Buying electricity, hoarding electricity and building power plants, US AI giants are almost forced to turn into power companies.

A nuclear power plant that has been out of operation for 6 years is preparing to restart because of AI. 

On September 8, the U.S. Department of Energy announced that it will provide a loan of up to 1.9 billion U.S. dollars to NextEra Energy for the restart of the Duane Arnold Nuclear Power Plant in Iowa. This nuclear power plant ceased operation in 2020. It is the only nuclear power plant in Iowa with an installed capacity of 615 megawatts. 

According to the plan, it will be reconnected to the grid in 2029, provided that it passes the approval of the U.S. nuclear regulatory authority. Behind this restart, there is also a signed long-term contract: Google promises to purchase most of the electricity generated by this nuclear power plant after restart for 25 years. 

This is not the first nuclear power plant in the United States that is preparing to be "resurrected" because of AI. Microsoft has long signed a 20-year long contract to promote the restart of Unit 1 of the Three Mile Island Nuclear Power Plant, which was shut down in 2019. 

In addition to traditional nuclear power, Meta, Amazon, and Google are also accelerating investment in small modular reactor nuclear power projects. 

The more radical Elon Musk is no longer satisfied with purchasing electricity. xAI is building its own 1.2 GW natural gas power generation facility, and Musk also plans to let SpaceX produce key components of gas turbines on its own. 

From nuclear power to geothermal energy, from natural gas to small nuclear reactors, AI giants are extending their reach to positions closer and closer to the upstream of the energy industry.

This technological competition of AI, which seems to compete in models, chips and computing power, has unexpectedly come to the stage of competing for electricity in the end. 

AI is advancing rapidly, and the U.S. power grid can no longer withstand the pressure 

A large number of AI data centers are pouring into the U.S. power grid at the same time. 

According to Reuters, in some power markets in the Midwest, Mid-Atlantic and southern regions of the United States, large power-consuming projects are queuing up to apply for grid connection, and the total applied power supply capacity has exceeded 700 gigawatts, the vast majority of which comes from data centers. This figure is already 10 times the actual electricity load of all data centers in the United States. 

Of course, applying for so much power does not mean that it will actually be used that much in the end. A large number of projects that have not really landed are mixed in, and even some data centers just grab a grid connection quota first. Now, it is almost impossible to figure out "how many data centers really need electricity".

Texas is the most typical example. As of this year, ERCOT has received grid connection applications for large power-consuming projects with a total required power supply capacity of hundreds of gigawatts, about 90% of which are related to data centers. However, the peak load of the Texas power grid itself only reached 91.1 gigawatts on July 22 this year. 

Texas has suspended some new grid connection applications for data centers, requiring these projects to be re-reviewed to see which ones are really going to be built and which ones are just occupying power grid resources in advance. 

"Occupying the resources without using electricity" is a term that has frequently appeared in the U.S. energy market recently: "phantom demand". However, even if there is exaggeration in the application figures, the actual electricity demand in the United States is indeed growing. 

The International Energy Agency predicts that from 2025 to 2030, the overall electricity demand in the United States will grow by nearly 2% on average every year, more than twice the average growth rate of the past decade. A considerable part of this growth is attributed to data centers. 

By 2030, the electricity consumption of U.S. data centers is expected to continue rising rapidly, and may even account for nearly half of the new electricity demand across the United States. 

What is more troublesome is that a large number of power generation projects in the United States are under preparation, but every link from power generation to power grid and then to data centers is queuing up. By the end of 2025, there are still about 8,200 projects in the United States waiting to be connected to the grid, involving 1,312 gigawatts of power generation capacity and about 749 gigawatts of energy storage capacity. 

These projects are not AI data centers themselves, but various new power supply projects such as wind power, solar power, natural gas power generation and energy storage. In other words: while the United States is frantically increasing electricity demand, it has not even had time to connect new power sources to the power grid. 

What is more troublesome is that the aging U.S. power grid is almost unable to hold on.

A large number of transmission lines, transformers and switchgear in the United States have been in operation for decades. Data cited by the American Society of Civil Engineers shows that 70% of power transformers have been in operation for more than 25 years, and 70% of transmission lines have also exceeded 25 years of service life. 

But now, it is impossible to replace the transformers of the power grid in time. In 2024, the procurement cycle for new transformers in the United States has reached 80 to 210 weeks, with an average of about 120 weeks. 

So what is in front of U.S. AI now is not simply "whether the power generation capacity is enough". Even if power is found, whether it can be delivered to the data center is another matter. 

All means are used to compete for power supply 

When it is getting harder and harder to wait for the public power grid, AI giants have to find other ways. The most direct way is to seize the power supply near the data center first.

At the beginning of September, Google signed a 396 MW power purchase agreement with geothermal energy company Fervo, which is one of the largest enhanced geothermal power purchase agreements in the world so far. It is expected to supply power to the data center planned by Google in Utah starting from 2028. 

396 megawatts is 396,000 kilowatts. But what is interesting is that the data center in Utah has not even been fully confirmed to be completed, but Google has already locked in the future power supply in advance. 

Amazon went a step further and directly bought the data center park next to the power station.

