Meta, Google, and Microsoft Invest in Nuclear Power to Fuel AI Expansion

Recently, U.S. AI companies have once again been busy investing in power plants.

Recently, Meta signed a long-term power purchase agreement with U.S. power company Vistra, directly procuring electricity from multiple of its currently operating nuclear power plants. Previously, Meta had also partnered with advanced nuclear energy companies such as Oklo and TerraPower to promote the commercial deployment of small modular reactors (SMRs) and fourth-generation nuclear technologies.

According to information disclosed by Meta, if the aforementioned collaboration proceeds as planned,By 2035, the maximum scale of nuclear power supply that Meta could potentially secure may reach up to approximately 6.6 GW (gigawatts, 1 GW = 1,000 MW/megawatts = 1 billion watts).

In the past year, bold moves by North American AI companies in the power sector have become less surprising: Microsoft is promoting the restart of retired nuclear power plants, Amazon is deploying data centers around nuclear power stations, and Google, xAI, and others continue to increase their long-term power purchase agreements.Against the backdrop of the intensifying computing power competition, electricity is shifting from a cost item to a strategic resource that AI companies must secure in advance.

On the other hand, the energy demand driven by the AI industry continues to put pressure on the U.S. power grid.

According to foreign media reports, driven by a surge in demand for AI, PJM, the largest power grid operator in the United States, is facing a severe supply and demand challenge. This power network, which spans 13 states and serves approximately 67 million people, is approaching its operational limits.

PJM expects electricity demand to grow at an average annual rate of 4.8% over the next decade, with almost all the additional load coming from data centers and AI applications, while power generation and transmission construction clearly cannot keep up with this pace.

According to the International Energy Agency (IEA) forecast, AI has become the most significant driver of electricity consumption growth in data centers. It is projected that global data center electricity consumption will rise to approximately 945 TWh by 2030, doubling the current level.

The misalignment in reality lies in the fact that the construction cycle for AI data centers typically lasts only 1–2 years, whereas a new high-voltage transmission line often requires 5–10 years to be completed.Against this backdrop, AI companies have started to take direct action, launching an alternative "major infrastructure" wave by investing in and building power plants.

Over the past decade, the main approach of AI companies in the energy sector has been "buying electricity" rather than "generating electricity." They have primarily acquired wind, solar, and some geothermal power through long-term power purchase agreements to lock in prices and meet their carbon reduction goals.

Take Google as an example. This AI/internet giant has signed long-term power purchase agreements for wind and solar power on a scale of tens of gigawatts globally, and has also partnered with geothermal companies to secure stable and clean electricity for its data centers.

In recent years, as AI electricity consumption has surged and grid bottlenecks have become apparent, some companies have begun to participate in power plant construction or form deep partnerships with nuclear power stations. Their role has shifted from being mere electricity consumers to participants in energy infrastructure.

One way to participate is to "revive" retired power plants. In September 2024, Microsoft signed a 20-year power purchase agreement with nuclear power operator Constellation Energy to support the restart of an 835-megawatt retired nuclear reactor and ensure its long-term electricity generation.

The U.S. government also joined Microsoft in this initiative. In November last year, the U.S. Department of Energy announced the completion of a $1 billion loan transaction for the project, providing partial financial support. The unit was renamed the Crane Clean Energy Center (previously Unit 1 of the Three Mile Island Nuclear Station).

In fact, Crane is not the only power plant to be "laid off and re-employed." In Pennsylvania, the Eddystone fossil fuel power plant was originally scheduled to retire by the end of May 2024, but was then urgently ordered by the U.S. Department of Energy to continue operating to avoid a power shortfall in the PJM region.

On the other hand, Amazon's cloud computing division, AWS, has taken a different approach by directly acquiring a data center located next to a nuclear power plant. In 2024, the power company Talen sold a 960-megawatt data center campus situated near the Susquehanna nuclear power plant in Pennsylvania to AWS. In June of last year, Talen further expanded the collaboration, announcing plans to supply up to 1,920 megawatts of carbon-free electricity to AWS data centers.

In the new power plant sector, in recent years, Amazon has participated in the development of the SMR (Small Modular Reactor) nuclear power project in Washington State through investments and partnerships, advanced by organizations such as Energy Northwest. Each unit has a capacity of approximately 80 megawatts, and the overall system can be expanded to hundreds of megawatts, aiming to provide long-term, stable baseload power for data centers.

On the Google side, in 2024, it partnered with the U.S. nuclear energy company Kairos Power to advance the development of new advanced nuclear reactor projects. The plan aims to bring the first units into operation around 2030 and establish a stable, carbon-free nuclear power supply of approximately 500 megawatts by 2035, to support the long-term operation of data centers.

In the wave of building nuclear power plants, Meta is one of the most aggressive participants. To date, the planned nuclear power resources it has secured amount to 6.6 gigawatts. For comparison, the total installed capacity of currently operating nuclear power plants in the United States is approximately 97 gigawatts.

