Emerging from stealth, Eindhoven-based Euclyd is reimagining inference from the ground up, with custom processors, custom memory and advanced 2.5D/3D packaging. The startup’s leadership brings together many of the brains from the illustrious Silicon Hive team.
Euclyd has introduced what it calls a breakthrough inference architecture for agentic AI. The architecture, currently in advanced design, is engineered to deliver the lowest power and cost per token in the industry. “Our philosophy reimagines inference from the ground up – custom processors, custom memory and advanced 2.5D/3D packaging,” says Bernardo Kastrup, founder and CEO of the startup that recently emerged out of stealth mode. “We’ve engineered every gate for maximum efficiency and minimal power draw.”
At the architecture’s core is the Craftwerk system-in-package, a palm-sized powerhouse featuring 16,384 custom SIMD processors delivering up to 8 PFLOPS (FP16) or 32 PFLOPS (FP4). Paired with 1 TB of custom ultra-bandwidth memory (UBM) offering a staggering 8,000 TB/s bandwidth, the SiP is claimed to set a new benchmark for data center inference efficiency. Craftwerk powers Euclyd’s flagship rack-scale system, the CWS 32, which integrates 32 SiPs to deliver 1.024 exaflops of FP4 compute and 32 TB of UBM. In multi-user mode, this system is projected to achieve 7.68 million tokens per second at just 125 kW for the state-of-the-art Llama 4 Maverick model, representing a 100x improvement in power efficiency and cost per token over leading alternatives.
This article is exclusively available to premium members of Bits&Chips. Already a premium member? Please log in. Not yet a premium member? Become one for only €15 and enjoy all the benefits.
