SoftBank's 75 Billion Euro France Bet: The Energy Wall Meets Sovereign Compute

On May 30, 2026, at the Choose France summit in Versailles, SoftBank Group announced a commitment to develop and operate 5 gigawatts of AI data-center capacity in France, an investment that could reach 75 billion euros (roughly 87.5 billion dollars). The headline traveled fast: SoftBank's largest AI-infrastructure commitment in Europe, thousands of high-skilled jobs, three campuses rising across the Hauts-de-France region. The first phase alone carries a 45 billion euro price tag and a target of 3.1 gigawatts energized by 2031, anchored at Dunkirk's Loon-Plage, plus sites at Bosquel and Bouchain.

If you read only the press release, you would conclude that the bottleneck on artificial intelligence is money, and that SoftBank has solved it. That conclusion is wrong, and the way it is wrong tells you almost everything about where the AI build-out actually stands in 2026. The constraint is not capital. Capital is abundant and getting more so. The constraint is energy: specifically, the number of gigawatts that can be physically energized, cooled, and connected to a transmission grid within a commercially viable window. The gap between gigawatts-announced and gigawatts-energized is the most important number in the industry, and almost nobody puts it on a slide.

This article is about that gap. It is about why SoftBank chose France instead of Texas or Ireland, why nuclear baseload has quietly become the most strategic asset in the AI economy, why a 45 billion euro check does not buy you a 2027 data center, and why "sovereign compute" stopped being a slogan and became an industrial-policy project that the European Union is now funding directly. The SoftBank announcement is a useful lens precisely because it is so large. Five gigawatts is roughly half of France's entire net export capacity. When a single private commitment is sized against a country's power exports, you have left the world of cloud procurement and entered the world of energy geopolitics.

What SoftBank Actually Committed To

Let us be precise about the numbers, because the press coverage blurred them. SoftBank committed to 5 GW of capacity for an investment of "up to" 75 billion euros. The word "up to" is load-bearing. The firm commitment is the first phase: 45 billion euros for 3.1 GW, delivered by 2031, across three named sites in Hauts-de-France. The remaining 1.9 GW and the additional 30 billion euros are described as subject to extension, conditional on the first phase performing and on additional sites being identified across France.

That structure matters. A 45 billion euro firm tranche delivering 3.1 GW by 2031 implies a build cadence of roughly 500 megawatts per year over six years, assuming a 2026 start. That is aggressive but not fantastical. It is also entirely dependent on a chain of physical and regulatory events that the capital commitment does not control: land permits, transmission-grid connection, cooling-water authorization, equipment supply, and the gradual commissioning of power blocks. The money is the easy part.

The partner roster is more revealing than the dollar figure. Schneider Electric will integrate power modules and build a power-module integration facility at the Port of Dunkirk alongside a SoftBank-operated enclosure-manufacturing facility. EDF, the French state nuclear utility, will supply electricity for the Bouchain facility. SB Energy supports development. Local universities and engineering schools anchor R&D. This is not a hyperscaler renting space; it is a vertically integrated industrial cluster that manufactures its own enclosures and power blocks next to a deep-water port. SoftBank is not buying a data center. It is building a factory that makes data centers, and locating it where the electrons are clean and the harbor is deep.

Why France: Nuclear Baseload as a Strategic Asset

The single best explanation for why this 87.5-billion-dollar bet landed in northern France rather than in the United States is the French grid. About 70 percent of French electricity comes from nuclear power; on any given day the figure ranges between roughly 60 and 75 percent. That is not a marketing line. It is the foundation of the entire proposition, and it produces three advantages that no American site can currently match.

The first advantage is dispatchable, low-carbon baseload. AI training clusters do not want intermittent power; they want flat, predictable, 24/7 electricity at utilization rates above 90 percent. Nuclear delivers exactly that. A wind-and-solar grid with gas backup can match the megawatt-hours on paper but not the carbon profile or the round-the-clock reliability without expensive storage. France can.

