Bull and Foxconn will manufacture the Nvidia Vera Rubin NVL72 platform in France: Europe takes control of its AI infrastructure
🔎 VivaTech 2026: Europe attempts its industrial AI catch-up
On June 17, 2026, Jensen Huang takes the stage at GTC Paris during VivaTech with a compelling argument. Exactly one year ago, he promised over 20 AI factories in Europe. This time, he is no longer making promises: he is parading factories.
Bull (Eviden/Atos) and Foxconn commit to developing and manufacturing the NVIDIA Vera Rubin NVL72 platform on European soil. The systems will first be tested in Foxconn facilities in the Czech Republic, then assembled and validated at the Bull factory in Angers, France.
The announcement is spectacular on paper. The Vera Rubin NVL72 is Nvidia's 3rd-generation rack-scale AI supercomputer: 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs, all connected by the 6th-generation NVLink fabric in a liquid-cooled rack. It is the tool that all AI players want to run their heaviest models — from OpenAI's GPT-5.5 to Anthropic's Claude Opus 4.7.
Except there is a catch. And it's a major one.
The essentials
- Bull and Foxconn will produce the NVIDIA Vera Rubin NVL72 in Europe, with final assembly in Angers and testing in the Czech Republic, according to the official NVIDIA Blog announcement (June 2026).
- The Vera Rubin NVL72 promises 5x higher inference performance and a 10x lower cost per token compared to the Blackwell architecture, with availability expected in the second half of 2026.
- France is attracting these investments thanks to its cheap nuclear energy and talent pool, but "European manufacturing" hides a more nuanced reality: assembly, not chip production.
- Jensen Huang takes stock of a year of promises: more than 20 AI factories announced for Europe, with Mistral AI as the designated champion of sovereign computing.
Recommended tools
| Tool | Main usage | Price (June 2026, check on site.com) | Ideal for |
|---|---|---|---|
| Hostinger | Web hosting for AI projects | Starting from 2,99 €/month | AI developers and startups |
| NVIDIA AI Enterprise | Software suite for AI datacenters | On quote | Companies deploying Vera Rubin NVL72 |
| Mistral AI Platform | Model inference and fine-tuning | Pay-as-you-go | European companies wanting sovereign |
The Vera Rubin NVL72: What Europe Will Actually Assemble
The NVIDIA Vera Rubin NVL72 platform is not just a simple server. It is a rack-scale supercomputer designed for the agentic era, where models like GPT-5.5 or Claude Opus 4.7 (Adaptive) require thousands of GPUs to run in production.
According to specifications detailed by StorageReview at CES 2026, the NVL72 unifies in a single rack: 72 Rubin GPUs equipped with HBM4 memory, 36 next-generation Vera CPUs, ConnectX-9 SuperNICs for networking, and BlueField-4 DPUs for infrastructure data processing. All of this is connected by 6th-generation NVLink fabric, with integrated liquid cooling.
GenMediaLab reports that Nvidia promises 5x higher inference performance and a 10x lower cost per token compared to the Blackwell architecture. This is precisely the generational leap that AI factory operators are waiting for to make their infrastructures profitable.
The connection to the NVIDIA Vera Rubin and N1X ARM architecture presented at GTC Taipei is direct: it is the same silicon base, now available in a rack-scale version for European datacenters. What changes is the assembly location.
Industrial assembly: Angers and the Czech Republic, but no foundry
The NVIDIA Blog press release (June 2026) is precise about the production line. The Vera Rubin NVL72 systems will be "manufactured and initially tested at Foxconn facilities in the Czech Republic, then assembled, integrated, and fully validated at the Bull plant in Angers, France."
This is a two-stage assembly process. Foxconn brings its mass production logistics. Bull brings its expertise as a European integrator and its final validation capacity. The end result bears the Bull brand — a strong symbol for the French and European public sector.
