In 2006, if you wanted to build an internet company, your biggest challenge was compute. You either bought servers or begged for data center space. Then Amazon Web Services launched, and the cloud changed everything. Companies that moved fast on compute won the next decade. In 2026, the bottleneck has shifted. It is no longer about compute. It is about power. TerraPower, the nuclear startup Bill Gates founded in 2008, just became the first next-generation nuclear company to receive U.S. approval to build a reactor. The Nuclear Regulatory Commission voted unanimously to issue a construction permit for TerraPower's Natrium plant in Kemmerer, Wyoming. It is the first new commercial nuclear reactor permit in nearly a decade. This is not a science story. It is a business strategy story. And it may be the most important infrastructure shift since the cloud.

AI has an energy problem

Every major AI company is now, at its core, an energy company in denial. A single hyperscale AI data center demands 300 to 500 megawatts of electricity, roughly equivalent to a mid-sized city. Data centers currently consume about 4.5% of all U.S. electricity. By 2030, that figure is projected to nearly double to somewhere between 6.7% and 12%, according to the Lawrence Berkeley National Laboratory. The International Energy Agency projects global data center electricity consumption will more than double from around 460 TWh in 2024 to over 945 TWh by 2030. This is not a slow-moving trend. It is a crisis of access. The companies building the most capable AI models are the ones burning the most power, and the grid was never designed to handle this kind of demand growth.

The nuclear land grab

Big tech is not waiting around for grid upgrades. They are buying their own power. Microsoft signed a 20-year power purchase agreement worth an estimated $16 billion to restart the Three Mile Island nuclear plant in Pennsylvania, targeting 835 megawatts of capacity by 2028. Google partnered with Kairos Power on the first corporate small modular reactor (SMR) fleet deal in the U.S., targeting 500 megawatts of capacity by 2030. Amazon has invested over $20 billion in nuclear-adjacent infrastructure, including plans to convert its Susquehanna facility into a major AI campus. Meta issued a request for proposals seeking 1 to 4 gigawatts of new nuclear capacity, then signed a deal with TerraPower for up to eight advanced reactors, with the first two aiming to come online by 2032. In early 2026, Meta signed three additional nuclear deals totaling up to 6.6 gigawatts. In total, big tech companies signed contracts for more than 10 gigawatts of new nuclear capacity in the U.S. over the past year alone. These are not pilot projects. This is infrastructure at scale.

Bill Gates' long bet pays off

TerraPower's approval did not come out of nowhere. It came out of nearly two decades of patience. Gates founded TerraPower in 2008 with a conviction that the nuclear industry needed a complete rethink. Traditional reactors are massive, bespoke projects that cost tens of billions and take over a decade to build. The only two American reactors built in the last 30 years, Vogtle Units 3 and 4 in Georgia, cost $35 billion, more than double initial estimates, and arrived seven years behind schedule. TerraPower's Natrium design takes a different approach. The reactor pairs a 345-megawatt sodium fast reactor with a molten salt thermal battery that can boost output to 500 megawatts for over five hours. The system is designed to be built from fabricated components in roughly three years, not assembled on-site over a decade. The company broke ground on its demonstration plant in Kemmerer, Wyoming in 2024, with the goal of splitting atoms by the end of 2030. Its construction permit was approved in just 18 months, far faster than the initially estimated 27 months. TerraPower CEO Chris Levesque has said the company plans to build "hundreds of Natrium reactors." The timeline from founding to first approval: 18 years. That is the kind of long-term bet that only looks obvious in hindsight.

The cloud parallel

There is a useful analogy here. In the mid-2000s, the smartest companies realized that owning or renting compute would determine their speed and scale. Amazon saw this first and built AWS. The companies that adopted cloud early, Netflix, Airbnb, Slack, gained a structural advantage over those that did not. Today, the same logic applies to energy. The companies that secure reliable, scalable, carbon-free power will be the ones that can train the largest models, run the most inference, and deploy AI at the widest scale. Power is the new compute. The difference is that energy is harder to abstract. You cannot spin up a nuclear reactor on demand. Lead times are measured in years, not minutes. That means the decisions being made right now, the deals signed in 2025 and 2026, will determine who has the energy advantage in 2032 and beyond.

The Singapore question

Not every country can play this game. Singapore is a case study in the constraints. The city-state has no nuclear power, limited land, and a tropical climate that makes cooling data centers expensive. Yet it remains deeply committed to being an AI hub. In 2025, Singapore announced plans for a 700-megawatt data center park on Jurong Island and introduced baseline energy efficiency standards for AI data centers. Singapore's approach is different from the American model. Rather than chasing raw power capacity, it is betting on efficiency, green mandates, and regional energy partnerships. The government has mandated green energy requirements for new data centers, positioning itself as a leader in sustainable AI infrastructure. But the fundamental question remains: in a world where AI advantage increasingly depends on energy access, how does a small island nation with no nuclear and limited land compete? Singapore's answer will matter, because it is a preview of the constraints most countries will eventually face.

Second-order effects

If AI companies become the world's largest energy buyers, they do not just consume power. They reshape the politics of power. Microsoft, Google, Amazon, and Meta accounted for roughly half of all global clean energy purchase deals in 2025. When four companies control that much of the clean energy market, they become political players in energy policy whether they intend to or not. Their procurement decisions influence which technologies get funded, which regions get investment, and which regulatory frameworks get prioritized. There is also a deep irony at work. The "clean energy" narrative, which emphasizes patience, environmental review, and long timelines, now collides with the "move fast" ethos of AI development. Nuclear plants take years to permit and build. AI models are being trained on timelines measured in months. The tension between environmental caution and deployment speed is real, and it is only going to intensify. Nuclear also carries genuine risks. Sodium-cooled reactors like TerraPower's Natrium have had historical incidents with coolant leaks. Regulatory scrutiny exists for good reason. And the promise of modular, factory-built reactors has yet to be proven at scale. This is not a done deal. It is a bet, and a big one.

The takeaway

The era of taking electricity for granted is over. For two decades, the tech industry treated power as a commodity, something that showed up when you plugged in. AI has shattered that assumption. The companies that recognized this shift early, that signed nuclear deals and invested in energy infrastructure before it was obvious, will have a structural advantage for the next decade. The ones that waited will find themselves in bidding wars for scarce power, just like the companies that were late to the cloud found themselves scrambling for compute. Nuclear is not the only answer. But it is the only carbon-free energy source that can deliver reliable, baseload power at the scale AI demands. And with TerraPower's approval, it just became a lot more real. The new cloud is not in the sky. It is underground, splitting atoms.

References