The second advantage is headroom. France was Europe's leading net electricity exporter in 2025, with a record 92.3 TWh exported, equivalent to a sustained capacity of just over 10 GW flowing out of the country. President Macron made the political subtext explicit in early 2026: "Thanks to our nuclear plants, we have the ability to open data centers, to build computing capacity, to be at the heart of the artificial intelligence challenge." Translating exports into compute is the entire French AI strategy in one sentence. The Flamanville 3 reactor (1.6 GW), connected in December 2024 and commissioned through 2025 and early 2026, adds to that headroom.

The third advantage is price. France's wholesale electricity is structurally cheaper than its neighbors'. The French regulator CRE anticipated wholesale prices around 58 euros per MWh in France for 2026 versus roughly 88 euros per MWh in Germany, a gap of about 30 euros per MWh. Over the life of a 3.1 GW campus running flat out, a 30-euro-per-MWh advantage is worth billions. That is the math that moved SoftBank's capital north.

There is one important caveat the French pitch tends to skip. European industrial electricity is still roughly double the US level for energy-intensive users. France is the cheapest seat in an expensive theater. Against Texas or Louisiana, France's nuclear advantage narrows considerably, because American gas and the new wave of behind-the-meter generation are cheaper still. France wins the European race decisively. It does not automatically win the global one, which is why the French and EU strategy leans so heavily on a second argument that has nothing to do with price: sovereignty.

The Energy Wall: The Real Constraint on AI Scaling

For most of the past three years the industry talked about AI scaling as a chip problem. Then it became a capital problem. In 2026 it is, unambiguously, an energy problem. Call it the energy wall: the point at which the marginal constraint on building more AI capacity is not GPUs and not money but megawatts that can be delivered to a specific location at a specific time.

The numbers make the wall concrete. By late 2027, US AI data centers are projected to need somewhere between 20 and 30 GW of combined power. A single Meta campus, Hyperion in Louisiana, is planned at 5 GW. OpenAI's Stargate program targets 10 GW across multiple sites. xAI's Colossus 2 is pushing toward 2 GW on a single Tennessee-Mississippi campus. These are not data centers in the old sense; they are private power stations with computers attached. The compute follows the power, not the other way around.

What makes this a wall rather than a hill is that power infrastructure scales on a fundamentally different clock than capital or silicon. You can raise 45 billion euros in a quarter. You can order GPUs and take delivery in months. You cannot conjure a 400 kV transmission line, a substation, and a grid connection in that time. In Western Europe, standard interconnection queues imply connection timelines of seven to ten years. Seven to ten years. A data center developer can secure land, permits, capital, and an anchor tenant and still be unable to energize the asset within any window that makes financial sense.

This is why the gigawatts-announced number and the gigawatts-energized number diverge so sharply, and why I keep insisting they are different metrics. Announcing 5 GW is a press release. Energizing 5 GW is a decade-long civil-engineering and regulatory campaign. The market rewards the announcement. The grid enforces the reality. Anyone evaluating these commitments should mentally discount the announced figure by the probability and timeline of actual energization, and that discount is large.

The chart above is illustrative, not a SoftBank forecast, but the shape is the point. Announced capacity is a flat line set the day of the press conference. Energized capacity is a slow ramp gated by grid connection and commissioning. The area between the two curves is risk, capital tied up in idle steel, and the single most important thing investors and policymakers underweight.

Grid Timelines vs Capex Speed

France understands the energy wall better than most, because it has been running a state-coordinated grid for decades. Its response is instructive. In May 2025, the French regulator CRE introduced a fast-track procedure (deliberation n° 2025-120) for connecting very large electricity consumers, specifically hyperscale data centers, to the high-voltage transmission grid at 400 kV. Under this regime, connections can be completed in three to four years instead of seven to ten. The catch is that the fast track applies only to sites the state designates as favorable, with feasibility confirmed by RTE, the transmission operator.

Read that again, because it is the whole industrial-policy story in miniature. France did not deregulate its way to faster data centers. It centralized. The state picks the sites, the state-owned transmission operator confirms feasibility, the state-owned utility supplies the power, and a national champion (Schneider) supplies the gear. The SoftBank campuses sit inside this designated-site fast track. That is not an accident; it is the entire reason the 2031 timeline is even arguable. Outside the fast track, a 3.1 GW build by 2031 would be implausible.

