Bull speed ahead! Can France re-build AI sovereignty with Bull acquisition?

Against the backdrop of prevailing Silicon Valley wisdom driving the current AI mania, the French government has recently placed a bet on fostering Sovereign AI. Specifically, the French government recently acquired all of Bull's assets in hopes of rebuilding a sovereign AI stack from the ground up. To be clear, it maintains some elements of American tech in its supply chain, such as AMD chips. However, the driving thesis is that the rest of the stack might be better aligned with European interests when stewarded in Europe.

The deal in late March valued Bull at €404 million. The company posted revenues of €720 million in 2025. For reference, Bull has about 3,000 staff, with half in France, 1,600 patents, and thirteen percent of revenue committed to R&D. It also returns Bull to state ownership for the first time since it was sold off to private capital in the 1990s, after a previous 1982 nationalization experiment.

I caught up with Bull CEO Emmanuel Le Roux to unpack the significance of this recent acquisition and its implications for European tech sovereignty. This notion has taken on increasing importance among European enterprise tech leaders and politicians in a pushback against the notion that only Americans know how to do it better. It also reflects some interesting fault lines emerging in the transatlantic relationship -- the US Cloud Act compels American vendors to share customer information with the US Government, and recent moves by US President Donald Trump have dismantled foundations of trust that had shaped US-Europe relations since World War II.

As background, the French Government had previously nationalized Bull in 1982, and that experiment didn't pan out as expected. The company was eventually privatized in the mid-to-late 1990s and spent the following three decades passing through a series of owners, ultimately ending up with Atos before this year's deal returned it to the French State, which is stewarded by the Agence des participations de l'État (APE).  So, what's different this time? Le Roux explains:

The context today is fundamentally different from the past. We are not the same Bull as we were 45 years ago, nor is APE the same agency it was previously. Today, the issue is broader; securing long-term sovereignty over the infrastructure that will underpin AI, high-performance computing and quantum innovation across Europe.

This is not a defensive rescue of an unproven asset. Bull is not a newcomer starting out, but already strongly established with Europe’s only supercomputer manufacturing plant in Angers and leadership in exascale systems such as JUPITER and the upcoming Alice Recoque. With APE’s investment, we are well-positioned to capitalize on opportunities to accelerate innovation. This new chapter will provide an ideal foundation for expanding across the entire AI spectrum, further complementing our systems expertise with components, software platforms and AI use cases.

The lesson learned is that state ownership works best when it provides strategic continuity and long-term investment, rather than short-term political intervention. This acquisition is at the core of APE’s strategy and mission, with a clear emphasis on stability, long-term investment, and critical technologies for the French State and Europe as a whole. The emphasis is on preserving critical expertise, securing industrial capabilities and accelerating future growth.

Controlling the full stack

Under the new stewardship, the French government and Bull management are hoping to build an integrated value chain from design to manufacturing at its Angers HQ. This stands in contrast to the prevailing trend in driving Western computing investments, in which computing hardware is designed in one region, fabricated in another, and assembled in a third.

The thesis seems to be that this approach can leverage lower-cost operations or domain expertise available in other countries. However, even when this approach can drive down short-term costs, it can also erode the long-term value of learning throughout the process. Le Roux elaborates:

In practice, controlling the full stack means Bull can design, integrate, manufacture, deploy and optimize critical layers of the system architecture within Europe, from design and interconnect to server assembly, software layers and deployment services. Our Angers facility, which is the only supercomputer manufacturing plant in Europe and our R&D centers (two in Europe, one in the US, one in India) are major differentiators and give us genuine control over system integration, testing, validation and industrial-scale assembly. On the networking side, technologies such as BullSequana eXascale Interconnect (BXI) exemplify our approach for sovereign European IP and are central to performance at scale.

BXI V3, it is worth noting, is built in collaboration with France's Alternative Energies and Atomic Energy Commission (CEA) and based on the Ultra Ethernet Consortium standard, of which Bull is a member. That qualifies it as sovereign European IP developed inside a broader open standard rather than a closed proprietary fabric.

That said, there are also practical constraints on doing everything in-house. For example, Le Roux acknowledges that dependencies remain at the semiconductor layer, especially for high-end GPUs and some CPU architectures, where the global ecosystem is heavily concentrated outside Europe. He notes:

The strategic objective is not to claim total self-sufficiency, but to maximize European control over the highest-value layers of the stack while progressively strengthening regional component partnerships, such as with SiPearl.

That translates directly into specific hardware: Alice Recoque itself, the next French exascale supercomputer now being assembled in Angers for staged delivery in late 2026 and 2027. It is being built in partnership with AMD, yet Bull controls the design, integration, interconnect and assembly.

One benefit of government ownership is the opportunity to step back and adopt a more long-term, patient-capital approach to value creation, rather than focusing on short-term profit extraction. For example, in Bull's case, they plan to invest thirteen percent of revenue in research and development, along with three hundred data scientists working on practical AI use cases.

In the US, most of the current AI scaling debate seems to be dominated by big AI companies building ever-larger data centers. In contrast, Bull seems to be positioning this differently, closer to what some might call embedded or distributed AI rather than centralized scaling. Le Roux explains that their current focus lies in exploring an alternative to the common discourse driving AI investment in the US:

Today’s discourse is often focused on hyperscale centralized compute, but with energy consumption for AI systems on the rise and closed architectures creating a lack of transparency, many customers need something different: trusted, sovereign and energy-efficient AI infrastructure aligned with specific regulatory and operational requirements. For regulated sectors, the priority is often not simply scale for its own sake, but control, transparency and the ability to deploy AI workloads close to critical data and operational environments.