These projects have all been incorporated into Meta's "Meta Compute" framework—a top-level strategy introduced by Meta at the beginning of this year to unify the planning of computing power and power infrastructure required for future AI.

According to the International Energy Agency, global data center electricity consumption will double by 2030, with AI being the primary driving factor. The United States will account for the largest share of this increase, followed by China.

The U.S. Energy Information Administration's (EIA) previous prediction of "stable" power generation capacity by 2035 is clearly being disrupted by the AI boom.

According to aggregated public information, by 2035, AI giants such as Microsoft, Google, Meta, and AWS are expected to directly or indirectly secure nuclear power generation capacity exceeding 10 gigawatts, with new infrastructure projects still being continuously disclosed.

AI is becoming the new "financier" for the revival of nuclear power. On one hand, it represents a practical choice for enterprises ——Compared to wind and solar power, nuclear power has the advantages of stable 7×24 hour power generation, low carbon emissions, and does not rely on large-scale energy storage.It is also closely related to the policy environment.

In May 2025, U.S. President Trump signed four executive orders on "Nuclear Energy Revival," proposing to quadruple the United States' nuclear power capacity within 25 years, positioning it as part of the nation's national security and energy strategy.

In the following year, the overall stock prices of companies related to nuclear power showed a significant upward trend. For example, nuclear power operators like Vistra generally saw their stock prices rise by more than 1.5 times. Companies focusing on small modular reactors (SMRs), such as Oklo and NuScale, experienced even more aggressive growth, with cumulative increases reaching several times their original value.

For a time, under the financial push from the AI industry and government-level promotion, nuclear power returned to the core discussions of U.S. energy and industrial policy.

Although a "nuclear renaissance" has boosted investment sentiment, nuclear power currently accounts for only about 19% of electricity generation in the United States, and the construction or restart of new plants typically takes a decade or more. In other words, the risk of AI straining the power system has not decreased.

PJM has warned in multiple long-term forecasts that almost all of the new load over the next decade will come from data centers and AI applications. If the construction of power generation and transmission cannot be accelerated, the reliability of power supply will face significant challenges.

As one of the largest regional transmission organizations in the United States, PJM covers 13 states and the District of Columbia, serving a population of approximately 67 million. Its stable operation is directly related to the core economic regions of the U.S. East and Midwest.

On one side, numerous capitals are investing in power infrastructure; on the other side, the power shortage remains unresolved for a long time.

Behind this contradiction lies a significant mismatch between the expansion speed of the U.S. AI industry and the pace of power infrastructure development. The construction cycle for an ultra-large-scale AI data center typically takes 1–2 years, while building new transmission lines and completing grid connection approvals often require 5–10 years.

The electricity consumption of data centers and AI is continuously increasing, but the newly added power generation capacity cannot keep up. Under the ongoing strain on power resources, the direct consequence is a sharp rise in electricity prices.

In regions with a high concentration of data centers, such as Northern Virginia, residential electricity prices have risen significantly over the past few years, with some areas experiencing increases of more than 200%, far exceeding the inflation rate.

Some market reports indicate that in the PJM region, power capacity market costs have significantly increased due to a surge in data center loads:The total capacity cost for the 2026–2027 auction is approximately $16.4 billion, and data center-related costs have accounted for nearly half of the total costs in recent rounds. These rising costs will be passed on to ordinary consumers through higher electricity bills.

As public sentiment turns increasingly dissatisfied, the strain on power resources has quickly spilled over into a broader societal issue. Regulatory authorities in New York State and other regions have clearly stated that large data centers must take on more responsibility for their surging electricity demands and the associated grid connection and expansion costs. This includes paying higher connection fees and fulfilling long-term capacity obligations.

"Before the emergence of ChatGPT, we had never seen such a surge in demand," Tom Falcone, chairman of the U.S. large public power committee, once publicly stated. "This is an issue that affects the entire supply chain, involving utility companies, industries, workforce, and engineers—people who don't just appear out of nowhere."

In November last year, PJM's market regulator filed a formal complaint with the Federal Energy Regulatory Commission (FERC), recommending that PJM should not approve any new large data center interconnection projects until relevant procedures are improved, citing concerns over reliability and affordability.

To address the massive electricity consumption of AI data centers, some U.S. states and utility companies have begun establishing dedicated "data center electricity rate categories." For example, in November 2025, Kansas implemented new rate rules that set long-term contract requirements, cost allocation for electricity rates, and infrastructure cost-sharing obligations for large power users (such as data centers) with consumption of 75 megawatts or more. These rules ensure that such large users bear a greater share of grid fees and upgrade costs.

Microsoft President Brad Smith recently stated in an interview that,Data center operators should "pay their way," by paying higher electricity rates or corresponding fees for their own power consumption, grid connection, and grid upgrades, to avoid shifting the costs onto regular electricity users.

Meanwhile, overseas in recent years, regions outside the United States such as Amsterdam, Dublin, and Singapore have suspended many new data center construction projects, mainly due to a lack of adequate power infrastructure.

