SemiAnalysis: US grid capacity may turn negative in 2027, AI data centers forced into "self-supplied power" era

The AI computing race is pushing the U.S. power grid to a breaking point.

A new energy research report from SemiAnalysis points out that the "Headroom" available on the U.S. power grid for new large-scale loads could drop to negative values as early as 2027, meaning that future new AI data centers will increasingly find it difficult to rely on the public grid for stable power supply.

As grid expansion lags far behind the growth in AI computing demand, more and more data centers will have to build their own power generation systems (Behind-the-Meter, BTM), potentially ushering the U.S. AI infrastructure into an era of "self-supplied power."

The report estimates that after 2028, more than half of new data centers in the U.S. will adopt BTM power supply; by 2029, the market size for data center BTM equipment alone is expected to exceed 50GW annually, becoming one of the fastest-growing new areas in the entire AI infrastructure investment chain.

AI demand surges, but grid expansion cannot keep up

SemiAnalysis believes that the current pace of data center construction in the U.S. has far exceeded the speed at which the grid can provide new capacity.

The report estimates that the new power demand from U.S. data centers will rapidly grow from 21GW in 2026 to 84GW by 2030. However, during the same period, the U.S. grid can only add approximately 15GW of new capacity annually that truly provides reliable power (ELCC), potentially increasing to over 20GW only by the end of this decade.

More importantly, this new capacity is not solely for data centers; it must also meet other new loads such as manufacturing, semiconductor factories, and residential needs.

In other words, in the coming years, new power supply in the U.S. will increasingly struggle to cover the explosive growth in demand from AI data centers.

SemiAnalysis's model shows that after accounting for peak load and reserve capacity requirements, the remaining headroom on the U.S. grid to accommodate new large-scale loads is nearly exhausted and is expected to turn negative around 2027. This means that continuing to rely on the traditional grid to build large AI campuses will become increasingly constrained by power bottlenecks.

The problem is not just power generation, but the entire grid construction speed

The report notes that the market generally underestimates the complexity of U.S. grid construction.

The biggest constraint currently comes not just from generation capacity but from the entire supply chain.

On one hand, the construction cycle for natural gas power plants typically takes 4 to 6 years, and new natural gas power projects in the U.S. over the next two years are very limited. After tracking approximately 40k power generation assets, SemiAnalysis estimates that from 2026 to 2027, the U.S. will add less than 10GW of new natural gas capacity annually, with significant improvement only after 2028.

On the other hand, lead times for key equipment such as high-voltage transformers, gas turbines, and circuit breakers have generally extended to 3 to 4 years, far above historical averages. At the same time, issues like project approvals, grid connection queues, financing, and community permits further slow down construction.

Many data center developers have already encountered similar situations: utilities initially promised to provide hundreds of megawatts of load by 2027, but later notified that it could only be delayed to 2029 or even later, and utilities often bear no responsibility for the delay.

For AI companies that trade computing power for revenue, such uncertainty is almost unacceptable.

Renewable energy struggles to fill the AI load gap

SemiAnalysis emphasizes that although solar and energy storage installations in the U.S. will continue to grow rapidly in the coming years, this installed capacity does not equate to the actual power supply that can reliably support the continuous operation of large data centers.

Using the ELCC (Effective Load Carrying Capability) metric commonly used in the power industry, the report finds that due to the intermittent nature of solar and wind power and their highly coincident generation times, their contribution to system reliability is far lower than their nominal installed capacity.

As the share of renewables increases, their marginal contribution will continue to decline.

While energy storage systems can alleviate short-term load fluctuations, they also suffer from diminishing marginal returns. When a large amount of 4-hour storage is deployed, system risks gradually shift to longer-duration supply gaps, and relying solely on storage becomes insufficient to meet the around-the-clock operational needs of AI data centers.

Therefore, in the coming years, dispatchable power sources such as natural gas will remain the core support for AI infrastructure expansion.

"Self-supplied power" becomes the fastest and most certain solution

Against the backdrop of an increasingly strained public grid, Behind-the-Meter is rapidly becoming a new choice for large AI data centers.

BTM refers to data centers directly building or equipping dedicated power generation facilities to supply electricity on-site, rather than relying entirely on the public grid.

SemiAnalysis believes that compared to waiting for a lengthy and uncertain grid connection, the biggest advantage of BTM is speed and certainty.

For AI labs like OpenAI and Anthropic, computing power directly determines model training and inference capabilities, as well as future revenue growth. The report points out that in the total cost of ownership (TCO) of AI cloud services, the proportion of electricity costs is not high, but obtaining stable power supply could correspond to billions of dollars or more in revenue. Therefore, companies are more willing to bear the cost of building self-supplied power rather than wait years for grid interconnection.

At the same time, some AI data centers are beginning to lower the traditional requirement for "five nines" (99.999%) power reliability in exchange for faster deployment. For example, some hyperscale AI data centers are accepting lower levels of power redundancy, further improving the economics of BTM solutions.

AI infrastructure competition shifts from GPUs to energy

SemiAnalysis believes that in the coming years, the key constraint on competition in the U.S. AI industry will no longer be just GPU supply, but the ability to secure power resources.

The report predicts that as grid capacity continues to tighten, more AI data centers will adopt a hybrid model of "self-supplied generation + public grid," and U.S. power infrastructure will be restructured accordingly. Investment opportunities around gas-fired generation equipment, fuel cells, on-site power generation systems, and related power equipment are expected to become important beneficiaries of the next phase of AI capital expenditure.

For the entire AI industry, this means that the focus of competition is gradually extending from chips and servers to energy infrastructure. In the future, whoever can secure stable, reliable, and scalable power resources first will be more likely to gain an advantage in the next round of AI computing competition.

Risk Warning and Disclaimer

        Market risk exists, investment requires caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment at your own risk.