In 2024, Amazon spent 650 million U.S. dollars to buy the data center park with a planned capacity of 960 megawatts next to the Susquehanna Nuclear Power Plant in Pennsylvania from U.S. energy company Talen Energy. In 2025, the two sides further expanded their cooperation. The scale of nuclear power supply that Amazon plans to obtain will eventually reach 1920 megawatts, and the contract will last until 2042. However, the expanded cooperation will shift to power supply through the public power grid. 

Another way to increase power supply is to restart the power stations that have been shut down.

In 2024, Microsoft signed a 20-year power purchase agreement with Constellation Energy to promote the restart of Unit 1 of the Three Mile Island Nuclear Power Plant in Pennsylvania. This unit has an installed capacity of 835 megawatts, which is equivalent to the electricity consumption of 700,000 households. 

It was shut down in 2019 for economic reasons. The company currently plans to resume power generation in 2027, and still needs to complete relevant approvals. Microsoft will purchase all the electricity generated by the unit after restart. 

From geothermal energy to nuclear power, from purchasing data centers to restarting shut-down nuclear power plants, the power purchase agreements signed by AI giants are more and more like a long-term energy lock. Agreements that last for 10 or 20 years almost reserve most of the power generation capacity of a large power station for their own data centers in advance. 

When it comes to Elon Musk, his approach is more radical. He even thinks buying power is not fast enough, and plans to generate power on his own.

xAI's Colossus data center in the Greater Memphis area previously used a large number of gas turbines directly for power supply. Now, xAI is gradually replacing the temporary power supply equipment located in Southaven, Mississippi with a permanent 1.2 GW natural gas power generation facility, and plans to install 41 permanent gas turbines. Previously, the Southaven power supply facility once deployed as many as 69 temporary gas turbines. 

But Musk is not satisfied, and even plans to produce power generation equipment on his own.

Recently, The Wall Street Journal revealed that SpaceX is building a casting facility in Bastrop, Texas, and plans to produce the blades and deflectors required for large natural gas turbines. 

Musk hopes to take this key supply chain into his own hands to shorten the production cycle of gas turbines. 

Is the end of U.S. AI development infrastructure construction? 

But the problem is, they are buying power, generating power, and even planning to build gas turbines. Why do AI giants go through so much trouble? 

Because even if you really find a power station that is willing to supply power to you, sending that power to the data center still has to go through the U.S. aging power grid in most cases. 

In recent years, the terrifying electricity demand brought by AI not only overwhelms the old power grid, but also makes the equipment needed for power grid transformation in short supply.

Data from Wood Mackenzie shows that between 2019 and 2025, the U.S. demand for generator step-up transformers increased by 274%, while the demand for substation power transformers increased by 116%. The delivery cycle of some large transformers has reached 3 to 5 years; the waiting time for some high-voltage switchgear has also exceeded two years. 

What is worse is that Trump signed an executive order during his first term, requiring restrictions on power equipment transactions with "foreign adversaries" that pose security risks; in 2025, U.S. senators even wrote letters demanding to "kick China out" of U.S. critical infrastructure. 

However, the U.S. power equipment supply still cannot do without China. In 2024, the total value of electrical products such as transformers, converters and inductors imported by the United States from China was about 4 billion U.S. dollars. According to Bloomberg, in 2025, the United States is still the largest buyer of Chinese transformers. 

Since the upgrade and transformation of the power grid cannot be waited for, it is better to move the power plant next to the data center. To be more radical, just like xAI, build a natural gas power plant next to the data center on your own. 

For AI companies, this is actually bypassing some of the slowest links: reducing the dependence on new transmission lines and public power grid expansion, and not pinning all hopes on when the public power grid can deliver power. 

But self-generating power cannot be built immediately whenever you want, first of all you have to get the gas turbines. 

Right now, gas turbines in the United States are also becoming scarce. As AI data centers begin to shift to natural gas power generation in large numbers, the industry supply is becoming increasingly tight, and the delivery cycle of some gas turbines has even been extended to several years. At the same time, the production capacity of major manufacturers such as GE Vernova has been largely locked by the demand for AI infrastructure. 

Unable to wait, Elon Musk directly asked SpaceX to manufacture gas turbine blades. 

One of the bottlenecks in the production of natural gas power generation equipment is the casting of blades and deflectors. It requires the use of high-temperature resistant nickel-based superalloys, and some even involve single crystal blade manufacturing. There are very few suppliers in the world that can mass-produce such parts, and almost all of them are fully occupied by orders now. 

According to Elon Musk, if SpaceX succeeds, the gas turbines can be put into operation up to 18 months earlier. 

From buying power to building data centers next to power stations; from restarting nuclear power plants to building natural gas power plants on your own; and finally even making key parts of gas turbines by yourself. 

This AI technology competition has unexpectedly come all the way to compete for power plants, transmission lines, gas turbines, and even a metal blade. 

Interestingly, Trump wants manufacturing to return to the United States, but the power shortage of AI has completely exposed the true weakness of U.S. manufacturing, and even its infrastructure.

This article is from the WeChat official account "Blue Print Plan", author: Chester, published by 36Kr with authorization.