The Megawatt Siege: Sovereign AI and the New Geography of Power

I. The French Signal

In May 2026, during President Emmanuel Macron’s Choose France summit, SoftBank Group announced a commitment of up to €75 billion to develop approximately 5 gigawatts of AI-oriented data center capacity across France over the next five years. The first phase alone injects roughly 3.1 gigawatts of compute infrastructure into northern France.

At first glance, this appeared to be another technology investment. In reality, it was a geopolitical signal.

The decision did not simply reflect confidence in European demand for AI services. It reflected confidence in French energy infrastructure specifically — in the stability, density, and long-term security of a national grid anchored by one of the world’s most extensive nuclear fleets. Masayoshi Son was not buying software capacity. He was buying megawatts under the protection of the French state.

To grasp the scale involved: 5 gigawatts is roughly equivalent to the output of several large nuclear reactors operating simultaneously. For most of modern history, concentrations of electricity at that level were associated with steel mills, chemical complexes, aluminum smelters, and military-industrial production. Today, they are increasingly associated with artificial intelligence.

Capital markets are beginning to allocate nation-state levels of energy consumption to computation — and the geography of that capital is being determined not by software capability, but by which states can guarantee the physical conditions that make computation possible.

Training frontier AI models requires continuous, high-density electricity delivered without interruption for months at a time. Massive GPU clusters cannot tolerate grid instability or prolonged outages. They require firm baseload generation, deep transmission capacity, long-term price visibility, and the kind of political stability that guarantees none of this will be arbitrarily revoked. France is one of the very few large European economies that offers all four.

Masayoshi Son was not buying software capacity. He was buying megawatts under the protection of the French state.

II. The Return of Baseload

One of the structural ironies of the AI revolution is that it has revived concepts that Western energy policy spent a decade trying to retire.

The dominant discourse of the 2010s prioritized renewable deployment, emissions targets, and the gradual phase-out of dispatchable baseload generation. The implicit assumption was that intermittency could be managed through storage, demand response, and grid interconnection.

That assumption was never tested against industrial-scale computation — because industrial-scale computation did not yet exist at the densities now being built.

AI has introduced a constraint that renewable-first energy policy was not designed to handle: continuity. Data centers do not care whether electricity is generated by wind or nuclear or hydro. They care whether it is available — not on average, not most of the time, but continuously, at high density, for years without interruption.

The numbers are unambiguous. Global data center electricity consumption reached 415 terawatt-hours in 2024 — approximately 1.5% of global electricity use, growing at 12% per year, more than four times faster than total electricity demand.

The IEA projects this figure will approach 945 terawatt-hours by 2030, roughly equivalent to Japan’s entire current annual consumption. In 2025 alone, data center electricity use surged 17%, with AI-focused facilities growing faster still. The five largest technology companies increased their combined capital expenditure to over $400 billion in 2025 and are projected to increase it by a further 75% in 2026.

This is not a trend. It is a structural transformation of the global energy system, and it is happening faster than the physical infrastructure required to support it can be built.

The result has been a rapid reassessment of energy infrastructure that would have seemed politically impossible five years ago. Nuclear power, dismissed across much of Europe as obsolete and dangerous, is being reconsidered as the only large-scale low-carbon technology capable of meeting AI’s baseload requirements.

The pipeline of conditional offtake agreements between data center operators and small modular reactor projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts as of early 2026. Microsoft signed a $1.6 billion, 20-year agreement with Constellation Energy to revive the Three Mile Island nuclear plant. Amazon secured 1.92 gigawatts from Talen Energy’s Susquehanna nuclear plant through 2042. Google signed the first corporate SMR power purchase agreement with Kairos Power in 2025.

The strategic question has shifted from who can generate electricity to who can generate electricity reliably, continuously, and at scale.

III. The Hidden Bottleneck

Most commentary on AI infrastructure still focuses on the processor layer. This framing captures something real, but it increasingly misses the binding constraint.

Across Europe, North America, and significant parts of Asia, the most immediate limitations on AI infrastructure deployment are not semiconductors. They are transformers, substations, transmission corridors, and grid interconnection queues — the physical systems that deliver electricity from generators to data centers.

These systems are expanding far more slowly than the demand they are being asked to serve.

The data is striking. Standard power transformers average 128 weeks for delivery, according to Wood Mackenzie’s 2025 survey. Generator step-up transformers average 144 weeks. Some orders extend to four years. Transformer prices have risen 77% since 2019. Substation lead times averaged 140 weeks in 2023, rising to 160 weeks by 2026.

Sightline Climate tracked 12 gigawatts of US data center capacity announced for 2026 across 140 projects — and found that only 5 gigawatts is actually under construction, with 25% of projects having disclosed no power strategy at all.

