Elon Musk Warns of AI "Power Shortage": At Least 15 Gigawatts Gap in 2027
On September 10-11, Google announced it would invest at least 13 billion euros in Finland over two years to build AI infrastructure and signed a 22-year nuclear power purchase agreement with local energy company Fortum; In the same week, Musk's xAI campus was building a 1.2 GW permanent natural gas power plant; Meanwhile, A-share power and "computing-power synergy" strengthened against the trend on September 10, with Mindong Power achieving two consecutive limit-ups. The capital market reacted ahead of public opinion.
In the past three years, AI's bottleneck has shifted from HBM and advanced packaging to power equipment and grid connection qualifications. GPUs can be delivered in months, but a large transformer takes two to five years. The real meaning of the 15 gigawatt gap is that a batch of machines that have already been paid for will next wait in warehouses for power.
How Is 15 GW Calculated?
The lifecycle of a large model consists of two stages: training and inference: On the training side, tens of thousands of graphics cards run in parallel for long periods, demanding a stable power supply. For example, GPT-4 consumes over 50 million kWh of power per training session; On the inference side, as AI connects to various terminal applications, data centers ×need to compute 24/7 without interruption to continuously extract power.
The combination of these two demands triggers the Jevons paradox in energy economics: technological breakthroughs have greatly reduced the energy consumption per computing cycle, but lowering the threshold has led to hundreds, thousands, or even thousands of times increase in call volume, resulting in a rise in total societal electricity consumption. The International Energy Agency predicts that global data center electricity consumption will nearly double by 2030; The U.S. Energy Information Administration calculates even more specifically: by 2030, data centers may consume 11.8% of the nation's electricity, and in extreme cases, 15.3%. Wood Mackenzie estimates that U.S. data center capacity will grow from about 24 gigawatts in 2026 to 110 gigawatts in 2030, a roughly 3.6-fold increase over four years.
But the supply curve can't be lifted that quickly. Musk's "10% to 20%" refers to the annual growth rate of electricity supply outside China: it is constrained by three things: power source construction, grid expansion, and grid connection approvals, none of which operate at the speed of chip iteration. A demand line with a 40% annual growth collides with a supply line that grows 15% annually; the gap is not in power plants, but in the slope difference between the two.
This isn't the first time Musk has said this. When the AI boom first started in 2023, he predicted the US would shift from chip shortages to power shortages; He then added two layers: not only is power generation insufficient, but grid equipment like transformers is also insufficient; Later, he proposed that large-scale energy storage is the real cure for the grid. At the time, these judgments were considered marginal views, but now they have become numbers on project schedules.
The most convincing evidence comes from his own company: xAI's supercomputer consumes 1 GW of power, hardware is built very quickly, but connecting to the local grid takes a year and can only be temporarily powered by gas turbines; Currently, xAI campus is building a 1.2 GW permanent natural gas power station, and licenses for 41 permanent gas turbines have been obtained. An even more extreme step upstream in the supply chain: Musk discovered that high-temperature alloy blades were blocking gas turbine output, so he decided to build an internal foundry at SpaceX to manufacture this component. According to his estimate, vertical integration could shorten gas turbine delivery times by about 18 months
A company is pushing from models all the way to gas turbine blades to obtain electricity, which is the most direct confirmation that "electricity is the final hurdle for AI." Industry confirmation is also emerging: GE Vernova's data center electrification orders in the first half of this year have exceeded $5 billion, more than double the total for all of 2025; Its gas turbine backlog is 116 GW, with company guidance to raise this to 125 GW by year-end. AI orders are spreading from chips to gas turbines, transformers, power grid equipment, and cooling systems.
The Real Bottleneck Is Not Power Generation, But "Power Supply"
If you look only at power generation, the U.S. is not short of electricity. What really gets stuck is the chain that delivers electricity to the cabinet. By the first quarter of 2026, the delivery cycle for U.S. generator step-up transformers has exceeded 160 weeks, and the wait time for high-voltage circuit breakers has extended from 77 weeks in 2023 to 125 weeks in the second half of 2025; The average waiting time for gas turbines is nearly 5 years, and prices have risen from about $800 per kilowatt to over $2,500—more than tripling. Globally, the delivery cycle for large high-power transformers is generally over 2.5 years, with some models reaching up to 5 years.
Meanwhile, about 2,300 gigawatts of power generation and storage capacity in the U.S. are queued for grid connection, exceeding the total installed capacity of the entire U.S. grid; PJM Grid alone has 31 gigawatts of data center demand awaiting approval. In August, the governor of Texas signed an executive order suspending all new data center grid connection approvals, with the total queued capacity exceeding 474 gigawatts—five times the state's peak load; New York State became the first state in the U.S. in July to implement a statewide data center construction ban. Meanwhile, nearly 70% of Americans oppose data centers, with only 14% willing to accept building data centers near their communities.
The result is that money has already been spent, but projects remain unmoved. In the first quarter of 2026, the total value of data center projects locked or delayed across the U.S. reached $130 billion; JPMorgan Chase's satellite imagery analysis shows that over 60% of data center projects scheduled for completion in 2027 have not yet started, and another 7% are delayed. This is the difference between "compute power manufactured" and "power that is lit up": the former can appear in financial reports within hours, while the latter must first pass through transformers, grid connection permits, and community hearings.
Why Is China Singled Out As An "Exception"?
When discussing power constraints, China is often singled out. In July, data from the National Energy Administration showed that national photovoltaic installations reached 1.286 billion kilowatts, officially surpassing coal power with 1.285 billion kilowatts, becoming the country's largest power source; China's advantage is not only in power generation but also in its ability to shift computing load from the east to green power-rich areas in the west—relying on unified national UHV cross-regional dispatch capability, which the United States does not have. Data from the National Energy Administration shows that domestic data center electricity consumption will increase from 170 billion kWh in 2025 to 800 billion kWh in 2030; During the 15th Five-Year Plan period, investment in the power sector is expected to exceed 20 trillion yuan, with West-to-East power transmission scale exceeding 420 million kW.
But "having an advantage now" does not mean "always having an advantage." Hardware constraints know no borders:Domestic intelligent computing capacity has reached 780,000 PFLOPS, and transformer delivery cycles have also extended to 12 to 18 months. The expansion of core power equipment capacity cannot keep pace with the pace of new computing cluster construction. A more hidden risk lies in power quality: Traditional grids granularize outage times in minutes or hours, but AI intelligent computing GPUs have almost zero tolerance for voltage drops of 10 milliseconds to 2 seconds. By 2025, there have already been cases of production interruptions caused by voltage drops in computing power facilities in Shanghai, Hainan, and other places, which have been specifically called for rectification by the National Energy Administration. A 99.99% power supply reliability rate cannot mask millisecond-level instability.
What To Watch Next
First, the delivery cycle of power equipment. Transformers falling from 160 weeks and gas turbine schedules shortening are the real signals of easing the power shortage; Before these numbers shift, any capital expenditure plan is just a theoretical sum.
Second, grid connection approvals and community resistance. Texas's approval suspension, New York state's ban, and residents' opposition rate close to 70% indicate that restrictions have shifted from "whether power can be generated" to "whether connection is allowed." The speed of resolution depends on politics, not engineering.
Third, the difference between actual power supply capacity and nominal computing power in 2027. The market is currently pricing in high growth in computing power demand, but if data centers delay grid connection, depreciation will continue and revenue will not be generated. AI asset pricing layers will start with these projects.
