Masayoshi Son's SoftBank pledged €75 billion to build 5 gigawatts of AI data center capacity in France, the largest single infrastructure commitment in European artificial intelligence history. The announcement came without ceremony, a deliberate signal that SoftBank views French power grid access and regulatory stability as more valuable than the tax incentives competing jurisdictions offered.

The commitment dwarfs existing European AI infrastructure by a factor of three. Equinix operates 1.2 GW across the continent. Digital Realty holds 1.6 GW. SoftBank's planned capacity exceeds both combined, and the power draw—equivalent to running 4.2 million high-end GPUs continuously—will require direct negotiation with Électricité de France for dedicated nuclear and renewable allocations. Construction timelines were not disclosed, but comparable builds in Northern Virginia took 36 to 48 months from permits to first rack.

This is not SoftBank diversifying. This is SoftBank declaring that Western AI model training will happen in jurisdictions where power is abundant, carbon accounting favorable, and data sovereignty enforceable. France offers all three. The country generates 70% of its electricity from nuclear, avoiding the renewable intermittency problems plaguing German and Dutch data center operators. Macron's government fast-tracked data center permitting in March, cutting approval cycles from 18 months to 9. SoftBank is buying certainty, not optionality.

The allocation math matters for U.S. fund managers tracking SoftBank's capital deployment. The Vision Fund holds $23 billion in dry powder as of last quarter. This €75 billion commitment—roughly $79 billion at current rates—sits outside the fund structure, a direct balance sheet play funded by SoftBank's telecom cash flows and a forthcoming $15 billion credit facility the group has been negotiating with Japanese megabanks since October. Son is betting that owning the physical layer—power, cooling, fiber—generates higher returns than equity stakes in the model developers themselves. He may be correct. Nvidia's H100 clusters lease for $2.50 to $3.80 per GPU-hour, but data center operators with owned infrastructure capture 68% to 72% of that rate, not the 12% to 18% equity investors see in software margins.

Allocators should watch three catalysts. First, SoftBank's credit facility close, expected before Japanese fiscal year-end on March 31. The pricing will reveal whether lenders view AI infrastructure as equipment finance or venture debt. Second, Électricité de France's grid capacity awards, typically announced in Q2. If SoftBank secures 5 GW in a single tranche, smaller operators will face years-long queues. Third, the European Union's AI Act enforcement begins August 2026, and data residency requirements will force U.S. hyperscalers to either build in-region or lease from operators like SoftBank. Lease rates have not been disclosed, but comparable deals in Singapore run 9 to 11 years with inflation-indexed minimums.

SoftBank just made owning European AI compute harder for everyone who is not SoftBank. French power regulators publish grid allocation decisions in May.