AI in the US "competes with the public for electricity," and nuclear power has become Silicon Valley's "hope of the entire village"

American AI companies have recently started busying themselves with investing in power plants again.

Recently, Meta signed a long-term power purchase agreement with U.S. electric utility Vistra, directly procuring electricity from several of its active nuclear power plants; previously, Meta also collaborated with advanced nuclear energy companies such as Oklo and Terra Power to promote the commercialization of small modular reactors (SMRs) and fourth-generation nuclear technologies.

According to information disclosed by Meta, if these collaborations proceed as planned, by 2035, Meta could lock in a nuclear power supply capacity of up to approximately 6.6 GW (gigawatts, 1 GW=1000 MW/megawatt=1 billion W).

Over the past year, the major investments by North American AI companies in the power sector are no longer news: Microsoft is pushing for the restart of decommissioned nuclear plants, Amazon is deploying data centers around nuclear power stations, and Google, xAI, and others are continuously increasing their long-term power purchase agreements. Against the backdrop of the intensifying computing power race, 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 stimulated by the AI industry is also putting continuous pressure on the U.S. power grid.

According to foreign media reports, driven by a surge in AI demand, the largest U.S. grid operator PJM is facing severe supply and demand challenges. This power network, covering 13 states and serving about 67 million people, is nearing its operational limits.

PJM forecasts that electricity demand will grow at an average annual rate of 4.8% over the next decade, with nearly all new loads coming from data centers and AI applications, while power generation and transmission construction clearly lag behind this pace.

According to the International Energy Agency (IEA), AI has become the most important driver of electricity growth in data centers, and global data center electricity consumption is expected to reach about 945 TWh by 2030, doubling the current level.

The reality mismatch lies in the fact that building AI data centers typically takes only 1–2 years, while a new high-voltage transmission line often requires 5–10 years to be completed. Against this background, AI companies are beginning to take matters into their own hands, initiating a wave of investments in and construction of power plants—an unconventional “infrastructure” boom.

01 AI Giants “Race” to Build Nuclear Power Plants

Over the past decade, the main activity of AI companies in the energy sector has been “buying electricity” rather than “producing electricity”: through long-term power purchase agreements for wind, solar, and some geothermal power, locking in prices and meeting carbon reduction goals.

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

In recent years, as AI electricity consumption surged and grid bottlenecks became apparent, some companies have started to participate in power plant construction or deeply integrate with nuclear power stations, shifting roles from mere electricity consumers to energy infrastructure participants.

One way of participation is to “revive” decommissioned plants. In September 2024, Microsoft signed a 20-year power purchase agreement with nuclear operator Constellation Energy to support the restart of an 835 MW decommissioned nuclear unit and provide long-term power supply.

Alongside Microsoft, the U.S. government also stepped in. In November 2023, the U.S. Department of Energy announced the completion of a $1 billion loan for this project, providing partial financing support. The plant was renamed the Crane Clean Energy Center (originally Three Mile Island Unit 1).

In fact, Crane is not the only plant “re-employed” after retirement. In Pennsylvania, the Eddystone oil and gas power plant was scheduled to retire at the end of May 2024 but was subsequently ordered by the Department of Energy to continue operation to avoid a power shortfall in PJM.

On the other hand, Amazon’s cloud division AWS has taken a different approach, directly purchasing data centers next to nuclear power stations. In 2024, Talen sold its approximately 960 MW data center campus near the Susquehanna nuclear station in Pennsylvania to AWS. In June last year, Talen announced an expanded partnership, planning to supply up to 1,920 MW of carbon-free electricity to AWS data centers.

Regarding new power plants, in recent years, Amazon has participated in the development of Washington State’s SMR small modular nuclear reactor project through investments and collaborations, led by organizations like Energy Northwest. The project involves about 80 MW per unit, with the potential to expand to hundreds of MW, aiming to provide long-term, stable baseload power for data centers.

Google, in cooperation with U.S. nuclear company Kairos Power in 2024, is advancing a new advanced nuclear reactor project, aiming to commission the first units around 2030 and to establish about 500 MW of stable, carbon-free nuclear power by 2035 to support long-term data center operations.

Amid the wave of nuclear plant construction, Meta is one of the most aggressive participants. So far, its planned locked-in nuclear resources have reached 6.6 GW. In comparison, the total installed capacity of operational nuclear plants in the U.S. is about 97 GW.

