Bull and Kalray partner on 1.6 Tbps AI networking silicon

Bull and Kalray have announced a strategic technology partnership to develop next-generation networking silicon for artificial intelligence and high-performance computing infrastructure. The agreement combines IP licensing, co-development and services, with the first commercially available products expected to reach the market at scale in 2028.

At the core of the deal, Kalray will licence its software and hardware technologies to Bull, including the Kalray Networking Engine subsystem. That subsystem has been enhanced to support bandwidth of up to 1.6 Tbps and is powered by 64 next-generation KV5 processor cores built on the RISC-V open instruction-set standard. Bull will integrate these IP blocks into an evolved version of its BXI interconnect fabric, adding cloud-native capabilities, security, virtualisation support, and compatibility with the Ultra Ethernet standard.

Bull is a founding member of the Ultra Ethernet Consortium, an industry body that now counts more than 170 member companies spanning networking, hyperscale computing and AI ecosystems. The consortium is developing an open networking standard designed specifically for large-scale AI data-centre deployments, where proprietary fabrics have historically created vendor lock-in and scaling constraints.

The deal

For Kalray, the agreement accelerates a licensing-led business model the Grenoble-listed fabless semiconductor company has been building for over a decade. Its MPPA manycore processor architecture, used here in the KV5 cores, is designed to handle parallel data streams without bottlenecks, positioning it as a complement to traditional GPU and CPU-based compute. The partnership gives Kalray a route to volume deployment through Bull's existing customer base in supercomputing, research organisations and enterprise AI.

Emmanuel Le Roux, chief executive of Bull, said the collaboration enables the two companies to deliver "a European alternative by producing a chip that incorporates the most advanced technological innovations available today" in a sector he described as largely dominated by non-European technologies.

The resulting interconnect solutions are intended to deliver a hybrid approach: ultra-low latency and advanced congestion management for HPC workloads, alongside Ultra Ethernet compatibility for AI and virtualised environments.

Market context and competitive positioning

The high-speed AI networking market is increasingly contested. NVIDIA's InfiniBand fabric, delivered through the Mellanox assets it acquired in 2020, currently holds a dominant position in GPU cluster interconnects, while Broadcom's Ethernet switching silicon and its own custom AI networking ASICs serve a large portion of hyperscaler buildouts. Cisco, Arista and a growing set of startups are also positioning for the shift toward Ethernet-based AI fabrics.

The Ultra Ethernet Consortium's progress represents a meaningful counter-weight to proprietary approaches. If the standard achieves broad interoperability, it reduces switching costs for buyers and lowers the barrier for new entrants, including European suppliers. Bull and Kalray are explicitly pitching European digital sovereignty as a differentiator, a positioning that resonates with public-sector procurement programmes in France and across the EU, where the European Chips Act and associated funding instruments favour domestically developed semiconductor IP.

Kalray's RISC-V-based core architecture also aligns with a broader European effort to build processor IP independent of US-origin instruction-set licences, an increasingly salient consideration given ongoing US export-control actions targeting advanced semiconductor technology.

With a 2028 commercial launch target, both companies will need to sustain R&D momentum through a period when hyperscalers and well-capitalised American rivals are moving quickly. Investors and customers will be watching for prototype performance data, named early adopters, and clarity on the OEM licensing pipeline that the two companies say will accompany the product.