Every AI answer has a hidden commute.
Before a model can produce the next word, enormous amounts of information must move between memory and processors. That movement consumes bandwidth, power and money. As AI shifts from models being trained occasionally to models answering requests millions of times a day, EUCLYD believes that journey is becoming one of the industry’s most expensive problems.
The Eindhoven semiconductor company has now raised more than €200 million in Series A funding, co-led by Samsung, Somerset Capital Partners, the Scaleup Europe Fund managed by EQT and Innovation Industries. EIFO, imec.xpand, Brabant Development Agency and Quadri also participated. Former ASML CEO Peter Wennink is joining as chairman.
Founded by Bernardo Kastrup and Atul Sinha, EUCLYD is building CRAFTWERK around a different assumption from conventional GPU infrastructure: rather than continually moving model data toward a smaller number of large processors, place far more memory closer to the processors that need it.
Its planned system-in-package combines 16,384 custom SIMD processors with 1 TB of custom Ultra-Bandwidth Memory. EUCLYD has projected that a 32-package CWS 32 rack could reach 1.024 exaflops of FP4 compute and 7.68 million tokens per second at 125 kW. Those figures come from modeled performance rather than independently validated commercial deployments, an important distinction while the architecture remains on its road to production.
That also makes Samsung’s role more interesting than the funding headline suggests.
“Samsung can help us in more ways than money” Kastrup told in CNBC Interview. Samsung brings memory manufacturing, systems engineering, supply-chain experience and a large semiconductor network to a company whose thesis depends heavily on redesigning the relationship between processors and memory.
EUCLYD is targeting two routes to market: selling physical inference systems to enterprises that want self-hosted AI, and licensing its underlying technology to companies developing custom silicon. Kastrup told the company aims to begin rolling out physical systems in 2028, with an ambition to serve thousands of enterprise customers by 2030.
That timeline matters. EUCLYD has raised substantial capital around an architecture whose toughest evidence still lies ahead: manufactured silicon, production workloads and independently demonstrated economics.
The company calls its long-term vision “Abundant Intelligence.” But abundance will not be decided by how many calculations AI can perform.
It may be decided by how little energy is wasted moving the data those calculations need.
