- Euclyd CRAFTWERK SiP claims bandwidth levels far beyond Nvidia’s designs
- The rack-scale system is said to reach over one exaflop performance
- Power efficiency claims suggest massive gains, though independent testing is missing
European startup Euclyd has announced new hardware aimed at large-scale AI inference.
The system, called CRAFTWERK, was introduced at the KISACO Infrastructure Summit 2025 in Santa Clara.
The company describes it as designed specifically for agentic AI workloads, with specifications that set it apart from current accelerators.
Inside the CRAFTWERK architecture
At the core of the release is the CRAFTWERK SiP, a system-in-package that fits into the palm of a hand.
It integrates 16,384 custom SIMD processors alongside 1TB of custom ultra-bandwidth memory, or UBM.
Euclyd claims this memory can deliver 8,000 terabytes per second of bandwidth.
The compute performance is listed at up to 8 petaflops in FP16 and 32 petaflops in FP4 precision.
These figures place the module above what established companies such as Nvidia currently advertise.
“Our Crafted Compute philosophy reimagines inference from the ground up, custom processors, custom memory, and advanced packaging,” said Bernardo Kastrup, CEO of Euclyd.
“We’ve engineered every gate for maximum efficiency and minimal power draw, by far the lowest in the industry.”
The company also revealed the CRAFTWERK STATION CWS 32, a rack-scale platform built from 32 SiPs.
In this configuration, Euclyd states the system reaches 1.024 exaflops of FP4 compute, backed by 32TB of UBM.
It is said to generate 7.68 million tokens per second in multi-user mode, and its power consumption is reported at 125 kilowatts.
According to the company, this is a hundred-fold gain in both energy use and cost efficiency compared to present alternatives.
The benchmark used to establish these improvements was modeled performance with Llama 4 Maverick.
Headquartered in Eindhoven, the Netherlands, the company also maintains offices in San Jose, California.
It promotes environmentally responsible engineering and efficiency gains in data center infrastructure.
“I believe AI inference will dominate datacenter silicon. CRAFTWERK’s economics will accelerate agentic AI adoption and usher in an era of abundant inference,” said investor Peter Wennink, former CEO of ASML.
While the specifications of CRAFTWERK are ambitious, the claims remain untested outside the company’s own framework.
Startups in the semiconductor space often face challenges in manufacturing at scale, building reliable software support, and ensuring integration with existing data center infrastructure.
Euclyd’s announcement suggests a design that could, on paper, outperform leading accelerators, but whether it can deliver in practice will depend on results seen in real deployments.
Until those results emerge, the hardware remains a set of impressive numbers with an uncertain path to widespread adoption.