Now contrast that with the speed of capital. SoftBank can deploy capex at a rate the grid cannot match. The mismatch creates a specific failure mode: campuses where the buildings, GPUs, and cooling are finished while the high-voltage connection is still in queue, leaving stranded assets earning nothing. The French model attacks this by making the grid connection a precondition of site designation rather than an afterthought. It is slower to start and faster to finish, which is exactly the right trade for infrastructure that has to last twenty years.

There is a national-resource-adequacy dimension too. RTE's own projections, in its 2025-2035 outlook, see data-center electricity consumption reaching 15 to 20 TWh by 2030 (roughly 3 percent of French consumption) and around 35 TWh by 2035 in the rapid-decarbonization scenario. France consumed under 1 TWh on dedicated data-center sites in 2025. So the SoftBank build, plus the broader AI wave, represents a step-change in domestic demand. The 10-plus GW of export headroom is real, but a meaningful chunk of it is now spoken for. Sovereign compute, in energy terms, means converting electricity France used to sell abroad into compute France sells to the world. That is a deliberate reallocation of a strategic resource, and it is the kind of choice only a state-coordinated grid can make cleanly.

The slope of that curve is the part worth sitting with. France goes from a rounding error to roughly 35 TWh of dedicated data-center demand inside a decade. Against 92.3 TWh of 2025 net exports, that is more than a third of the headroom consumed by a single use case. The political economy of this is delicate. Every terawatt-hour fed to a data center is a terawatt-hour not exported to Germany, Italy, or Britain, and not available to electrify French transport and heating, which the country's own climate plan requires. France's nuclear fleet is large but not infinite, and the new EPR2 reactors that would expand it are a decade and tens of billions of euros away from delivering power. The SoftBank campuses are being underwritten, in effect, by export revenue France is choosing to forgo and by climate headroom France is choosing to allocate to AI rather than to domestic decarbonization. That is a defensible choice, but it is a choice, and it deserves to be named as one rather than waved through as free.

The capex-versus-grid mismatch in numbers

It helps to put concrete tempos on the two clocks I keep invoking. On the capital side, SoftBank's first phase implies roughly 7.5 billion euros of deployment per year for six years. A firm that orchestrated the original Vision Fund can move money at that pace in its sleep; the constraint there is governance and discipline, not availability. On the grid side, the relevant tempo is measured in substations and kilometers of 400 kV line. A single hyperscale connection at that voltage involves transformer procurement with multi-year lead times (the global transformer shortage is real and acute), right-of-way acquisition, environmental review, and physical construction that proceeds at the speed of heavy civil engineering, not software. Even under France's fast-track, three to four years is the optimistic case, and it assumes RTE's own build program stays on schedule while serving every other connection request in the queue simultaneously.

The chart makes the asymmetry unmistakable. Everything except the grid connection can be compressed into roughly two years with enough money and supplier muscle. The grid connection cannot. It is the long pole in the tent by a factor of two or more even in the friendliest regime on the continent. This is why I argue that the announced-to-energized gap is not a temporary friction that better project management will dissolve. It is structural. It is the grid clock asserting itself against every other clock in the system, and no amount of capital velocity changes its tempo.

Sovereign Compute as Industrial Policy

The word "sovereign" is doing a lot of work in 2026, and it is worth unpacking. Sovereign compute means compute capacity located within a jurisdiction, subject to that jurisdiction's law, and ideally owned or controllable by domestic actors, so that a foreign government or hyperscaler cannot revoke access during a crisis. After three years of watching the US-China chip war and the concentration of frontier compute in a handful of American firms, European policymakers concluded that depending on rented American GPUs for the foundational technology of the next economy was a strategic vulnerability on par with energy dependence.