All these projects are incorporated into Meta’s “Meta Compute” framework—a top-level strategy announced earlier this year to unify planning for future AI computing power and electrical infrastructure.

According to IEA data, by 2030, global data center electricity consumption will double, with AI being the main driver. The U.S. accounts for the largest share of this increase, followed by China.

However, the U.S. Energy Information Administration (EIA)'s previous forecast of maintaining stable power capacity by 2035 has been clearly disrupted by the AI surge.

Based on publicly available information, by 2035, the nuclear capacity directly or indirectly locked in by AI giants like Microsoft, Google, Meta, and AWS is expected to exceed 10 GW, with ongoing new infrastructure projects still being disclosed.

AI is becoming the new “financial backer” for nuclear revival. On one hand, companies’ practical choices—compared to wind and solar—are nuclear power’s advantages of 24/7 stable output, low carbon emissions, and independence from large-scale energy storage; on the other hand, it is closely related to policy environment.

In May 2025, U.S. President Trump signed four “nuclear revival” executive orders, proposing to quadruple U.S. nuclear capacity within 25 years, positioning it as part of national security and energy strategy.

Within a year, the stock prices of nuclear-related companies generally strengthened significantly: represented by operators like Vistra, with stock increases of over 1.5 times; companies focusing on SMRs like Oklo and NuScale saw even more aggressive gains, rising several times.

For a time, under the combined influence of AI industry funding and government promotion, nuclear power has returned to the core of U.S. energy and industrial policy discussions.

02 Fast Model, Slow Power Plant Construction

Although the “nuclear revival” has boosted investment sentiment, nuclear power’s share in the U.S. power generation structure remains only about 19%, and the cycle for new or restarted plants generally spans a decade. In other words, AI’s pressure on the power system has not decreased.

PJM has repeatedly warned in long-term forecasts that in the next ten years, nearly all new loads will come from data centers and AI applications. If power generation and transmission construction cannot be accelerated, the reliability of power supply will face serious challenges.

As one of the largest regional transmission organizations in the U.S., PJM covers 13 states and Washington D.C., serving about 67 million people. Its stable operation is directly related to the core economic zones of the eastern and central U.S.

On one side, many capital investments are flowing into power infrastructure; on the other, power shortages are still unresolved.

Behind this contradiction is a severe mismatch between the rapid expansion of the U.S. AI industry and the pace of power system construction. Building a large-scale AI data center typically takes 1–2 years, while new transmission lines and grid connection approvals often require 5–10 years.

As data centers and AI workloads continue to grow, but new power generation capacity cannot keep pace, the resulting power resource squeeze leads directly to soaring electricity prices.

In areas like Northern Virginia, where data centers are highly concentrated, residents’ electricity prices have risen sharply over the past few years, with some regions experiencing increases of over 200%, far exceeding inflation.

Some market reports show that in the PJM region, as data center loads surge, the capacity market costs have risen significantly: the total capacity auction cost for 2026–2027 is about $16.4 billion, with nearly half of recent costs related to data centers. These rising costs will be passed on to consumers through higher electricity bills.

As public sentiment grows increasingly dissatisfied, the power resource squeeze has quickly spilled over into social issues. Regulators in states like New York have explicitly required large data centers to bear more responsibility for their surging electricity demand and the costs of grid connection and expansion, including higher connection fees and long-term capacity obligations.

“Before ChatGPT appeared, we had never seen such load growth,” said Tom Furlong, chairman of the U.S. Public Power Association. “This is a supply chain issue involving utilities, industry, labor, 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), suggesting that PJM should not approve any new large data center interconnection projects before improving related procedures, citing concerns over reliability and affordability.

To cope with the massive electricity consumption of AI data centers, some U.S. states and utilities have begun establishing dedicated “data center electricity rate categories.” For example, Kansas, in November 2025, introduced new rate rules requiring long-term contracts, electricity cost sharing, and infrastructure cost sharing for large electricity users (such as data centers) with capacities of 75 MW or more, ensuring these large users bear more of the grid connection and upgrade costs.

Microsoft President Brad Smith recently stated in an interview that data center operators should “Pay our way,” paying higher electricity prices or fees for their own power, grid connection, and upgrades, to avoid passing costs onto ordinary electricity consumers.

Overseas, in recent years, regions outside the U.S., such as Amsterdam, Dublin, and Singapore, have suspended many new data center projects mainly due to a lack of sufficient power infrastructure.