The Bull plant in Angers already exists. It has been assembling servers for decades for government agencies and businesses. But let's be clear: we are talking about component integration, not semiconductor manufacturing. The Rubin GPUs come from TSMC in Taiwan. The HBM4 comes from Samsung or SK Hynix in South Korea. The Vera CPUs are designed by Nvidia, probably manufactured by TSMC.
The European value added lies at the level of system integration, liquid cooling, NVLink cabling, firmware, and validation. This is honorable, but it is not silicon sovereignty. We will return to this.
Euronews points out that this announcement is part of a series of major agreements signed at VivaTech 2026 between Foxconn, Nvidia, and Mistral AI, consolidating France's position as the continent's AI infrastructure hub.
Why France? Nuclear power and talent, the two pillars of attractiveness
The Euronews article (June 18, 2026) asks the blunt question: why France, and not Germany or the Netherlands? The answer comes down to two words that Americans and Taiwanese understand very well: energy and brains.
The French nuclear fleet provides decarbonized electricity at a competitive price — about 40% cheaper than the European average according to 2025 Eurostat data. For an AI datacenter that consumes as much as a city of 50,000 inhabitants, the energy bill makes or breaks profitability.
This is all the more critical given that the Vera Rubin NVL72 is a liquid-cooled rack, power-hungry but thermally more efficient than air. The current context of tension on European power grids makes the stability of the French energy mix strategic.
The second pillar is talent. France trains around 100,000 engineers per year, with a strong concentration in applied mathematics and computer science. Research laboratories (INRIA, CNRS) directly feed the teams at Mistral AI, Hugging Face, and now the Bull-Nvidia integration teams.
The contrast with Germany is revealing. German industry remains polarized around automotive and traditional industry. France made an explicit bet on digital technology and AI starting in 2017, with the Villani plan followed by successive investments. The result can be measured today in infrastructure contracts, not in speeches.
Jensen Huang at VivaTech: Taking Stock of a Year of Promises
Keep the context in mind. In June 2025, Jensen Huang came to Paris for the first GTC Paris. He promised more than 20 AI factories in Europe and designated Mistral AI as the champion of the continent's sovereign computing, with factories powered by the Blackwell architecture.
A year later, the GTC Paris keynote at VivaTech (June 17-20, 2026, Paris Expo Porte de Versailles) serves as a progress report. TechTimes (June 10, 2026) describes it as an accounting exercise: how many of these 20 factories actually exist?
The answer is mixed. Several AI factories are operational or in the process of being deployed, initially powered by Blackwell, with a planned migration to Vera Rubin. But the figure of 20 seems to include projects at various stages of maturity, not just concreted and cabled datacenters.
The Bull-Foxconn announcement changes the game because it adds an industrial component to the purely real estate aspect. It is no longer just "we are going to build buildings to house American servers" — it is "we are going to assemble these servers in Europe."
Mistral AI, for its part, is signing major infrastructure deals during this same VivaTech. The European champion, which runs some of the highest-ranked models (its architecture positions itself competitively against GPT-5.4 and Claude Sonnet 4.6 on generalist benchmarks), needs computing power for training and inference. These agreements with the Nvidia-Bull-Foxconn ecosystem guarantee it priority access to Vera Rubin racks.
Assembly vs. sovereignty: the geopolitical limit of the announcement
This is the critical point that no one really settled at VivaTech. When Bull "manufactures" the Vera Rubin NVL72, what does this word actually mean?
The value chain of an AI supercomputer breaks down schematically as follows:
| Stage | Actor | Location | Share of value |
|---|---|---|---|
| Architecture design | NVIDIA | Santa Clara, USA | ~35% |
| Chip manufacturing (GPU, CPU) | TSMC | Hsinchu, Taiwan | ~30% |
| HBM4 memory | Samsung / SK Hynix | South Korea | ~15% |
| Rack integration, cooling, validation | Bull / Foxconn | Angers (FR) / Czech Republic | ~12% |
| Software (CUDA, drivers, stack) | NVIDIA | USA + global | ~8% |
Europe takes on the integration stage. It is necessary — without local integration, there is no rapid deployment, no reactive maintenance, no optimized GDPR compliance. But it is the share with the lowest barriers to entry and the lowest margin.
