SoftBank Bets 75B Euros to Build 5GW France AI Hub

Masayoshi Son just pledged more capital to French data centers than the French government spends on its entire annual defense budget. SoftBank will pour up to 75 billion euros, roughly 87 billion dollars, into 5 gigawatts of AI compute capacity across northern France. It is the largest infrastructure commitment SoftBank has ever made on European soil, and the size of the figure forces a question the continent has dodged for a decade: who will actually own the physical layer of Europe's AI future, and what does it cost to find out?

What Actually Happened

On May 31, SoftBank Group confirmed it would invest up to 75 billion euros to build and operate as much as 5 gigawatts of additional data center capacity in France. The first phase commits the company to roughly 3.1 gigawatts spread across three sites in the Hauts-de-France region: Dunkirk at Loon-Plage, Bosquel, and Bouchain. SoftBank wants this first wave online by 2031, a timeline that is aggressive for projects of this electrical scale. The company described the program as its single largest AI infrastructure investment anywhere in Europe, a deliberate signal that Son views the continent as a core part of his compute empire rather than an afterthought.

The deal did not emerge from a procurement process. It grew out of direct diplomacy between French President Emmanuel Macron and Son, who met during Macron's visit to Japan earlier this year. France brought industrial partners to the table to make the power and construction work. Schneider Electric will act as the industrial partner at the Dunkirk hub, handling the electrical backbone and cooling architecture that a gigawatt campus demands. State backed utility EDF will supply low carbon nuclear power to the Bouchain facility, which matters because a 5 gigawatt buildout is, before anything else, an electricity problem dressed up as a technology story.

SoftBank is not a neutral landlord here. It is both an investor in OpenAI and one of its largest customers through the Stargate compute program. That dual role means the French capacity is unlikely to sit on an open market for any enterprise to rent. It is far more probable that a large share of these gigawatts feed frontier model training and inference for the OpenAI ecosystem and SoftBank's own AI ventures. The 75 billion euro headline is therefore not just a French win. It is SoftBank securing a European beachhead of compute it can direct, at a moment when access to power and land has become the true bottleneck of the AI race.

Why This Matters More Than People Think

The instinct is to read this as another large number in an industry drowning in large numbers. That misses what 5 gigawatts physically represents. Five gigawatts is roughly the output of five full size nuclear reactors, or close to one percent of France's entire installed generating capacity, dedicated to running silicon. When a single private buildout consumes that much of a national grid, AI stops being a software story and becomes an energy and industrial policy story. France can offer this because its grid is more than 60 percent nuclear, giving it spare low carbon power that Germany, post nuclear phaseout, simply cannot match. SoftBank chose France for the electrons, not the engineers.

For Europe, the deal cuts in two directions at once. It is a validation that the continent can attract frontier scale capital, and it is a warning that the capital arrives on foreign terms. The hardware, the models, and the orchestration layer remain American or Japanese. France contributes land, power, and permits. That is the same arrangement that has defined Europe's relationship with the cloud era for fifteen years, now repeated one layer down at the level of raw compute. Macron can claim a jobs and investment victory while the strategic dependency quietly deepens, and both things are true simultaneously.

There is also a hard financial subtext. SoftBank is funding a 75 billion euro program while carrying heavy existing commitments to OpenAI and the broader Stargate effort, which together run into the hundreds of billions. The company is effectively betting that demand for AI compute will keep compounding fast enough to fill 5 gigawatts of French capacity at prices that justify the spend. If inference demand grows as projected, these sites print money for two decades. If a model efficiency breakthrough collapses the compute needed per token, SoftBank could be left holding gigawatts of half utilized concrete in Dunkirk. The downside is as large as the ambition.

The Competitive Landscape

SoftBank is not building into empty territory. Microsoft, Amazon Web Services, and Google have all announced multibillion euro European data center programs, and Mistral is constructing a 13,800 GPU campus near Paris to anchor France's homegrown model effort. What separates the SoftBank commitment is the willingness to underwrite an entire 5 gigawatt fleet up front rather than expanding in cautious tranches. That scale is a deliberate moat. Power interconnection queues in Europe stretch years long, and whoever locks in grid capacity first effectively locks competitors out of the same region. SoftBank is buying time and electrons, two things money cannot conjure quickly.

The closest historical parallel is the railway and electrification booms of the late nineteenth and early twentieth centuries, when a handful of industrialists raced to claim rights of way and generating capacity that defined regional economies for generations. The firms that secured the corridors early extracted rents for decades, while latecomers paid premium prices or were locked out entirely. AI compute is following the same physics. The scarce input is no longer capital, which is abundant, but interconnection rights, cooling water, and land near low carbon power. SoftBank is playing the land grab phase, and France's nuclear grid is the equivalent of a perfectly placed coal seam in 1900.

The named competitors will not respond passively. Expect AWS and Microsoft to accelerate their own French and Nordic announcements, and expect Gulf sovereign funds, already circling European compute, to push their own multi gigawatt proposals. The Nordics offer cheaper hydropower and natural cooling, while Spain and Portugal pitch solar abundance. France's advantage is the combination of dense nuclear baseload and a government willing to fast track permits for a marquee investor. The competitive question for the next 18 months is whether other European states match France's speed of approval, or whether bureaucratic caution hands SoftBank a multi year head start it never has to defend.

