US AI companies have recently been busy investing in power plants again.
Recently, Meta signed a long-term power purchase agreement with US power company Vistra to directly procure electricity from several of its operational nuclear power plants; previously, Meta also partnered with advanced nuclear energy companies like Oklo and Terra Power to promote the commercial deployment of small modular reactors (SMRs) and fourth-generation nuclear power technologies.
According to information disclosed by Meta, if the above collaborations proceed as planned, by 2035, Meta could lock in a nuclear power supply scale of up to approximately 6.6GW (gigawatts, 1GW=1000MW/megawatts=1 billion watts).
Over the past year, major energy sector investments by North American AI companies are no longer news: Microsoft pushed for the restart of a decommissioned nuclear power plant, Amazon deployed data centers around nuclear power plants, and Google, xAI, among others, continued to increase long-term power purchase agreements. Against the backdrop of an intensifying computing power race, electricity is transforming from a cost item into 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 sustained "pressure" on the US power grid.
According to foreign media reports, driven by surging AI demand, PJM, the largest grid operator in the US, 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 expects electricity demand to grow at an average annual rate of 4.8% over the next decade, with almost all new load coming from data centers and AI applications, while power generation and transmission construction are clearly not keeping pace with this rhythm.
According to predictions by the International Energy Agency (IEA), AI has become the most important driver of electricity consumption growth in data centers, and it is expected that global data center electricity consumption will rise to about 945TWh by 2030, doubling from current levels.
The现实错位 (practical misalignment) lies in this: the construction cycle for an AI data center typically takes only 1-2 years, while a new high-voltage transmission line often takes 5-10 years to complete. In this context, AI companies have started to get directly involved, initiating a wave of alternative "big infrastructure" projects by investing in and building power plants.
01 AI Giants "Rush to Build" Nuclear Power Plants
Over the past decade, the main action of AI companies on the energy side has been "buying electricity" rather than "making electricity": procuring wind, solar, and some geothermal power through long-term power purchase agreements to lock in prices and meet decarbonization goals.
Taking Google as an example, this AI/internet giant has signed dozens of gigawatts of wind and solar long-term power purchase agreements globally and partnered with geothermal companies to obtain stable clean electricity for its data centers.
In recent years, with the surge in AI electricity consumption and the emergence of grid bottlenecks, some companies have begun to shift towards participating in power plant construction or deeply integrating with nuclear power plants, transforming their role from mere electricity consumers to participants in energy infrastructure.
One way to participate is to "resurrect" already decommissioned power plants. In September 2024, Microsoft signed a 20-year power purchase agreement with nuclear power operator Constellation Energy to support the restart and long-term power supply of an 835-megawatt decommissioned nuclear unit.
The US government also joined in: last November, the US Department of Energy announced the completion of a $1 billion loan disbursement for the project for partial financing support. The unit was renamed the Crane Clean Energy Center (formerly the Three Mile Island Unit 1 nuclear plant).
In fact, Crane is not the only power plant getting a "second career". In Pennsylvania, the Eddystone oil and gas power plant was originally scheduled to retire at the end of May 2024 but was subsequently ordered by the US Department of Energy to continue operating to avoid a power shortfall in PJM.
On the other hand, Amazon Web Services (AWS) took a different approach by directly purchasing a data center next to a nuclear power plant. In 2024, power company Talen sold its approximately 960-megawatt data center campus adjacent to the Susquehanna nuclear power plant in Pennsylvania to AWS. In June last year, Talen announced an expanded collaboration, planning to supply up to 1,920 megawatts of carbon-free electricity to AWS data centers.
Regarding new power plant construction, Amazon has recently participated in the development of an SMR small modular nuclear power plant project in Washington state through investment and cooperation, advanced by Energy Northwest and other institutions. The single-unit scale is about 80 megawatts, expandable to hundreds of megawatts overall, aiming to provide long-term, stable baseload power for data centers.
As for Google, in 2024 it partnered with US nuclear company Kairos Power to advance plans for new advanced nuclear reactor projects, aiming to put the first batch of units into operation around 2030 and form a stable carbon-free nuclear power supply of about 500 megawatts by 2035 to support long-term data center operation.
In the wave of building nuclear power plants, Meta is one of the most aggressive participants. So far, the scale of nuclear power resources it has planned to lock in has reached 6.6 gigawatts. For comparison, the total installed capacity of operational nuclear power plants in the US is about 97 gigawatts.
These projects are all incorporated into Meta's "Meta Compute" framework—a top-level strategy proposed by Meta earlier this year to uniformly plan the computing power and power infrastructure required for future AI.