Under stricter constraints on power and land, the expansion of data centers has become a stress test for a country's fundamental infrastructure and its ability to mobilize capital. Apart from the two major economies, the United States and China, most other economies struggle to simultaneously match such engineering capabilities.

Even from the current electricity shortages in the United States, it is not hard to see that simply investing money to build new power plants may not be enough to resolve the energy crisis of the AI era.

Beyond the power plant side, a larger structural issue contributing to the electricity shortage lies in the long-term lag in the development of the U.S. transmission grid.

Some industry reports indicate that in 2024, the United States added only 322 miles of high-voltage transmission lines (345 kV and above), making it one of the slowest construction years in the past 15 years; in contrast, this figure was close to 4,000 miles in 2013.

Lagging transmission capacity means that even if more power plants come online, electricity may not be effectively delivered to densely populated areas due to the inability to transmit it over long distances.

Between 2023 and 2024, PJM repeatedly issued public warnings that the inability to accelerate transmission construction and the lack of sufficient generation resources have forced the grid operator to adopt unconventional measures to maintain system stability. These measures include proposing options such as disconnecting certain data centers or requiring them to use self-generated power during periods of extreme demand. Otherwise, the risk to grid reliability will further intensify.

By contrast, China, known as the "infrastructure wizard," has consistently maintained a high growth rate and technological advancement in power grid construction. In recent years, China has continuously accelerated the development of ultra-high-voltage (UHV) transmission. Between 2020 and 2024, multiple ±800kV and 1000kV UHV transmission lines were put into operation, with thousands of kilometers of new transmission lines added annually.

In terms of installed capacity, China's total installed capacity is expected to exceed 3,600+ gigawatts in 2025, showing a steady increase from 2024, with plans to add 200–300 gigawatts of new renewable power generation capacity for the entire year.

This gap in power grid infrastructure capabilities cannot be easily bridged in the short term by U.S. policies or capital investment alone.

Against the backdrop of a surge in AI workloads, the U.S. Federal Energy Regulatory Commission (FERC) officially issued Order No. 1920 in May 2024, completing its regional transmission planning reforms initiated in 2021.The new regulations require utilities to conduct 20-year forward-looking planning and include new types of loads, such as data centers, in cost allocation discussions.

However, due to the lengthy process of implementing regulations, project approvals, and construction cycles, this policy is more of a medium- to long-term "infrastructure enhancement" tool. In reality, the pressure from shortages in power resources will continue. Against this backdrop, deploying computing power in space has become a new direction that the industry is focusing on.

In recent years, the global technology industry has been promoting the concept of "space computing," which involves deploying computing nodes or data centers with AI training and inference capabilities in low Earth orbit (LEO). This aims to address bottlenecks related to energy consumption, heat dissipation, and connectivity faced by ground-based data centers.

Represented by SpaceX, low Earth orbit satellites and inter-satellite laser communication are considered the foundation for building a distributed "orbital computing network." SpaceX is leveraging its Starlink constellation to explore in-orbit edge computing, which is used for remote sensing processing and real-time inference, thereby reducing the pressure on ground transmission and energy consumption.

On the other hand, startup Starcloud has launched the Starcloud-1 satellite in November 2025, equipped with an NVIDIA H100 GPU and successfully completing in-orbit inference validation. This case demonstrates that deploying computing power in space is moving closer to practical implementation.

China is also accelerating its layout in space computing power. The "Three-Body Computing Constellation," led by the Zhejiang Lab, has successfully launched its first batch of 12 satellites. According to official plans, the overall computing power will reach the level of 1000 POPS, used for edge computing in orbit, massive data preprocessing, and AI inference.

However, whether it's space-based computing power or the next-generation energy system, both are still in the early verification stages. This also explains why, over the past year, major U.S. AI companies have been vying to invest in power infrastructure such as nuclear power plants.

"We need a clean and reliable source of power that can operate continuously 24 hours a day, seven days a week," said Fatih Birol, the Executive Director of the International Energy Agency, in a previous interview. He added that "nuclear energy is once again moving to center stage globally."

Given the reality that grid expansion and power generation construction cannot keep up in the short term, the current strain on electricity resources in the United States cannot be quickly alleviated. Therefore, continuing to make large-scale capital investments in electricity, especially in the nuclear power industry, remains the only viable option at present.

Wood Mackenzie pointed out in its latest forecast that, with data centers and artificial intelligence workloads continuously driving up electricity demand, nuclear power generation in the United States is expected to increase by about 27% from current levels after 2035.

According to foreign media reports, the U.S. government is supporting nuclear equipment suppliers such as Westinghouse through Department of Energy loans, export credits, and demonstration projects, promoting the construction of new reactors and the life extension and upgrading of existing units, and rebuilding the country's nuclear industry capabilities.

Against the dual backdrop of industry trends and policy drivers, U.S. AI giants will be closely tied to the nuclear energy industry for a considerable period of time in the future.