The skeptics point out a timing risk the headline conveniently ignores. SoftBank has a documented history of buying near the top of a cycle, from its 100 billion dollar Vision Fund that backed WeWork at a 47 billion dollar valuation to late entries in chip and robotics bets that took years to recover. A 75 billion euro buildout announced at the peak of AI infrastructure euphoria invites the obvious comparison. The bear case is that compute supply is being added faster than durable, paying demand, and that 2027 could bring a glut of gigawatts chasing a market whose growth was extrapolated from a two year boom. However, even critics concede that low carbon power near population centers is a genuinely scarce asset that holds value across cycles, which is why the choice of nuclear heavy France, rather than a cheaper but dirtier grid, reads less like hype and more like a deliberate hedge against exactly that risk.

Hidden Insight: The Grid Is the Real Frontier Model

The non obvious truth buried in this announcement is that the frontier of AI has quietly migrated from the model to the megawatt. For three years the industry obsessed over parameter counts and benchmark scores. The binding constraint has now shifted entirely to power. A company can license a frontier model in an afternoon, but it cannot conjure 5 gigawatts of low carbon electricity and the substations to deliver it. SoftBank understands this, which is why it is spending 75 billion euros on transformers, cooling, and grid connections rather than on a flashier model lab. The moat is no longer the algorithm. It is the electrical interconnection agreement.

This reframes who the real winners are. The model labs capture attention, but the durable economic rents may accrue to whoever controls the power and the physical sites. EDF, Schneider Electric, and the turbine and transformer makers behind this buildout are positioned to earn steady margins for twenty years regardless of which model wins. In every gold rush the reliable fortunes went to the suppliers of picks, shovels, and water rights. In this one the picks are GE Vernova turbines, the shovels are Schneider switchgear, and the water rights are EDF power purchase agreements. SoftBank is paying all three, and those are the businesses with the most defensible position.

There is a deeper geopolitical layer. By anchoring 5 gigawatts in France, SoftBank is also making a quiet bet on regulatory geography. The European Union is simultaneously tightening data sovereignty rules, and physical compute located inside the bloc becomes far more valuable if regulators eventually require sensitive workloads to run on European soil. SoftBank may be positioning to sell sovereign compliant compute to European banks, hospitals, and governments that will soon be barred from processing sensitive data on US located infrastructure. The 75 billion euros is not only a power play. It is a regulatory arbitrage on rules that have not finished being written.

Consider the second order effect on European labor and industry. A 5 gigawatt buildout is one of the largest construction undertakings the Hauts-de-France region has seen in a generation, drawing thousands of electricians, steelworkers, and civil engineers to sites near Dunkirk and Bouchain. The political appeal is obvious, because Macron can point to concrete jobs in a region that has felt deindustrialization acutely. Yet the permanent headcount inside a hyperscale data center is small, often a few hundred technicians for a facility that draws a gigawatt. The lasting economic value is the construction phase and the tax base, not a durable employment engine, and that gap between the jobs promised at the announcement and the jobs that remain at steady state is the part of the story local politicians rarely emphasize.

The uncomfortable insight for European policymakers is that sovereignty bought with foreign capital is a contradiction they have chosen not to examine. France will host the gigawatts, employ the construction crews, and collect the tax receipts, yet the strategic direction of what those gigawatts compute sits in Tokyo and San Francisco. Europe is becoming the world's preferred location to run AI without becoming a place that controls AI. That distinction will not matter while the electrons flow and the jobs arrive. It will matter intensely the first time a geopolitical dispute makes a foreign owner reconsider what runs on French soil.

What to Watch Next

In the next 30 days, watch for the grid interconnection filings with RTE, France's transmission operator. The speed at which 3.1 gigawatts of connection requests move through the queue will reveal whether the government's fast track promises are real or rhetorical. Also watch whether SoftBank names a named anchor tenant. If OpenAI or a Stargate entity is formally attached to the Dunkirk site, that confirms the capacity is captive rather than commercial, which changes the economics for every other European cloud buyer hoping to rent these gigawatts.

Over 90 to 180 days, the markers to track are financing structure and power pricing. SoftBank has not detailed how much of the 75 billion euros is its own balance sheet versus project debt and partner co investment, and that split determines how exposed the company is if AI demand softens. Watch the EDF power purchase agreement terms at Bouchain, because the price per megawatt hour locked in now will define the site's competitiveness for two decades. Watch also for copycat announcements from Spain, the Nordics, and the Gulf, which will signal whether France's nuclear advantage is unique or merely first.

The mental model for evaluating this is simple. If AI inference demand keeps compounding and European sovereignty rules tighten, SoftBank will look prescient and these French gigawatts will be among the most valuable industrial assets on the continent. If a model efficiency breakthrough slashes compute per query, or if interconnection delays push the 2031 timeline out by years, the same buildout becomes a cautionary tale about confusing a capital glut for a demand signal. The bet is binary, the stakes are 75 billion euros, and the first real evidence arrives the moment those substations energize.

The frontier of artificial intelligence has quietly moved from the model to the megawatt, and SoftBank just spent 75 billion euros to prove it.

---