Data from the International Energy Agency shows that by 2030, global data center electricity consumption will double, with AI being the main driving factor. The US accounts for the highest proportion of this increase, followed by China.
The US Energy Information Administration's (EIA) previous prediction of "maintaining stability" in power generation capacity by 2035 has clearly been broken by the AI wave.
Based on public information汇总 (summary), by 2035, the nuclear power capacity directly or indirectly locked in by AI giants like Microsoft, Google, Meta, and AWS is expected to exceed 10 gigawatts, and new infrastructure projects are still being disclosed continuously.
AI is becoming the new "golden sponsor" for the nuclear power revival, partly due to practical corporate choices—compared to wind and solar power, nuclear power has the advantages of 24/7 stable output, low carbon, and not relying on large-scale energy storage; it is also closely related to the policy environment.
In May 2025, US President Trump signed four "Nuclear Energy Revival" executive orders, proposing to quadruple US nuclear power production capacity within 25 years, positioning it as part of national security and energy strategy.
In the following year, stock prices of related nuclear power companies generally strengthened significantly: represented by nuclear power operators like Vistra, cumulative stock price increases were普遍 (generally) over 1.5 times; while companies focused on small modular reactors (SMRs) like Oklo and NuScale saw more激进 (aggressive) gains, rising several-fold cumulatively.
For a time, under the monetary offensive of the AI industry and promotion at the government level, nuclear power returned to the core of discussions on US energy and industrial policy.
02 Models Run Fast, But Power Plants Aren't Built Fast
Although the "Nuclear Revival" has boosted investment sentiment, nuclear power currently still accounts for only about 19% of the US power generation mix, and the cycle for building new or restarting plants is generally measured in decades. In other words, the risk of AI crowding out the power system has not decreased.
PJM has warned in multiple long-term forecasts that almost all new load growth in the next decade will come from data centers and AI applications. If power generation and transmission construction cannot accelerate, power supply reliability will face severe challenges.
As one of the largest regional transmission organizations in the US, PJM covers 13 states and Washington D.C., serving a population of about 67 million. Its stable operation is directly related to the core economic zones of the eastern and central US.
On one hand, numerous capital is投入 (invested in) power infrastructure; on the other hand, the electricity squeeze is迟迟得不到缓解 (slow to be alleviated).
Behind this contradiction lies a serious mismatch between the expansion speed of the US AI industry and the construction pace of the power system. The construction cycle for a hyperscale AI data center typically takes 1-2 years, while building new transmission lines and completing grid approval often takes 5-10 years.
Data center and AI power consumption loads continue to increase, while new power generation capacity cannot keep up. Under the持续 (sustained) crowding out of power resources, the direct consequence is soaring electricity prices.
In areas with highly concentrated data centers like Northern Virginia, residential electricity prices have risen significantly over the past few years, with increases exceeding 200% in some areas, far above inflation levels.
Some market reports show that in the PJM region, with the surge in data center load, power capacity market costs have risen sharply: the total capacity cost for the 2026-2027 auction was about $16.4 billion, with data center-related costs accounting for nearly half of the total cost in recent rounds. These increased costs will be borne by ordinary consumers through higher electricity bills.
As public discontent grows, the crowding out of power resources has quickly spilled over into a social issue. Regulatory agencies in states like New York have explicitly demanded that large data centers take on more responsibility for their surging electricity demand and the costs of new grid connections and expansion, including higher connection fees and long-term capacity obligations.
"We've never seen load growth like this before ChatGPT appeared," Tom Falconé, chairman of the Large Public Power Council, has publicly stated. "This is a problem involving the entire supply chain, involving utility companies, industry, labor, and engineers—these people don't just appear out of thin air."
Last November, PJM's market monitor filed a formal complaint with the Federal Energy Regulatory Commission (FERC), suggesting that PJM should not approve any new large data center interconnection projects until relevant procedures are improved, citing reliability and affordability issues.
To cope with the massive electricity consumption of AI data centers, some US states and power companies have begun to establish special "data center electricity price categories." For example, in November 2025, Kansas passed new electricity price rules, setting long-term contract, electricity price sharing, and infrastructure cost-sharing requirements for large power users (like data centers) of 75 megawatts and above, ensuring these large users bear more grid fees 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 corresponding fees for their own electricity use, grid connection, and grid upgrades, avoiding passing the costs on to ordinary electricity users.
Overseas, in recent years, regions outside the US like Amsterdam, Dublin, and Singapore have suspended many new data center projects, mainly due to a lack of corresponding power infrastructure.
