Bull and Kalray: in AI infrastructure the choice that binds you most is the network

On 28 July 2026, from Grenoble, Bull and Kalray (Euronext Growth Paris: ALKAL) announced a strategic partnership on networking technologies for AI and high-performance computing infrastructures. The agreement, a footnote specifies, had already been announced by Kalray on 16 July.

Beneath the announcement lies a lesson that goes beyond the two companies. When you buy AI infrastructure, the choice that binds you most is not the model, and it is not even the accelerator. It is the interconnect: the fabric that links the nodes to one another. You choose it once, it lasts as long as the data hall, and it determines which accelerators you can install tomorrow and from whom you will have to buy them.

What was announced

Kalray is a fabless company — it designs chips, it does not make them — registered in Montbonnot-Saint-Martin, near Grenoble (RCS Grenoble 507 620 557). Its product is the MPPA® processors, which it classifies as DPUs, for applications handling massive data flows: data centres, “AI Gigafactories”, telecoms, aerospace, defence.

Bull describes itself as a global player in high-performance computing, AI and quantum technologies: around €720 million in revenue, 3,000 professionals in 32 countries, 1,600 patents. One fact the release does not recall: since 31 March 2026 Bull is no longer part of the Atos group — the French State acquired 100% of its capital, at an enterprise value of up to €404 million.

Kalray licenses hardware and software technologies and intellectual property blocks, in particular the Kalray Networking Engine, a subsystem modified to sustain 1.6 Tbps and built on an array of 64 KV5 cores compatible with the RISC-V standard. Bull integrates them into the evolution of BXI, its proprietary network, towards the Ethernet standard and Ultra Ethernet compatibility. The model is intellectual property licensing, co-development and services; the first solutions are expected to be commercially available at scale in 2028. Amount, duration and exclusivity are not disclosed: this is an announcement of industrial intent, not a contract already in operation.

Emmanuel Le Roux, CEO of Bull, speaks of a contribution “to a strong European drive toward technological sovereignty”; Éric Baissus, Chairman of Kalray’s Management Board, of “the foundations for sustainable technological sovereignty”. The parties’ own positions, reported as such.

Why the interconnect decides

Training or serving a large model does not mean putting one accelerator to work: it means making tens or thousands of accelerators talk continuously, exchanging gradients, weights and activations at every step. Bandwidth and latency between nodes determine how much of the compute you bought is actually used. If the fabric is slow or congested the accelerators wait: expensive hardware, idle for a significant fraction of the time you paid for.

The least visible lock-in

A proprietary interconnect ties the buyer to a single supplier for the entire life of the data hall. You do not change it with an upgrade: you replace adapters, switches, cabling and often the nodes. It is a dependency that shows up in no accelerator price comparison — memory, peak performance and power draw get compared, the network is taken for granted — and that surfaces at the first expansion.

The most eloquent fact in the release is this, and it is not us saying it: Bull is moving its own proprietary interconnect towards an open standard. The technical description is explicit: a hybrid approach, BXI mode for the extreme latency of HPC and Ultra Ethernet mode for virtualised AI workloads, up to 1.6 Tbps per card. According to SDxCentral, the third generation of BXI will be the first Bull offering to use Ethernet in place of the proprietary link protocol, targeting eight million endpoints against the 64,000 nodes of the previous one. When the owner of a proprietary fabric puts a standard on top of it, the market’s direction has been stated.

What Ultra Ethernet is, and where the limit lies

The Ultra Ethernet Consortium is a standards organisation incorporated under the Linux Foundation. The aim: to make Ethernet — an already existing multi-vendor ecosystem — suitable for the AI and HPC workloads that until now required dedicated fabrics. The public 1.0.3 specification, dated 16 July 2026, runs to 573 pages and covers the whole stack, from the software API to the physical layer, plus management, compliance and interoperability. It defines three transport profiles — AI Base, AI Full, HPC — and does not require replacing the Ethernet switches already on the market. The Bull-Kalray release speaks of more than 170 players and contributors, with the goal of replacing RoCE v2 and the proprietary InfiniBand network.

For a buyer: an open standard makes it possible to put network suppliers in competition and to avoid tying the data hall to a single stack.

The limit must be stated just as clearly. A standard is not interoperability: it becomes so when certified products from several vendors genuinely talk to each other. The specification distinguishes mandatory requirements from optional features — two “compliant” products may not implement the same things. And here the products are expected in 2028: until then, openness is a commitment, as with accelerators.

RISC-V: a third layer, not a guarantee

The 64 cores of the networking engine are RISC-V compatible, an open instruction set architecture: no licence is paid on the instruction set, and extending it does not depend on a single owner’s roadmap. For a buyer of infrastructure that is one more point in favour of long-term continuity. No more: the foundry, the development tools and the supply chain remain. An open architecture does not make the product sovereign — it moves a dependency, it does not remove it.

What to write in the tender

A data hall is paid for in accelerators and used in bandwidth. Five questions on the interconnect fabric, to be put in writing before signing.

Which standard and which version, with the revision number. “Ultra Ethernet compatible” without a version is not a requirement.
Who certifies interoperability, under which test programme and against which other vendors: the body and the date of the test, not a supplier declaration.
What happens if you change accelerator supplier mid-life: do adapters and switches remain usable, or does expansion force you to buy the fabric again?
Who owns the adapter firmware and on what terms security updates keep arriving if the relationship ends or the supplier changes ownership.
How effective node-to-node bandwidth is measured at acceptance, not the declared figure: which test, which message size, how many nodes communicating simultaneously, which threshold triggers non-conformity.

The contract rule applies: a limit that is neither written nor verifiable does not exist.

Why this matters to us

Anyone who chooses to keep AI inside their own perimeter also buys the infrastructure that runs it, and that choice lasts longer than the model running on top: models are swapped in months, a network fabric stays for the life of the hall. And announced capacity is not available capacity.

That is why a dedicated AI system must be delivered in two modes, never just one: on-premise in the client’s own environment, or CSIDIA dedicated cloud — an environment reserved for the single client, access over a dedicated VPN, data centre in Italy, premises staffed directly by us. Infrastructure choices are written down beforehand, not discovered afterwards. Method matters more than the brand of the component.

Do you need to write an AI infrastructure tender, or assess one you have been offered? Let us talk it through in thirty minutes.