Under stricter constraints on power and land, data center expansion has become a test of a country’s underlying infrastructure and capital mobilization capacity. Besides China and the U.S., most economies find it difficult to match such engineering capabilities simultaneously.

Even from the current U.S. power squeeze, it is clear that simply pouring money into new power plants may not resolve the energy crisis of the AI era.

03 Building the Grid, Also “Watching the Sky”

Beyond power plants, a larger structural issue in the power squeeze lies in the long-term lag in U.S. transmission grid construction.

Some industry reports show that in 2024, the U.S. added only 322 miles (about 345 kV and above) of high-voltage transmission lines, making it one of the slowest years in the past 15 years; in 2013, this figure was nearly 4,000 miles.

The lag in transmission capacity means that even if more power plants come online, electricity may not be effectively delivered to high-demand areas due to inability to transmit over long distances.

Between 2023 and 2024, PJM repeatedly warned that due to the slow pace of transmission construction and insufficient generation resources, the rapid growth of new data center loads has forced grid operators to adopt unconventional measures to maintain system stability, including proposals to disconnect some data centers during extreme demand or use backup generation, otherwise reliability risks will further increase.

In contrast, China, known as a “infrastructure maniac,” has maintained high growth and technological iteration in grid construction. In recent years, China has continued to ramp up ultra-high-voltage (UHV) construction, with multiple ±800 kV and 1000 kV UHV lines put into operation between 2020 and 2024, with annual new transmission mileage reaching thousands of kilometers.

In terms of installed capacity, China’s total capacity is expected to exceed 3,600+ GW by 2025, steadily increasing from 2024, with plans to add 200–300 GW of renewable generation capacity annually.

This gap in grid infrastructure capacity cannot be quickly bridged by U.S. policy or capital in the short term.

Amid the surge in AI loads, the U.S. Federal Energy Regulatory Commission (FERC) officially issued Order No. 1920 in May 2024, completing its regional transmission planning reform initiated in 2021. The new rules require utilities to conduct 20-year outlook planning and include new loads like data centers in cost-sharing discussions.

However, due to the long implementation, approval, and construction cycles, this policy functions more as a medium- to long-term “network supplement” tool, and the pressure of power resource shortages will likely persist. Against this backdrop, space-based computing power deployment has become a new industry focus.

In recent years, the global tech industry has been promoting the concept of “space computing power,” deploying AI training/inference-capable computing nodes or data centers in low Earth orbit (LEO) to address bottlenecks in ground-based data centers related to energy, heat dissipation, and connectivity.

Represented by SpaceX, low Earth orbit satellites and inter-satellite laser communication are seen as the foundation for building a distributed “orbital computing network.” SpaceX is exploring in-orbit edge computing with the Starlink constellation for remote sensing processing and real-time inference, reducing ground transmission and energy consumption.

On the other hand, startup Starcloud launched the Starcloud-1 satellite in November 2025, equipped with NVIDIA H100 and completed in-orbit inference validation. This case indicates that space deployment of computing power is entering the practical deployment stage.

China is also accelerating its space computing layout. Led by Zhejiang Laboratory, the “Trisolaran Computing Constellation” has successfully launched the first batch of 12 satellites, with an overall planned computing power reaching the 1000 POPS level, used for orbital edge computing, massive data preprocessing, and AI inference.

However, whether it is space computing or the new energy systems, both are still in early validation stages. This also explains why, over the past year, American AI giants have been competing to invest in nuclear power plants and other power infrastructure.

“We need a clean, reliable power source capable of operating 24/7,” said Fatih Birol, Executive Director of the International Energy Agency, in a recent interview. He stated that “nuclear energy is re-entering the global stage.”

Given the difficulty of expanding the grid and building new power generation in the short term, the current power resource squeeze in the U.S. cannot be quickly alleviated. Large-scale capital investment in the nuclear industry remains the only option at present.

Wood Mackenzie’s latest forecast indicates that, as data centers and AI workloads continue to push electricity demand higher, U.S. nuclear power generation is expected to grow by about 27% after 2035 compared to current levels.

According to foreign media reports, the U.S. government is supporting new reactor construction and unit life extension through loans from the Department of Energy, export credits, and demonstration projects for vendors like Westinghouse, aiming to reshape the nuclear industry capacity.

Under the dual influence of industry and policy, for a considerable period into the future, U.S. AI giants will remain closely tied to the nuclear energy industry.