Bull and Kalray have partnered to develop next-generation networking technology for AI and high-performance computing (HPC), targeting 1.6 Tbps throughput per network card and full compatibility with the emerging Ultra Ethernet standard. The collaboration, announced on July 28, 2026, aims to deliver commercial solutions by 2028 that combine ultra-low latency for HPC with Ethernet-based protocols for cloud-scale AI workloads, while strengthening European technological sovereignty.
The deal addresses the explosion of data traffic driven by AI adoption, which demands faster, more scalable, and cost-effective interconnects. Both companies emphasize reducing dependence on non-European networking technologies, particularly in critical infrastructure sectors. Bull and Kalray will design a chip based on the most advanced process nodes available today, with Kalray licensing its hardware and software IP developed over the past decade.
Emmanuel Le Roux, CEO of Bull, said: "The development of new interconnect solutions represents a strategic milestone in meeting the needs of future generations of supercomputers and AI infrastructures. The collaboration we've signed with Kalray enables us to develop advanced technologies-particularly for the cloud and AI-and to build high-performance solutions, all while contributing to a strong European drive toward technological sovereignty. In a sector largely dominated by non-European technologies, we are particularly proud to offer a European alternative by producing a chip built on the most advanced technology nodes available today."
Éric Baissus, Chairman of Kalray's Management Board, said: "The signing of this agreement with Bull is further recognition of the relevance of our technological vision and the quality of the innovations developed by our teams. At a time when AI is redefining the needs for computing power, networking, and energy efficiency, we are excited to contribute, alongside a player as prestigious as Bull, to the design of the infrastructure that will shape the digital economy for decades to come. This agreement also illustrates the ability of European leaders to join forces to develop strategic technologies, promote open standards such as Ultra Ethernet, and lay the foundations for sustainable technological sovereignty."
How the partnership works
The agreement combines Kalray's IP licensing model with co-development. Kalray will supply its 'Kalray Networking Engine' subsystem, which has been upgraded to support 1.6 Tbps by incorporating an array of 64 of its next-generation KV5 cores. These cores are RISC-V compatible and designed for high-throughput data processing. The resulting interconnect technology will be integrated into Bull's BXI networking platform, with first large-scale commercial availability slated for 2028.
BXI's evolution and the Ultra Ethernet push
Bull's BXI interconnect technology will evolve to support Ethernet-based protocols and the open security and virtualization management required for AI workloads in cloud environments. Full compatibility with the Ultra Ethernet standard is a core goal; the standard is backed by over 170 industry players including the Ultra Ethernet Consortium (UEC), OCP, and IEEE. Bull is the sole European founding member of the UEC.
A key differentiator is the planned hybrid HPC/AI mode: the network card can switch between BXI's native low-latency, congestion-optimized mode for HPC applications and Ultra Ethernet mode for virtualized AI workloads, using the full 1.6 Tbps bandwidth. Kalray will also be able to offer the jointly developed Ultra Ethernet technology blocks to other companies, broadening the ecosystem around the open standard.
Why this matters for IT and Development
For IT infrastructure teams and network engineers, the move toward open, high-speed Ethernet-based interconnects signals a shift in data center networking. Ultra Ethernet's modern transport, RDMA, and congestion management layers are designed to scale to hundreds of thousands of nodes, directly affecting how AI clusters are architected. The availability of 1.6 Tbps network cards-combined with a hybrid mode that supports both HPC and cloud-native AI-could simplify network design and reduce vendor lock-in.
Professionals looking to build skills in this area can explore the AI Learning Path for Network Engineers, which covers the intersection of AI and networking, including RDMA, congestion control, and Ethernet-based fabrics. The AI for IT & Development tag page also offers a broader set of courses on AI infrastructure, deployment, and management.
Your membership also unlocks:
