Eindhoven-designed agentic AI inference silicon · 16,384 × 128-way SIMD processors · 2.5D/3D system-in-package · CWS 32 rack scales to 1.024 exaFLOPS FP4
A palm-sized System-in-Package that pairs custom inference silicon with purpose-built memory.
Rack-scale token factory aggregating 32 CRAFTWERK SiPs into a single exascale-class system.
Euclyd’s custom memory tier, designed to sidestep both SRAM scaling limits and HBM bandwidth ceilings.
Euclyd B.V.
Eindhoven, Netherlands
San Jose, CA
CRAFTWERK SiP
Craftwerk Station CWS 32
Agentic AI inference accelerator
Systems-in-package & rack
16,384 × 128-way SIMD processors
8 PFLOPS FP16
32 PFLOPS FP4
1 TB Ultra Bandwidth Memory (UBM)
per SiP
8,000 TB/s per SiP
~3 kW per SiP
Advanced 2.5D & 3D system-in-package
Palm-sized, liquid-cooling ready
Craftwerk Station CWS 32
32 SiPs + 16 CPUs
1.024 exaFLOPS FP4
32 TB UBM
256 PB/s aggregate
~125 kW, liquid cooled
100× better power efficiency / cost per token vs leading solutions
244,000 tokens/s multi-user on Llama 4 Maverick (CWS 32, modeled)
ADTechnology (Korea)
semicustom development partnership
Peter Wennink (ex-ASML CEO)
Federico Faggin, Steven Schuurman
In advanced design
Announced at Kisaco AI Infra Summit
Quote
Euclyd is a European inference-silicon startup based in Eindhoven, the Netherlands, with a second office in San Jose, California. Its CRAFTWERK architecture rethinks inference “from the ground up” — custom processors, custom memory and advanced 2.5D/3D packaging — in pursuit of the industry’s lowest power and cost per token for agentic AI workloads. The premise is that memory bandwidth, not raw arithmetic, is the binding constraint on inference economics: Euclyd claims roughly 100× better power efficiency and cost per token than leading incumbent solutions.
The CRAFTWERK SiP is a palm-sized system-in-package carrying 16,384 custom SIMD processors and 1 TB of Ultra Bandwidth Memory (UBM) delivering a claimed 8,000 TB/s. Peak compute is 8 PFLOPS in FP16 or 32 PFLOPS in FP4 at approximately 3 kW. By placing a terabyte of memory in the package rather than over an HBM interface, Euclyd targets dense single-silicon multi-agent serving without the capacity and bandwidth ceilings that constrain HBM-based accelerators.
At rack scale, the Craftwerk Station CWS 32 aggregates 32 CRAFTWERK SiPs and 16 CPUs into a single liquid-cooled system rated at 1.024 exaFLOPS FP4 with 32 TB of on-package memory and 256 PB/s of aggregate bandwidth, drawing roughly 125 kW. Euclyd positions the rack as the world’s lowest-power exascale AI token factory. Silicon development is being carried out with Korean design house ADTechnology, and the company is backed by figures including former ASML CEO Peter Wennink and microprocessor pioneer Federico Faggin. CRAFTWERK figures are modeled projections from a design still in advanced stage; availability is quoted on request.
Need Euclyd CRAFTWERK or CWS 32?
Contact QS Compute for availability, configuration, and volume pricing.
Request Quote