To address this and ensure the delivery of AI systems that can be fully trusted and as energy efficient as possible, the infrastructure behind those systems must embody these core principles. So, from a Bull perspective, that means dedicating resources to systems that are energy-efficient, interoperable, and transparent. Moreover, for some verticals such as the public sector, research and defence industries, getting all the stacks, including use cases/data sciences, in Europe is critical, and Bull is the only company in Europe that can provide an end-to-end proposal from silicon to use cases.

Planning for turbulence

Bull has already delivered JUPITER, Europe's first exascale supercomputer, and is now building Alice Recoque. Much of the groundwork for both projects was laid before the recent geopolitical uncertainties, such as new and constantly changing US tariffs, export controls, and even the prospect of shifting alliances. Le Roux acknowledges these are all important concerns relating to the goal of building and maintaining an independent supply chain for large supercomputers to meet European sensibilities:

There are undoubtedly geopolitical uncertainties at present, but that only reinforces the fact that Europe cannot afford to rely entirely on non-European infrastructure.  We are working on two levels of “independencies”: The first is obviously to expand the design we are doing, to software, systems (cooling, power supply, electronic boards and components), but also manage our supply chain with 2 or 3 alternative suppliers, if possible, based on different countries, so that we limit the supply chain risk.

One concern is that dependence on overseas hardware ecosystems or a small group of technology providers can be limiting in many ways. A big one is in Europe’s ability to shape systems in accordance with its own regulatory and strategic priorities, including the EU AI Act.

However, the success in delivering JUPITER and soon, Alice Recoque, provides evidence of what is possible when European public institutions and industry leaders unite with a shared vision, transparent leadership and advanced technical expertise. They demonstrate that Europe can build world-class, large-scale and energy-efficient AI and HPC infrastructure.

Despite this progress, dependencies remain, particularly at the semiconductor layer around high-end accelerators and leading-edge chips, where Europe still relies on non-European ecosystems. This is one of the most important strategic challenges ahead. Le Roux explains:

That is precisely why system-level sovereignty spanning design, manufacturing, interconnect, software and deployment matters so much. Even where some component dependencies remain, maintaining control over the architecture, integration and operational stack is critical to reducing risk and ensuring Europe can continue to build and scale sovereign infrastructure. We’re confident that if this level of commitment is maintained, Europe will continue to deliver many more examples of its ability to build large-scale, sovereign AI infrastructure projects.

Since the acquisition, Bull has made a series of announcements that build on its recent pivot. First, it announced plans for 500 new hires concentrated in France, along with smaller hires in Germany and the Czech Republic. It also announced a memorandum of understanding with Dublin-based quantum vendor Equal1 to integrate its quantum technology into the Bull stack. And it launched a five-year, €30 million contract to supply the Mimer AI Factory at Sweden's NAISS national supercomputing facility. This will be the first system to use European AI accelerators from Axelera AI, a European AI chip specialist, with installation scheduled to begin in 2026.

What sovereignty means for enterprise buyers

So, what are the implications of a sovereign AI stack for European companies? What might they be able to do with an AI stack built in the EU that might not work as well with a US or Chinese alternative? Le Roux elaborates:

For an enterprise buyer, sovereignty means control over where workloads run, the capability to choose, design and scale on your own terms and the protection of sensitive use cases. However, sovereignty also offers a strategic advantage: it secures long-term access, reduces geopolitical risks, and allows organizations to develop architectures specifically designed for Europe, based on core values such as openness and trust.

With Bull, customers can deploy AI and HPC environments designed and operated within a European regulatory and industrial framework, aligned with requirements around data residency, operational transparency and long-term supply continuity. In practical terms, that means greater assurance of compliance with frameworks such as the EU AI Act, reduced exposure to geopolitical disruption and confidence that sensitive research, financial modelling or defense-adjacent workloads remain under European control.

The Mimer contract is the first fruit of Bull’s efforts to build a full sovereign AI stack from the ground up, including AI accelerator chips and European software.

Bull has nearly a century of history and survived multiple ownership changes, nationalizations, and near-death experiences. How does this frame Le Roux’s strategy and planning for the future? He says:

What is most useful from Bull’s history is the understanding that strategic computing capability must be built for decades. Nearly a century of innovation has created deep engineering expertise, industrial discipline and long-standing trust with customers in research, industry and the public sector.

The challenge and what keeps us focused is ensuring Europe does not lose control of the technologies that will define the next era of AI, HPC and quantum computing. The risk is not simply commercial; it is strategic dependence on external ecosystems for critical infrastructure.

My take

The current discourse about the future of AI seems to be captured by American sentiments that bigger is necessarily better, and by a certain quality of hubris. It also seems to be shaped by American social media paradigms in which the most outrageous, loudest, and sometimes de-stabilizing narratives are increasingly finding their way into the public discourse about what AI is to become.

As a consequence, the trade press seems rife with interesting pronouncements that AI will cause a SaaS-pocalypse or a security apocalypse, which seem more like veiled attempts at fear-driven marketing than productive debates about a future of AI that actually wants and needs us. The fact that France has decided to take meaningful control of its future seems significant. A few hundred million-euro investment in building its own base of AI craftsmanship across the entire stack seems like a pittance compared to the hundreds of billions of dollars big AI believes we need to build to support some vague AI-driven future. That simultaneously seems to exclude the value of human craftsmanship, sovereignty, and uniquely European concepts like the right to control your own data that seem increasingly edited out of the rhetoric emerging from Silicon Valley.

All that said, the history of governments buying up private companies is rife with examples in which government stewardship didn't pan out well. But perhaps this time, driven by turbulence in the global supply chain around AI infrastructure, and by a CEO who is unusually willing to name the limits of his own sovereignty pitch, things might be different.