A next-generation AI model can be designed in months and trained in weeks. A large power transformer requires two to four years from order to installation. A high-voltage transmission line may take a decade from permitting to energization.

The world can produce new AI models faster than it can build the electrical systems required to run them at sovereign scale. This asymmetry — digital acceleration against physical constraint — is the defining bottleneck of the AI era.

It has also produced a geopolitical irony that has gone largely unremarked. China controls approximately 60% of global power transformer manufacturing capacity. In 2025 alone, the United States imported more than 8,000 high-power transformers from China — up from fewer than 1,500 in all of 2022.

The country leading the AI race is structurally dependent on its principal strategic adversary for the physical equipment that makes the race possible. The semiconductor dependency that Washington has spent years trying to reverse has an electrical infrastructure mirror image that is, if anything, more acute — and far less visible.

The world can produce new AI models faster than it can build the electrical systems required to run them at sovereign scale.

IV. Portugal and the Atlantic Model

France is not the only European geography being revalued through this lens — but understanding Portugal’s emerging position requires looking past the obvious.

The conventional analysis of Portugal’s AI infrastructure potential focuses on Sines: its deep-water port, its Atlantic positioning, its renewable energy generation, and its role as a terminus for transatlantic subsea cable systems linking Europe with North America and South America. These are real advantages and they are attracting real capital.

But the deeper significance of Portugal’s position lies elsewhere — in a convergence of factors that, taken individually, appear incremental and, taken together, suggest the emergence of something structurally distinct.

Portugal’s 2009 claim to an extended continental shelf, if validated by the UN Commission on the Limits of the Continental Shelf, would give it the tenth largest maritime boundary in the world. That claim encompasses substantial stretches of the Mid-Atlantic Ridge — the most geologically active hydrogen generation zone in the Atlantic basin, where the continuous serpentinization of ultramafic rock has been producing natural hydrogen at depth for millions of years.

The Azores archipelago sits directly astride this ridge, placing it in the rare category of locations where the geology of the deep ocean is accessible from existing land-based and near-shore infrastructure.

This matters because geological hydrogen is not a future technology. It is a present geological reality whose commercial infrastructure remains to be built. In June 2025, the French government confirmed natural hydrogen reserves in Lorraine of approximately 46 million tons — roughly half of current global annual hydrogen production — alongside significant reserves in the Pyrenees and Aquitaine, with purity levels exceeding 90% at several sites.

These formations do not stop at the French border. The geological conditions that produced them extend across the Western Atlantic margin in ways that current exploration is only beginning to map.

Data center investments made today will operate for twenty to thirty years. The capital evaluating these sites is not asking only about current energy costs. It is asking about long-term energy optionality — about which geographies are positioned to host successive generations of energy surpluses, not just the present one.

A geography that offers renewable generation today, geological hydrogen optionality in the medium term, and deep Atlantic connectivity throughout is not merely an infrastructure site. It is a strategic position.

France represents the nuclear model of computational sovereignty — deep baseload power, immediate security, decades of operational experience. Portugal represents the Atlantic model — connectivity, layered optionality, and a geographic position at the intersection of transatlantic trade routes, subsea cable architectures, and emerging hydrogen geology that may prove as consequential in the twenty-first century as it was in the age of maritime exploration.

A geography that offers renewable generation today, geological hydrogen optionality in the medium term, and deep Atlantic connectivity throughout is not merely an infrastructure site. It is a strategic position.

V. Four Models, One Conclusion

The transformation is not confined to Europe. Across different geographies and political systems, the same structural logic is producing convergent conclusions.

In the United States, the Stargate initiative in Texas, hyperscale expansion across Virginia and Ohio, and the accelerating concentration of compute infrastructure around energy-rich regions all reflect the same reality: AI follows energy. Northern Virginia’s data centers already consume 26% of the state’s electricity. Dublin’s consume 79% of the Irish capital’s supply. The geography of computation is being determined by the geography of the grid.

In Canada, Quebec’s vast hydroelectric reserves have attracted growing interest from AI infrastructure developers seeking stable long-term power contracts — baseload renewable generation that offers continuity at scale, precisely the combination that most electricity markets cannot readily provide.

Across the Gulf, Saudi Arabia and the United Arab Emirates are investing billions to convert energy abundance into computational capacity, positioning themselves as future AI hubs rather than merely energy exporters. This ambition — directly relevant to the analysis of the Iran conflict offered in The New Tordesillas — illustrates how the transformation of energy geography and the transformation of computational geography are becoming the same process, viewed from different angles.