The EU's response is now concrete and funded. In January 2026 the Council amended the EuroHPC Joint Undertaking regulation to add AI gigafactories to its mandate, facilities planned around roughly 100,000 advanced processors each, about four times the current generation. The broader gigafactory program is sized in the tens of billions of euros. France's national piece, AI Factory France, federates existing national compute with the forthcoming Alice Recoque exascale system and is in talks to link with Germany's JAIF system toward a virtual 50,000-GPU supercomputer for federated learning. Layer on top of that the private frontier-model layer: Mistral AI plans 200 MW of its own data-center capacity in Europe by end of 2027, funded independently, after raising 830 million euros in institutional debt in early 2026.

The SoftBank commitment sits at the intersection of these two layers. It is private foreign capital (Japanese) building on French soil under French and EU rules, supplied by French nuclear power, using French-made power modules. It is sovereign in the sense that the physical asset and its electrons are European, even though the balance sheet is not. That is the pragmatic compromise European sovereignty has settled on: you do not need to own the capital if you control the land, the law, the power, and a chunk of the supply chain. Whether that is "sovereign enough" is the central political question, and reasonable people disagree. I explored the broader dynamics of this competition in my analysis of the AI sovereignty cascade across Cohere, Aleph Alpha, and Manus, and the pattern there, national champions racing to plant flags before the window closes, is the same one playing out in Hauts-de-France.

Competition With US Hyperscalers

It is tempting to frame this as France versus the United States, but the more accurate frame is two different theories of how to build AI infrastructure. The American theory is speed through abundance: cheap gas, behind-the-meter generation, deregulated siting, and a willingness to strain local grids and accept the environmental bill in exchange for energizing capacity fast. The French theory is speed through coordination: clean baseload, state-designated sites, a national transmission champion, and a slower start in exchange for a faster, more durable finish and a far better carbon profile.

The American model is winning on raw gigawatts today. Stargate's 10 GW, Meta's 5 GW Hyperion, xAI's 2 GW Colossus 2, Microsoft's gigawatt-class Fayetteville, all energizing through 2026 and 2027, dwarf anything in Europe. But the American model is paying for that speed with grid stress, rising consumer electricity prices in affected regions, water conflicts, and a carbon footprint that will become a liability as AI's emissions draw scrutiny. The French model is slower and smaller but cleaner and more politically durable. I dug into the American end of this in my piece on the SpaceX AI industrial complex and the Colossus build-out, and the contrast with France is stark: one system optimizes for time-to-energize at any cost, the other for legitimacy and longevity.

The strategic question for Europe is not whether it can out-build the United States on gigawatts. It cannot, not in this cycle. The question is whether it can build enough sovereign capacity to host its own frontier models, serve its own regulated industries (banking, health, defense, public administration), and avoid total dependence on American clouds. Five gigawatts in France, plus the EuroHPC gigafactories, plus Mistral's self-build, is plausibly enough to clear that lower bar. Sovereignty does not require parity. It requires a credible floor.

The SoftBank Question: Can the Financier Finish?

There is a counterparty risk in this story that the French government has every incentive to downplay and that prudent observers should not. SoftBank is not a utility, a hyperscaler with a balance sheet full of cloud profits, or a sovereign wealth fund. It is a holding company whose history includes both spectacular wins and spectacular write-downs, and whose AI thesis is enormous, concentrated, and partly funded by debt and asset sales. The same firm is simultaneously committed to the US Stargate program, to a large stake in its chip-design holdings, and now to up to 75 billion euros in France. The aggregate of SoftBank's announced AI commitments across geographies is large relative to its demonstrated capacity to fund them without selling assets or raising new capital on favorable terms.

This matters because the conditional structure of the France deal, 45 billion euros firm and 30 billion euros subject to extension, is exactly the kind of structure that can be quietly de-scoped if SoftBank's broader AI bet runs into a funding wall. The firm tranche is real and contractually anchored to named sites and partners. The extension is an option, and options get abandoned when the market turns. If the AI capital cycle cools (and there is no shortage of voices warning that it is overheated), the most likely casualty is not the 3.1 GW first phase but the conditional 1.9 GW and the broader "additional sites across France" language. France would still get a very large data-center cluster. It would not get the full 5 GW headline it announced. That, too, is a version of the announced-versus-energized gap, operating through finance rather than through the grid.

