second line matters more for the average builder than the first. If the architecture works, it can show up inside chips from vendors you already buy from, on a timeline decoupled from Euclyd's own hardware shipping, the same way IP licensing let ARM's designs end up in chips its own name never appears on.

The bigger pattern this round fits

Euclyd is not alone in chasing inference-specific hardware. NPUs, positioned as a lower-power alternative to GPUs for inference workloads, have been shipping in consumer devices for several product cycles already, with the same underlying thesis: training and inference have different hardware needs, and the industry has mostly been running inference on hardware optimized for the wrong job.

For a founder deciding what model to build against right now, the practical read is this: current published API pricing is the number that governs your unit economics today. Hardware bets like Euclyd's are worth tracking as a two-to-four-year signal about where the cost floor is heading, not as a reason to delay a decision you need to make this quarter.

Frequently asked questions

Is Euclyd's chip available now?

No. Euclyd's own timeline, per CNBC's reporting, targets physical chip systems shipping in 2028. There is nothing to buy or deploy today.

How is Euclyd different from Nvidia?

Nvidia's GPUs handle both AI training and inference, using general-purpose parallel architecture. Euclyd is building hardware specifically for inference, the step where a trained model responds to a query, with a processor-memory design aimed at cutting per-token cost for that one workload rather than training.

Will this make AI APIs cheaper for developers?

Not directly, and not soon. Euclyd sells hardware to enterprises and licenses IP to chipmakers; it does not run a public API. Any effect on the price you pay OpenAI, Anthropic, or another provider would depend on those providers or their infrastructure partners adopting inference hardware like this, which would take years even after Euclyd ships.