IN Brief:
- Semidynamics and SiPearl are developing a European rack-scale AI compute platform for cloud inference.
- The design combines SiPearl’s Arm-based CPU with Semidynamics’ RISC-V-based GPU/AI inference ASIC.
- The project supports Europe’s push for more sovereign AI infrastructure and OCP-aligned data centre systems.
Semidynamics and SiPearl have formed a strategic cooperation to develop a European rack-scale AI compute platform for large-scale inference workloads in cloud and enterprise environments.
The planned system will combine SiPearl’s Arm-based CPU technology with Semidynamics’ RISC-V-based GPU/AI inference ASIC. SiPearl’s processor will provide general-purpose compute, orchestration, and data plane hosting, while Semidynamics’ accelerator will handle AI inference workloads and future performance scaling.
The companies are targeting European AI infrastructure programmes, including AI Factory and Giga Factory applications, as well as public and private cloud deployments. The rack design is expected to follow Open Compute Project standards, aligning the platform with established data centre mechanical, electrical, and serviceability practices.
Roger Espasa, CEO of Semidynamics, said: “Combining SiPearl’s high-performance CPU with Semidynamics’ RISC-V-based GPU/AI inference technology gives Europe a credible path towards sovereign, rack-scale AI infrastructure built around European-controlled compute.”
SiPearl emerged from the European Processor Initiative and is developing processors for high-performance computing, AI, and data centre applications. Semidynamics, based in Barcelona, develops RISC-V processor IP and memory-focused AI infrastructure, including technologies designed to reduce performance losses caused by memory bottlenecks in inference workloads.
AI infrastructure is increasingly being engineered at rack level rather than around individual accelerators. Training systems have been defined by large GPU clusters, but inference creates different requirements: sustained throughput, predictable latency, lower energy per query, high utilisation, and software stacks that can support frequent model updates. Those requirements change the balance between host CPU, accelerator, memory subsystem, networking, orchestration, and cooling.
European compute strategy has also moved beyond general concerns over data location. Processor origin, system design control, procurement resilience, export exposure, and software dependency are now part of infrastructure planning for public-sector and industrial AI deployments. A CPU-plus-accelerator architecture built around European-controlled silicon gives cloud and enterprise operators a more direct route to regional hardware options.
The use of Open Compute Project alignment gives the proposed platform a practical route into data centre environments. Rack-scale AI systems must fit existing assumptions around power distribution, cooling, service access, networking, telemetry, and maintenance. Bespoke architecture with poor physical compatibility can struggle even when the underlying silicon is competitive.
The integration work will be demanding. Host processing, accelerator performance, memory bandwidth, firmware, interconnect, orchestration software, and thermal management all have to behave as a coherent platform. RISC-V gives Semidynamics architectural flexibility for accelerator development, while SiPearl’s Arm-based CPU provides a host environment closer to current infrastructure software practice.
The companies also plan to explore future chiplet-level integration. Chiplet architectures are becoming more prominent as AI and high-performance compute designs push package-level bandwidth, power efficiency, and yield economics beyond the limits of large monolithic devices. Combining heterogeneous compute elements at package level could help future systems reduce latency and improve energy efficiency, provided software and interconnect layers keep pace.
The cooperation gives Europe another route into AI inference hardware at a time when demand is moving from experimental deployments to operational infrastructure. If the platform advances from design cooperation into qualified rack systems, it would add a processor-led European option to a market currently dominated by non-European accelerator ecosystems.