My own base case, reflected in the chart above, is that France energizes something in the neighborhood of 2.6 GW by 2031, short of the 3.1 GW firm target and well short of the 5 GW headline, not because anyone acted in bad faith but because grid commissioning slips, one of the three sites encounters a permitting or transmission snag, and the conditional extension is deferred. That would still be, by a wide margin, the largest AI data-center build in Europe and a genuine strategic achievement for France. It would also vindicate the thesis that you should always read the announced number as a ceiling and the energized number as the thing that actually arrives.

Water, Cooling, and the Local Bill

Every gigawatt of AI compute lands somewhere specific, and the place it lands pays a bill that does not appear in the press release. Water is the most visible line item. A 100 MW data center using conventional cooling can consume on the order of 2 million liters per day. Scale that naively to 3.1 GW and the numbers get alarming fast, which is why the cooling technology choice is not a footnote but a determinant of whether a campus is socially viable.

The good news, and SoftBank's Schneider partnership leans into this, is that closed-loop liquid cooling dramatically reduces water draw. Schneider's own modeling for a Paris-region facility shows annual water consumption dropping from about 108,000 cubic meters with traditional air cooling to roughly 51,000 cubic meters with liquid cooling, a 53 percent reduction. A closed-loop system might consume only 5 to 10 percent of its water withdrawal on a net basis. An air-cooled facility consumes almost no water on-site but pays for it in electricity for chillers. The choice is a triangle between water, power, and capex, and the Hauts-de-France campuses appear to be choosing the liquid-cooling, lower-water corner.

But water is only the most photogenic part of the local bill. There is land, taken out of agricultural or industrial use. There is the visual and acoustic footprint of a gigawatt campus on a rural skyline. There is the grid build-out, new pylons and substations marching across the countryside, which communities experience as an imposition even when the data center itself is invisible behind a fence. And there is the labor question: "thousands of high-skilled jobs" is real during construction, but a finished, automated data center employs surprisingly few people relative to its footprint and power draw. The town gets the pylons and the water draw permanently; it gets the construction jobs temporarily and a modest operations headcount thereafter.

This is the part of the story that the gigawatt accounting erases, and it is the part I tried to render human in the accompanying short story, The Water Table, set in a Hauts-de-France town as the campus arrives. The macro numbers are clean. The local experience is not.

The Gap Between Announced and Energized

I want to return to the metric I opened with, because it is the one readers should carry away. Across the global AI build-out, the sum of announced gigawatts now vastly exceeds the sum that will be energized on the announced timelines. This is not cynicism; it is the structural reality of infrastructure that scales on a grid clock while being financed on a capital clock. SoftBank's 5 GW is real intent backed by real money. It is also, like every other announcement, subject to a discount for grid connection, permitting, equipment supply, and the simple physics of commissioning power blocks one at a time.

The investment implication is direct. The firms and regions that will win the next phase are not the ones with the largest announcements; they are the ones that can compress the announced-to-energized gap. France's fast-track regime, its clean baseload, and its state-coordinated grid are precisely a gap-compression strategy. SoftBank chose France not because it is the cheapest or the fastest in absolute terms, but because France offers the best odds of actually energizing the gigawatts it announced. In a world drowning in announcements, credibility of energization is the scarce asset.

The deeper compute-economics dimension, who actually gets to use these gigawatts once they are live, and how agent demand is already colliding with rate limits, I covered in my analysis of Anthropic, SpaceX Colossus, and the economics of renting compute. The short version: even 5 GW energized does not end scarcity, because demand for inference and agentic workloads is growing faster than any plausible build cadence. The energy wall does not disappear when France clears its grid queue. It moves.

The Three-Bloc Energy Map

Step back far enough and a three-bloc structure comes into focus, and energy is the axis that defines it. The United States is building on abundance and speed: cheap gas, behind-the-meter generation, deregulated siting, and a tolerance for grid stress and deferred environmental cost in exchange for energizing tens of gigawatts fast. China is building on state direction at a scale that dwarfs both, pairing aggressive coal-and-renewables expansion with a nuclear build-out and a manufacturing base that supplies its own equipment, and it is doing so behind a wall of export controls that has forced a parallel domestic chip ecosystem. Europe is building on coordination and clean baseload, slower and smaller, betting that durability and legitimacy beat raw speed over a twenty-year horizon.

