The U.S. power grid is breaking under the weight of artificial intelligence, and the world's most valuable companies have already decided to route around it. Microsoft restarted Three Mile Island. Amazon paid $650 million to co-locate directly with a nuclear plant. Google signed agreements for small modular reactors. Meta issued a request for proposals seeking up to 4 gigawatts of new nuclear capacity. These are not the moves of companies that believe the existing utility model will work.

What's Driving This Nuclear Pivot?

The numbers tell the story. A single ChatGPT query consumes roughly 10 times the energy of a Google search. Training the next generation of large language models (LLMs), which are AI systems trained on vast amounts of text data, requires the equivalent power draw of small cities. McKinsey forecasts that AI data center capital expenditure will reach roughly $5.2 trillion between now and 2030. Goldman Sachs Research projects global data center power demand will surge up to 165 percent by 2030 compared to 2023 levels.

The American electrical grid was designed for a world where electricity demand grew at one to two percent per year. That's a reliable, manageable rate that allowed utilities to plan their generation and transmission decades in advance. AI's energy appetite has no precedent in this framework. The grid was not built for it, the utilities running the grid have no plan for it, and the regulatory framework governing the utilities cannot move fast enough to keep up.

The evidence is mounting. Berkeley Lab recently found that more than 70 percent of grid interconnection requests in the United States are ultimately withdrawn because the grid simply cannot accommodate them. Kevin O'Leary, the Shark Tank investor and longtime infrastructure backer, has gone further, estimating that 50 percent of the data centers currently planned across the United States will never be built, not for lack of money or demand, but because the grid cannot deliver the power.

Why Are Local Communities Blocking Data Centers?

Even when the power is theoretically available and utilities are willing, local opposition is killing projects. In September 2025, a developer proposed a $12 billion data center complex in St. Joseph County, Indiana, which would have been the largest single project investment in state history. The developer had the capital, identified land, and support from county economic development officials. The Local Area Plan Commission voted 7-0 against the project. Community members raised concerns about water demands, electricity demands, tax impacts, and the conversion of farmland to industrial use. Sixteen single-family homes and two family farms would have been displaced.

This scenario is repeating across the American data center pipeline. The power load strains regional grids. Water demand for cooling competes with municipal supply. Property tax abatements shift cost burdens onto existing residents. Truck traffic and construction disruption rattle quiet communities for years. None of it sits well with neighbors who do not directly benefit from the AI boom but bear every one of the local costs.

How Are Tech Giants Securing Their Power Supply?

• Direct Nuclear Partnerships: Microsoft signed a 20-year deal to restart Three Mile Island, while Google signed agreements with Kairos Power for small modular reactors, which are nuclear reactors smaller than traditional plants but capable of powering large facilities.
• Co-Location with Existing Plants: Amazon paid $650 million for a data center campus to co-locate directly with the Susquehanna nuclear station in Pennsylvania, guaranteeing AI-grade power without waiting on utility queues.
• Massive Nuclear Procurement: Meta issued a request for proposals seeking up to 4 gigawatts of new nuclear capacity, essentially building the generation infrastructure itself rather than relying on utilities.
• International Power Agreements: Bitzero Holdings, a Canadian-listed Bitcoin miner with infrastructure in Norway, signed a binding letter for a 15-year, $2.6 billion lease, securing access to hydroelectric power in jurisdictions where the grid works.

These moves reflect a fundamental shift in how the largest technology companies view power security. Microsoft did not pay to restart Three Mile Island because it wanted to be in the nuclear business. It did it because the grid could not deliver the power it needed, on the timeline it needed, for any price. Amazon did not pay $650 million for a co-located data center because it loves nuclear chemistry. It did it because direct co-location was the only way to guarantee AI-grade power without waiting on a utility queue.

Smaller AI companies cannot afford to restart nuclear plants or co-locate with reactors. As the hyperscalers, or largest technology companies, route around the grid, every other AI player in America gets squeezed harder by what is left. The result is a pipeline of proposed AI capacity that gets thinner with every passing month. The big projects that survive get bigger. The marginal projects die in committee. The map of where AI data centers can actually get built shrinks.

Where Is AI Power Infrastructure Actually Being Built?

The Nordic countries, especially Norway and Finland, emerged as the escape valve for hyperscalers shut out of the U.S. grid. Both have massive hydroelectric and nuclear generation, cold climates that slash cooling costs, stable governments, and EU data sovereignty protections that make them ideal for AI workloads. For a few years, the Nordics looked like the solution. That window has now mostly closed.

Norway has capped new data center operators at just 5 megawatts of initial allocation, which is barely enough to run a small Bitcoin mining operation, let alone an AI training cluster. Finland and Sweden are tightening as well. The companies that secured significant Nordic power capacity before the AI boom kicked in are sitting on something that cannot be replicated, no matter how much capital is thrown at it. Bitzero is one of those companies. The company is a licensed grid operator in Norway with direct connections to hydroelectric power plants, all-in power costs of roughly 3 to 4 cents per kilowatt-hour, and more than 1 gigawatt of secured capacity across four sites in Norway, Finland, and the U.S. .

The broader implication is clear: as the U.S. grid strains under AI's power demands, the companies that already own AI-grade power capacity in jurisdictions where the grid works are gaining an unprecedented competitive advantage. The hyperscalers are betting billions on nuclear because they have concluded that the existing utility model cannot deliver. They are spending billions to route around it, leaving smaller competitors to fight over the scraps of available grid capacity.