€75 billion. That number makes every prior European AI announcement look like a pilot program. SoftBank's June 2026 commitment to build AI infrastructure in France — announced during President Macron's third annual AI summit in Paris — is larger than the entire EU AI Act compliance budget, larger than Microsoft's €4B French investment from February 2025, and represents roughly 15% of SoftBank's total assets being directed at a single geography.
The headline number covers a 4-year deployment through 2030. The French government is providing subsidized land, accelerated permitting, and what officials described as "sovereign data partnership" terms that give France priority access to compute capacity hosted on its own soil. In exchange, SoftBank gets favorable tax treatment and a fast track through France's historically slow infrastructure approval process.
What Is Actually Being Built
The €75B breaks down into three components per the joint press release:
| Component | Planned capacity | Timeline | Approximate budget allocation |
|---|
| GPU data centers (AI training + inference) | 3.2 GW | 2026–2028 | ~€45B |
| Fiber and edge network buildout | Coverage across 12 French regions | 2027–2029 | ~€18B |
| R&D facilities + talent programs | 8 research centers, 50,000 AI jobs target | 2026–2030 | ~€12B |
The 3.2 GW of GPU data center capacity deserves framing. The entire data center capacity of the EU today is approximately 12–14 GW across all workloads. This single commitment would expand European AI-specific compute by roughly 25% when fully built. For context: a 100MW GPU cluster can run approximately 5,000 H100 GPUs continuously. 3.2 GW is 160,000 H100 equivalents.
Why France and Why Now
Three factors pushed France ahead of Germany and the Netherlands, which both competed for the investment:
Nuclear power surplus. France gets 70% of its electricity from nuclear, with wholesale electricity prices averaging €45/MWh in Q1 2026 — roughly half Germany's €89/MWh average. AI data centers are electricity-intensive at a scale that makes per-MWh cost the single largest operational variable. SoftBank's internal models reportedly showed a €2.1B annual operating cost advantage versus equivalent German infrastructure.
EU AI Act compliance base. France is one of four EU nations with a fully operational AI regulatory sandbox under the AI Act. Building infrastructure in France means any model trained or fine-tuned there carries a compliance provenance that satisfies Article 53 requirements for high-risk AI systems. For SoftBank portfolio companies that serve EU enterprise customers — ARM, Coupang, ByteDance-adjacent ventures — having a France-based training infrastructure is a compliance asset, not just a cost play.
Masayoshi Son's Stargate thesis. SoftBank's $500B US Stargate commitment, announced in January 2026, was followed within months by parallel announcements in Japan (¥20T), the UAE ($14.5B), and now France. The pattern is clear: Son believes AI infrastructure is the defining capital allocation of the 2026–2035 decade and is building geographically distributed compute at scale in every jurisdiction that offers favorable terms.
Impact on EU AI Sovereignty
The EU has been explicit about its concern with hyperscaler dependency. The EU Cloud and AI Sovereignty guidelines published in March 2026 state that member states should have "access to domestically controlled compute for critical AI functions by 2028." France hosting 3.2 GW of SoftBank-controlled compute does not fully satisfy that criterion — SoftBank is Japanese, not European.
The French government addressed this directly. The joint announcement included a "sovereign access guarantee": France retains the right to requisition 20% of total capacity for government and public sector AI workloads during declared national emergencies, with pricing locked at pre-agreed rates. This is the same structure the UK used in its 2025 Microsoft agreement for National Health Service AI infrastructure.
For practical EU AI sovereignty purposes, this changes the calculus. French AI startups and enterprises previously had two options for compliant EU compute: use AWS/Azure/Google EU regions (which physically sit in EU but are US-controlled) or use small EU cloud providers with far less capacity. SoftBank's French buildout creates a third option: large-scale, EU-territory compute with explicit data residency guarantees backed by government agreement rather than just contractual terms.
What It Means for Developers
The practical implications arrive in two waves.
