- EUCLYD Craftwerk SIP affirms bandwidth levels far beyond Nvidia designs
- It is said that the rack scale system reaches exaflop yield
- Energy efficiency statements suggest mass profits, although independent evidence is missing
The Euclyd European startup has announced a new hardware aimed at the inference of large -scale AI.
The system, called Craftwerk, was introduced at the Kisaco 2025 infrastructure summit in Santa Clara.
The company describes it as specifically designed for the workloads of Agent, with specifications that distinguish it from current accelerators.
Within Craftwerk architecture
In the core of the launch is the Craftwerk SIP, a system in packaging that fits the palm of one hand.
Integrates 16,384 Simd Customized Processors together with 1 TB of Ultra Personalized Band Memory or UBM.
EUCLYD states that this memory can deliver 8,000 terabytes per second bandwidth.
The calculation yield is listed in up to 8 Petaflops in FP16 and 32 Petaflops in FP4 precision.
These figures place the module above what the established companies, such as NVIDIA, are currently announced.
“Our elaborated computing philosophy reimagines inference from scratch, custom processors, personalized memory and advanced containers,” said Bernardo Kastrup, CEO of Eucalyd.
“We have designed each door to obtain maximum efficiency and a minimum energy raffle, with much, the lowest in the industry.”
The company also revealed the Craftwerk Station CWS 32, a shelf scale platform built from 32 sips.
In this configuration, EUCLYD states that the system reaches 1,024 FP4 computing exachers, backed by 32TB of UBM.
It is said that it generates 7.68 million tokens per second in multi -user mode, and its energy consumption is informed to 125 kilowatts.
According to the company, it is a profit of a hundred times both in energy consumption and in profitability efficiency compared to current alternatives.
The reference point used to establish these improvements was modeled with flame 4 maverick.
Based in Eindhoven, the Netherlands, the company also maintains offices in San José, California.
Promotes engineering and environmental efficiency profits in the infrastructure of the data center.
“I think that the inference of IA will dominate Datacenter Silicon. Craftwerk’s economy will accelerate the adoption of agent and the beginning of an era of abundant inference,” said investor Peter Wennink, former CEO of ASML.
While Craftwerk’s specifications are ambitious, the statements remain not proven outside the company’s own frame.
Startups in the space of semiconductors often face challenges in scale manufacturing, creating reliable software support and guaranteeing integration with the infrastructure of existing data centers.
The Euklyd announcement suggests a design that could, on paper, overcome leading accelerators, but can deliver in practice it will depend on the results seen in real implementations.
Until these results arise, hardware remains a set of impressive numbers with an uncertain path towards generalized adoption.
