- Euclyd Craftwerk SIP claims bandwidth levels far beyond the conceptions of Nvidia
- The Rack scale system would achieve exaflop performance
- Energy efficiency affirmations suggest massive gains, although independent tests are missing
The European startup Euclyd has announced new equipment aimed at a large scale inference of the AI.
The system, called Craftwerk, was introduced to Kisaco Infrastructure Summit 2025 in Santa Clara.
The company describes it as specifically designed for AI original workloads, with specifications that distinguish it from current accelerators.
Inside the Craftwerk architecture
At the heart of the version is the Craftwerk SIP, a package system that is part of the palm of one hand.
It incorporates 16,384 personalized SIMD processors alongside 1 TB of personalized ultra-band, or UBM.
Euclyd claims that this memory can deliver 8,000 teraoctets per second of bandwidth.
Calculation performance is listed 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 announce.
“Our manufactured calculation philosophy reinstates inference from zero, personalized processors, personalized memory and advanced packaging,” said Bernardo Kastrup, CEO of Euclyd.
“We have designed each door for maximum efficiency and a minimal draw, by far the lowest in the industry.”
The company also revealed the Craftwerk CWS 32 station, a rack-scale platform built from 32 SIP.
In this configuration, EuclyD indicates that the system reaches 1.024 Exaflops of Calculation FP4, supported by 32 TB of UBM.
It is said that it would generate 7.68 million tokens per second in multi-user mode, and its energy consumption is reported at 125 kilowatts.
According to the company, this is a gain at a hundred both in energy consumption and profitability compared to current alternatives.
The reference used to establish these improvements has been modeled with performance with Llama 4 Maverick.
Based in Eindhoven, the Netherlands, the company also retains offices in San Jose, California.
It promotes engineering and efficiency gains responsible on the environment in the infrastructure of the data center.
“I believe that the IA inference will dominate the Datacenter Silicon. Craftwerk’s economy will accelerate the adoption of AI and inaugurate abundant inference, “said investor Peter Wennk, former ASML CEO.
Although Craftwerk’s specifications are ambitious, complaints remain not tested outside the company’s framework.
Startups of the semiconductor space are often faced with challenges in large-scale manufacturing, to create reliable software support and to ensure integration with the existing data center infrastructure.
Euclyd’s announcement suggests a design that could, on paper, surpass lead accelerators, but if it can deliver in practice will depend on the results observed in real deployments.
Until these results emerge, the material remains an impressive set of numbers with an uncertain path towards generalized adoption.