Under stricter power and land constraints, data center expansion has become a stress test for a country's underlying infrastructure and capital mobilization capabilities. Apart from the two major powers, China and the US, most economies can hardly match such engineering capabilities simultaneously.
Even from the current electricity squeeze in the US, it's not hard to see: merely throwing money at building new power plants may not necessarily resolve the energy crisis of the AI era.
03 Build the Grid, But Also "Watch the Weather"
Beyond the power plant side, the更大的结构性 (larger structural) problem of the electricity squeeze lies in the long-term lag in US transmission grid construction.
Some industry reports show that in 2024, the US added only 322 miles (345kV and above) of high-voltage transmission lines, one of the slowest construction years in the past 15 years; while in 2013, this number was close to 4000 miles.
Backward transmission capability means that even if more power plants come online, electricity may not be effectively delivered to power-intensive areas due to the inability to transmit it over long distances.
Between 2023 and 2024, PJM repeatedly warned externally that due to the inability to speed up transmission construction and power generation resources failing to keep up, the growth of new data center loads has forced grid operators to take unconventional measures to maintain system stability, including proposing options such as cutting power to some data centers or requiring them to use self-generation during extreme demand, otherwise reliability risks would further intensify.
In contrast, China, known as the "infrastructure maniac," has maintained relatively high growth rates and technological iteration in grid construction. In recent years, China has continued to加码 (ramp up) UHV (Ultra-High Voltage) construction, putting into operation multiple ±800kV, 1000kV UHV lines between 2020 and 2024, with an average annual新增里程 (new mileage) measured in thousands of kilometers.
In terms of installed capacity, China's total installed capacity is expected to exceed 3600+ gigawatts by 2025, growing from 2024, and plans to add 200-300 gigawatts of renewable generation capacity for the full year.
This gap in grid infrastructure capability is not something the US can弥补 (make up for) in the short term through policy or capital.
Against the backdrop of激增 (surging) AI load, the Federal Energy Regulatory Commission (FERC) formally 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 forward-looking planning and include new types of loads like data centers in cost allocation discussions.
However, due to the lengthy process of rule implementation, project approval, and construction cycles, this policy is more like a medium-to-long-term "grid repair" tool, and the pressure from the practical electricity resource squeeze will persist. In this context, deploying computing power in space has become a new direction targeted by the industry.
In recent years, the global tech industry has been promoting the concept of "spatial computing power," which involves deploying computing nodes or data centers with AI training/inference capabilities in low Earth orbit (LEO) to solve the bottlenecks of ground-based data centers in energy, heat dissipation, and connectivity.
Represented by SpaceX, low-orbit satellites and inter-satellite laser communication are seen as the foundation for building a distributed "orbital computing power network." SpaceX is exploring in-orbit edge computing leveraging the Starlink constellation for remote sensing processing and real-time inference, reducing the pressure on ground回传 (downlink) and energy consumption.
On the other hand, startup Starcloud launched the Starcloud-1 satellite in November 2025, equipped with an NVIDIA H100 and completed in-orbit inference verification. This case indicates that deploying computing power in space is有望进入 (expected to enter) the actual deployment stage.
China is also accelerating its布局 (layout) in space computing power. The "Three-Body Computing Constellation" led by the Zhejiang Lab has successfully launched its first batch of 12 satellites, with an official plan for overall computing power reaching the 1000 POPS level, used for orbital edge computing, massive data preprocessing, and AI inference.
However, whether it's space computing power or a new generation energy system, both are still in the early验证 (validation) stages. This also explains why US AI giants have been争先 (scrambling) to invest in power infrastructure like nuclear power plants over the past year.
"We need clean, reliable power sources that can operate continuously, 24/7," International Energy Agency Executive Director Fatih Birol said in an interview, adding that "nuclear power is returning to center stage globally."
Given the reality that grid expansion and power generation construction are difficult to keep up with in the short term, the current crowding out of power resources in the US cannot be quickly alleviated. Continued large-scale capital investment in the power industry, especially the nuclear power sector, remains the only choice for now.
Wood Mackenzie pointed out in its latest forecast that as data center and artificial intelligence loads continue to push up electricity demand, US nuclear power generation is expected to grow by about 27% from current levels after 2035.
According to foreign media reports, the US government is supporting nuclear power equipment suppliers like Westinghouse through Department of Energy loans, export credits, and demonstration projects, promoting the construction of new reactors and unit life extension upgrades, and重塑 (reshaping) nuclear power industrial capabilities.
Under the dual background of industry and policy drive, for a considerable period in the future, US AI giants will be tightly捆绑 (bundled) with the nuclear energy industry.