Near term (2026–2027): Nothing changes immediately. Data centers take 18–24 months to build from groundbreaking. The first phase of SoftBank's buildout will not be operational until late 2027 at the earliest. Developers should continue using existing EU-region cloud providers for GDPR-compliant workloads.
Medium term (2028+): This is where the developer implications become concrete:
EU inference pricing. Today, running inference via Anthropic, OpenAI, or Mistral in EU regions costs roughly 15–20% more than US-East due to lower competition and higher infrastructure costs. 3.2 GW of new compute entering the EU market by 2028–2029 will compress that premium. Conservative estimates suggest EU inference costs drop 25–40% versus today's rates by 2030.
Data residency without compromises. GDPR Article 44 restrictions on data transfers currently push some EU companies into architectures where they cannot use the best US-hosted models for personal data processing. French-hosted models eliminate that constraint entirely — personal data never leaves EU jurisdiction.
Model training for EU-specific use cases. Fine-tuning models on EU languages, regulatory documents, and sector-specific data requires EU-resident compute to satisfy data processing agreements. Most EU enterprises currently cannot do this at scale. 3.2 GW of France-based GPU capacity makes it economically viable by 2028.
The EU developer community has a parallel tool in the WOWHOW tools suite for estimating compute costs across cloud providers — worth bookmarking as EU pricing evolves over the next 18 months.
Comparison with US Hyperscaler EU Spending
SoftBank's €75B dwarfs individual announcements but sits within a broader acceleration:
| Company | EU AI infrastructure commitment | Announced |
|---|
| SoftBank (France) | €75B | June 2026 |
| Microsoft (EU-wide) | €14B | January 2026 |
| Google (EU-wide) | €9B | March 2026 |
| AWS (EU-wide) | €7.8B | February 2026 |
| Mistral AI (France, raised + ecosystem) | €600M | Ongoing 2025–2026 |
The combined EU AI infrastructure commitment from major players now exceeds €106B over 2026–2030. That is real infrastructure being built, not announced intentions. European developers who assumed they would always be second-class citizens on AI compute have roughly 18 months before that assumption needs updating.
People Also Ask
What will SoftBank build in France with the €75 billion?
Three main areas: 3.2 GW of GPU data centers for AI training and inference (the largest allocation at ~€45B), fiber and edge networking across 12 French regions (~€18B), and 8 research centers targeting 50,000 AI-related jobs (~€12B). Full buildout runs through 2030.
Does SoftBank's France investment affect GDPR data residency?
Yes, when operational. Infrastructure physically located in France and covered by French government sovereign access agreements satisfies GDPR Article 44 restrictions on personal data transfers. EU companies would be able to process personal data using French-hosted models without the transfer restrictions that currently apply to US-hosted services.
How does this compare to Stargate in the US?
SoftBank's US Stargate commitment was $500B (approximately €460B) announced in January 2026. The France commitment at €75B is the single largest ex-US AI infrastructure deal but about one-sixth the scale of Stargate. Son has described this as a parallel global buildout, not a reallocation from Stargate.
When will developers see cheaper EU inference pricing?
The first SoftBank data centers in France will not be operational until late 2027 at the earliest. Meaningful pricing competition from this new capacity is a 2028–2030 story. EU inference costs are already declining due to Mistral AI's EU-hosted API and Google's 2026 EU region expansions — but the large-scale impact from SoftBank's buildout is medium-term, not immediate.
Originally published at wowhow.cloud
Top comments (1)
This is a massive signal of how aggressively Europe is trying to secure AI infrastructure sovereignty. A €75B commitment for 5GW of data center capacity isn’t just an infrastructure play—it’s geopolitical positioning in the global AI stack. What stands out is how energy strategy (especially France’s nuclear base) is becoming a key competitive advantage in attracting compute-heavy investments. I also find it interesting that capital is flowing into “picks and shovels” rather than model development itself. The real question is whether this translates into European AI capability, or if it mainly becomes infrastructure serving external model providers. Either way, it reshapes the landscape.