The comparison with other sovereign strategies is illuminating. South Korea opted for a vertically integrated approach: LG provides the datacenter infrastructure, SK Telecom operates the network, SK Hynix produces the HBM memory, and Naver develops the models. Each link in the chain has a national champion.
India, for its part, is deploying "sovereign AI factories" with a logic of data and model control, relying on partnerships with Nvidia but demanding strong local anchoring. Data Center Dynamics reports, moreover, that GMI Cloud and Magna AI are building a global network of sovereign AI factories in Malaysia, Belgium, and Romania, also based on the Vera Rubin NVL72 architecture — showing that European demand is real but not exclusively French.
The French paradox is as follows: we are celebrating an "AI sovereignty" that relies on Taiwanese chips, a Californian architecture, and Korean memory, assembled in Maine-et-Loire. It is better than importing everything fully assembled. But it is not the sovereignty that South Korea or Japan are striving to build.
What the Vera Rubin NVL72 concretely changes for European users
Beyond geopolitics, the arrival of the Vera Rubin NVL72 in Europe has immediate practical consequences for businesses and developers on the continent.
Firstly, latency. Running a GPT-5.5 or Claude Opus 4.7 inference on a server based in Northern Virginia adds 80-120 ms of network latency per request. With local Vera Rubin racks, this latency drops below 10 ms. For real-time applications — AI agents, algorithmic trading, robotics — this is the difference between a viable product and a prototype.
Secondly, compliance. The GDPR imposes strict restrictions on the transfer of personal data outside the EU. Having Vera Rubin infrastructure physically located in France, operated by Bull (a European company), significantly simplifies contractual negotiations with DPOs. No international transfer clauses, no Privacy Shield to renegotiate.
Thirdly, cost. The 5x higher inference performance promised by the Vera Rubin compared to Blackwell concretely means that the same computing budget processes 5x more tokens. For a company doing fine-tuning on DeepSeek V4 Pro or deploying Claude Sonnet 4.6 in production, the economic equation changes radically.
The parallel with the global infrastructure dynamic is also relevant. When SpaceX becomes an AI infrastructure provider for Google with 110,000 NVIDIA GPUs for 920 million dollars per month, we can measure the scale at which American players are operating. Europe will not catch up to this scale with a single Bull factory. But it can position itself in the segments where geographical proximity is a decisive advantage.
Foxconn, Bull, and "physical AI": the link with robotics
The Vera Rubin NVL72 announcement should not be read in isolation. At this same VivaTech 2026, Foxconn presented its first humanoid robots in Europe, unveiling its "physical AI closed-loop" stack which combines simulation, GPU training, and physical deployment.
TechTimes (June 17, 2026) details how this stack precisely uses computing infrastructures like the Vera Rubin NVL72 for training robotics models. The massive parallelism of the rack (72 GPUs in NVLink) is suited to the complex physical simulations required by humanoid robotics.
The industrial setup then takes on a broader meaning. Foxconn is not coming to Europe merely to assemble servers. It is coming to deploy a complete ecosystem: AI factories to train the models, and robots to execute them in factories. Bull becomes the local integrator for this ecosystem.
This is a strategy of industrial encirclement. Foxconn controls robot manufacturing, Nvidia provides the computing, Bull ensures the European anchor. The factory of the future — the one Foxconn envisions — requires both: AI brains in Vera Rubin racks, and robotic arms on the production line.
For France, it is a risky but coherent bet. Accepting a dependency on US/Taiwanese silicon in exchange for an industrial anchor that creates highly skilled jobs in Angers, systems integration expertise, and an advanced robotic ecosystem.
Mistral AI's Role: Sovereign Champion or Captive Client?
