Bull and Kalray partner to build Ultra Ethernet AI interconnects

This move signals a shift away from proprietary interconnect lock-in, such as InfiniBand, toward open Ultra Ethernet standards that enhance GPU utilization for massive AI deployments. For the streaming ecosystem, this provides a more scalable networking foundation to handle the surge in data flows required for real-time AI-enhanced video processing and large-scale rendering. The collaboration positions Europe as a provider of independent, high-throughput infrastructure at a time when North American and Asian chipset dominance is under scrutiny. Watch for the initial release of the joint 1.6 Tbps chips in 2028 to see if they meet the ultra-low latency benchmarks required for competitive exascale computing.

The partnership between Bull and Kalray aligns with a broader push for European 'technological sovereignty' in the global AI race. Per the European Commission in June 2026, the EU unveiled a 'tech sovereignty' package designed to reduce dependence on foreign-owned Big Tech by fostering homegrown alternatives in semiconductors and cloud infrastructure. This initiative follows the French state's March 2026 completion of a €404 million acquisition of Bull from Atos, placing the continent’s only supercomputer manufacturing facility under national control to protect core strategic interests.

The technical backbone of this effort is the Ultra Ethernet Consortium (UEC), which launched its version 1.0 specification in June 2025. According to UEC, the standard is purpose-built for the AI era, providing a modernized Remote Direct Memory Access (RDMA) transport layer that allows data to bypass CPUs and move directly to application memory. By late 2026, the consortium had grown to over 170 members, including industry giants like AMD and Cisco, all seeking to replace legacy protocols that struggle at the scale of modern AI 'gigafactories.'

Kalray’s contribution centers on its newly modified Networking Engine, which utilizes 64 cores based on the RISC-V open standard to achieve 1.6 Tbps throughput. This hardware enables a hybrid networking mode, allowing cards to switch between low-latency HPC modes and virtualized AI workloads. As reported by Techzine in April 2026, Bull’s existing systems already power Europe’s first exascale supercomputer, Jupiter, and the upcoming Alice Recoque system, highlighting the immediate real-world infrastructure where these next-generation interconnects will likely be deployed.