News

Semidynamics looks to Nokia for the networking behind its AI racks

Exploratory agreement advances the company’s systems ambitions and gives European sovereignty pitch a practical infrastructure dimension

David Harold

Semidynamics and Nokia will explore integrating AI inference accelerators and liquid-cooled racks with Nokia’s data center networking. The MOU supports Semidynamics’ transition from processor IP to complete systems and gives Europe’s sovereign AI ambitions a practical infrastructure focus, with integration and commercial deployment still to be proven.

(Source: JPR AI)

Semidynamics and Nokia have signed a memorandum of understanding to explore combining the Barcelona company’s AI inference accelerators and liquid-cooled racks with Nokia’s data center networking. The agreement, announced on October 29, provides a framework for assessment, testing, validation, and, where appropriate, integration with Nokia’s Data Center Fabric solutions. Semidynamics will evaluate Nokia networking for its inference racks, and the findings will inform potential technical and commercial offerings.

The announcement follows Semidynamics’ move from licensing RISC-V processor IP toward developing its own silicon and integrated inference systems. Networking becomes a bigger part of the engineering problem when the product offered is a rack rather than a processor core.

Nokia brings a portfolio covering data center switching and network automation, addressing the connectivity needed to operate computing infrastructure at scale. Its Data Center Fabric offering targets both conventional cloud infrastructure and AI workloads.

The companies frame the agreement around Europe’s sovereign AI ambitions. Their proposed contribution is a broader choice of European computing and networking technologies for organizations building AI infrastructure.

Semidynamics CEO Roger Espasa said the agreement provides a framework to advance the company’s memory-centric inference platform alongside the networking required for large AI infrastructure. Nokia’s stated emphasis is on contributing to a broader, open ecosystem for trusted AI infrastructure.

The scope remains exploratory. The announcement does not provide a delivery schedule.

What do we think?

This is a logical next step in the business Semidynamics is trying to build—everyone wants to have a full stack. We have previously argued that RISC-V is becoming secondary to its story. The ISA provides architectural freedom; the commercial proposition rests on memory handling, processor utilization, and the economics of running inference.

In our recent interview, CTO Iakovos Stamoulis explained the reasoning behind Semidynamics’ use of LPDDR memory. More capacity per accelerator can keep larger models and KV caches local, potentially reducing communication between devices. The trade-off is lower bandwidth than HBM, which puts more responsibility on the architecture to use that memory effectively.

Working with a networking supplier is consistent with that approach. Keeping more work local can reduce communication requirements, but a rack still needs to distribute requests, connect computing resources, and operate reliably under load. The balance will depend on the models being served, how they are distributed, and the number of concurrent users. Moving from IP to systems changes what customers expect a supplier to deliver. They need an installation that can be deployed, managed, and supported. A relationship with Nokia could help Semidynamics address that requirement.

However, Semidynamics has no NIC or DPU like ConnectX or BlueField, and the public record doesn’t yet say how its racks connect to Nokia’s fabric. Nokia mentions Data Center Fabric (Ethernet leaf-and-spine switching and routing) and SR Linux with event-driven automation for configuration and monitoring.

The press release doesn’t name the interface standard, and we found no mention of PCIe, UEC, or UALink. One report mentions Ethernet.

The sovereignty argument becomes more useful when it identifies equipment that customers could actually buy and operate. European accelerator technology connected through European networking offers a more concrete proposition than a general call for technological independence. As Stamoulis previously told us, the potential market also extends beyond Europe. Some customers want greater choice between American and Chinese technology suppliers. A European infrastructure offering could serve that demand.

An MOU is an early step. Its significance here is that Semidynamics is addressing another of the practical requirements of becoming an AI systems supplier. 

LIKE WHAT YOU’RE READING? TELL YOUR FRIENDS; WE DO THIS EVERY DAY, ALL DAY.