Mistral AI occupies an ambiguous position in this architecture. Designated by Jensen Huang as the "champion of European sovereign computing" in June 2025, the French company signed new infrastructure agreements at VivaTech 2026. These agreements tie it even closer to Nvidia's CUDA-Vera Rubin ecosystem.
On one hand, this makes sense. Mistral AI's models, which rival GPT-5.4 (generalist score of 89) and Claude Opus 4.6 (87) on benchmarks, require thousands of GPUs for training and inference. The Vera Rubin NVL72 offers the best performance/cost ratio on the market. Not using it would be a competitive disadvantage.
On the other hand, dependency on CUDA locks Mistral into the Nvidia ecosystem. If Nvidia changes its commercial terms, licensing rates, or delivery priorities, Mistral has no Plan B in the short term. Alternatives (AMD MI400, Google TPU v6) exist but are not drop-in replacements.
This is the paradox of European AI sovereignty as it is taking shape in 2026: the designated champion depends on the American provider for its infrastructure, the Taiwanese manufacturer for its chips, and the Franco-Taiwanese integrator for its servers. Mistral's only true sovereignty lies in its model weights — its training data and algorithmic expertise. The rest is an assumed geopolitical dependency.
❌ Common mistakes
Mistake 1: Confusing assembly and semiconductor manufacturing
The Bull plant in Angers assembles servers from pre-fabricated components. It does not etch chips. Presenting this announcement as "Europe makes its own AI chips" is factually incorrect. The Rubin GPUs are produced by TSMC in Taiwan. The distinction is crucial for assessing the actual level of sovereignty achieved.
Mistake 2: Thinking that Vera Rubin solves the energy problem
The rack consumes more than the Blackwell in absolute terms — that is the price of 5x higher performance. Liquid cooling reduces the climate footprint compared to air, but the total energy bill of an AI factory remains colossal. The French nuclear advantage does not eliminate the need for software optimization and energy efficiency.
Mistake 3: Believing that 20 AI factories are operational
Jensen Huang's figure includes projects that are announced, under construction, and operational. Concluding that "Europe has caught up" based on this number is a misinterpretation. The actual deployment pace is more modest, and the transition from Blackwell to Vera Rubin will slow down some commissionings.
❓ Frequently Asked Questions
Where exactly will the Vera Rubin NVL72 be assembled?
The components are first tested in Foxconn facilities in the Czech Republic, then final assembly and full validation take place at the Bull factory in Angers, France, according to the official NVIDIA Blog (June 2026).
What is the difference between Blackwell and Vera Rubin?
The Vera Rubin NVL72 promises 5x higher inference performance and a 10x lower cost per token compared to Blackwell, featuring Rubin GPUs, Vera CPU, HBM4 memory, and 6th-generation NVLink, versus HBM3e and NVLink 5 for Blackwell.
Has Europe become independent from the United States for AI?
No. The chips are designed by Nvidia (USA) and manufactured by TSMC (Taiwan). Europe is gaining integration and local deployment capabilities, which is significant, but the silicon dependency remains total.
When will the first Vera Rubin racks be available in Europe?
Nvidia plans a deployment in the second half of 2026. The first deliveries to European customers (including Mistral AI) are expected in late 2026 or early 2027, according to GenMediaLab.
How does this announcement impact French developers?
Reduced latency for API calls, simplified GDPR compliance with local datacenters, and potentially lower infrastructure costs thanks to the efficiency gains of the Vera Rubin. Developers using models like Claude Sonnet 4.6 or DeepSeek V4 Pro via European platforms will benefit directly.
✅ Conclusion
The Bull-Foxconn announcement for the Vera Rubin NVL72 in France is a real industrial step, not a PR stunt. Europe gains world-class integration capacity in Angers, a latency shortcut for its companies, and a GDPR compliance argument that has concrete market value. But calling this "AI sovereignty" is an understatement: as long as the chips come from Taiwan and the architecture from California, Europe will assemble, not manufacture. To follow the evolution of this infrastructure and its implications for AI in France, the next chapter will play out in Mistral AI's upcoming contracts and future GTCs.