Each bloc's strategy is downstream of its energy endowment and its political system. America has cheap molecules and a fragmented, market-driven grid, so it gets fast, dirty, distributed builds. China has a command economy and a vast state grid, so it gets enormous coordinated builds with a heavy carbon footprint it is racing to clean up. France, and through it Europe, has clean nuclear baseload and a centralized grid, so it gets slow, clean, coordinated builds that energize what they announce. The SoftBank deal is Europe's strategy made concrete: foreign capital channeled through a state-coordinated, nuclear-backed, designated-site regime to produce sovereign-but-not-sovereign capacity. It is the only one of the three models that can credibly claim its gigawatts are both low-carbon and likely to actually arrive.

The competitive question for the decade is not which bloc announces the most. It is which bloc's energized capacity grows fastest relative to its demand, and which bloc's model proves politically durable when the local bills, water, land, pylons, electricity prices, come due. America's model is fastest to energize but most exposed to local backlash as consumer power prices rise near data-center clusters. China's is largest but most carbon-exposed and most opaque. Europe's is slowest but cleanest and most consent-friendly, which in democracies is not a soft factor but a hard constraint on whether you can build the next campus at all. France is betting that legitimacy compounds. Over one cycle, abundance wins. Over three, coordination might.

This is why the SoftBank France announcement is more than a regional infrastructure story. It is a test of whether the coordinated, clean, sovereignty-driven model can move at a tempo that keeps Europe in the game at all. If France energizes even 2.6 to 3 GW of clean, sovereign AI capacity by 2031 while the American model chokes on grid backlash and rising power prices, the European bet looks prescient. If France's grid clock proves even slower than feared and the capital wanders back to Texas, the bet looks like a beautifully engineered way to arrive late. Both outcomes are live, and the next two years of grid-connection milestones, not capex announcements, will tell us which way it breaks.

What To Watch

Five things will tell you whether the SoftBank France bet is tracking. First, grid-connection milestones: watch for RTE confirmations and substation commissioning at Loon-Plage, Bosquel, and Bouchain, not for additional capex announcements. Second, the EDF supply contract for Bouchain, which is the first concrete test of converting nuclear export headroom into dedicated compute power. Third, the Schneider power-module facility at Dunkirk, whose ramp is the leading indicator of whether the integrated-cluster model actually accelerates the build. Fourth, the conditional 1.9 GW extension, which will either be triggered (signaling the model works) or quietly deferred (signaling the energy wall bit harder than expected). Fifth, local consent: water permits, agricultural-land conversions, and the politics of new transmission lines through Hauts-de-France communities.

If those five line up, France will have demonstrated that a state-coordinated, nuclear-backed, sovereignty-driven model can energize gigawatts on a credible timeline. That would be a genuine alternative to the American abundance model, and a template the rest of the EU would follow. If they slip, the SoftBank announcement will join the growing pile of gigawatts that were announced and never energized, and the energy wall will claim another headline.

My own read, developed in the prediction that accompanies this piece, is that European sovereign-compute programs will collectively announce a great deal more capacity than they energize over the next two years, and that the gap will be the defining metric of the era. You can read the specific, dated claim here: EU sovereign-compute announced vs energized through 2027. The energy wall is real, it is the binding constraint, and France has chosen the one strategy, coordination over abundance, that has any chance of getting over it cleanly. Whether 87.5 billion euros and a deep-water port are enough to beat the grid clock is the question 2031 will answer.

Further Reading

• The AI sovereignty cascade: Cohere, Aleph Alpha, China, and Manus
• The SpaceX AI industrial complex, Cursor's IPO, and Colossus
• Anthropic, SpaceX Colossus, and the economics of renting compute
• Prediction: EU sovereign compute, announced vs energized through 2027
• Short story: The Water Table