China is pursuing a different path altogether, integrating artificial intelligence into a state-directed framework that links power generation, manufacturing capacity, transmission networks, rare earth processing, and national industrial strategy into a unified system. The model is distinct from anything in the Western experience: not market-driven investment following demand signals, but sovereign infrastructure deployment executing a long-range plan in which AI capability, energy independence, and industrial dominance are treated as aspects of a single strategic objective.

VI. The Chinese Dimension

No analysis of the AI-energy nexus is complete without confronting the Chinese trajectory — and the picture is more complex than the standard framing of Western competition suggests.

China’s data center buildout, grid expansion, renewable deployment, and retention of coal-fired baseload generation — maintained explicitly to support industrial-scale computation during the transition period — represent a level of physical infrastructure investment that no market-driven economy has matched. China accounts for 25% of global data center electricity consumption today and is projected to account for the largest share of global growth through 2030, alongside the United States.

Yet China’s mineral endowment, while substantial in rare earths and battery materials, does not include the geological hydrogen formations concentrated along the Atlantic margin and in the cratonic zones of the Western Hemisphere. Its rare earth dominance — examined in detail in The New Tordesillas — gives it unparalleled leverage over the electric vehicle and battery storage transition. It gives it no equivalent leverage over a hydrogen economy whose primary resource base lies in geographies it does not control and, under the emerging Monroe Doctrine framework, cannot access on its own terms.

This structural gap is not coincidental. It is the geological foundation of the US-China energy contest — and the reason why the bipolar condominium proposed at the Beijing summit is, from the American perspective, a temporary management arrangement rather than a permanent settlement.

China controls the materials for the electric transition. The United States is positioning to control the resources for the hydrogen transition. The transition from one to the other — and the currency denomination of each — is the deepest layer of the competition that the summit was beginning to manage.

China controls the materials for the electric transition. The United States is positioning to control the resources for the hydrogen transition.

VII. The Monetary Layer Beneath the Infrastructure

The energy-AI nexus has a monetary dimension that almost no mainstream commentary has registered.

The petrodollar system worked because oil was simultaneously the world’s primary energy source and the denominator of the world’s primary reserve currency. The two functions reinforced each other: countries needed dollars to buy oil, so they held dollars, so dollar demand remained structurally elevated regardless of American fiscal or monetary policy.

This was not deliberate design. It was a structural consequence of denominating the world’s most essential commodity in a single currency.

The AI infrastructure race is beginning to create the conditions for an analogous dynamic — but with energy abundance itself, rather than any specific fuel, as the denominating commodity.

The nations that control the infrastructure capable of sustaining sovereign AI at industrial scale will increasingly set the terms on which others access that infrastructure. If that infrastructure is built under American financial and regulatory frameworks — denominated in dollars, financed through dollar-denominated capital markets, governed by American legal and contractual structures — then the dollar’s reserve function acquires a new foundation precisely as its petrodollar foundation erodes.

SoftBank’s €75 billion commitment to French data center infrastructure is denominated in euros. But the capital markets that financed SoftBank’s capacity to make that commitment, the semiconductor supply chains that will equip those data centers, and the AI model architectures that will run on them are overwhelmingly dollar-denominated.

What we described in The New Tordesillas as the possibility of a future hydrogen-linked monetary architecture runs directly through the infrastructure being built in France and Portugal today — whether or not any individual investor in that infrastructure is aware of it.

This is how structural monetary transitions work. They do not require coordination or conspiracy. They require that the dominant power position itself at the intersection of the world’s most essential commodity and the world’s primary financial architecture — and that it do so before the transition is complete, when the terms can still be set.

VIII. What the AI Revolution Is Actually Revealing

The dominant narrative about artificial intelligence remains a story about algorithms, models, and the companies that build them. That narrative is increasingly a distraction.

Consider what the AI revolution has actually illuminated in the last three years. France’s nuclear fleet — built over fifty years for reasons that had nothing to do with artificial intelligence — is one of the most consequential infrastructure assets in the world for sovereign computation at continental scale. Portugal’s Atlantic geography, shaped by five centuries of maritime positioning, sits at the intersection of transatlantic data connectivity, renewable energy generation, and geological hydrogen optionality in ways that no deliberate twenty-first century planning could have engineered. Quebec’s hydroelectric system, built for industrial purposes in the mid-twentieth century, is now among the most strategically valuable power assets in North America for AI infrastructure development.

None of this was designed. All of it was accumulated — through geology, through historical decisions about infrastructure, through geographic accidents that took on strategic weight only when the dominant technology shifted.

Artificial intelligence is not creating a new geography. It is revealing the strategic value of an old one.

Artificial intelligence is not creating a new geography. It is revealing the strategic value of an old one.