SemiAnalysis: U.S. Grid Capacity May Turn Negative by 2027, Forcing AI Data Centers into a "Self-Supplied Power" Era

```

The AI computing power race is pushing the US power grid to the brink.

A new energy research report by SemiAnalysis points out that the "headroom" of the US power grid to support additional large loads will drop to negative as early as 2027. This means that it will become increasingly difficult for new AI data centers to rely on the public grid for stable power supply in the future.

As grid expansion lags far behind the growth in AI computing power demand, more and more data centers will have to build their own power generation systems (Behind-the-Meter, BTM), indicating that US AI infrastructure may enter the era of "self-supplied power."

The report estimates that after 2028, more than half of newly built US data centers will adopt the BTM model for power supply; by 2029, the market scale for BTM equipment in data centers alone is expected to exceed 50GW per year, becoming one of the fastest-growing new fields in the entire AI infrastructure investment chain.

AI demand is surging, but grid expansion can't keep up

SemiAnalysis believes that the current pace of US data center construction has far exceeded the speed at which the grid can supply new capacity.

The report estimates that new power demand from US data centers will rapidly increase from 21GW in 2026 to 84GW in 2030, while the US grid can only add about 15GW of capacity per year with reliable power supply capability (ELCC) during the same period, and may not gradually increase to above 20GW until the end of this century.

More importantly, this additional capacity is not solely for data centers; it must also meet new loads from manufacturing, semiconductor plants, and residents.

In other words, in the next few years, new US power supply will find it increasingly difficult to keep up with the explosive growth in demand from AI data centers.

SemiAnalysis's model shows that after deducting peak load and reserve requirements, the US power grid's remaining capacity for additional large loads is nearly depleted and will officially turn negative around 2027. This means that continuing to rely on the traditional grid to build large AI campuses will increasingly be limited by power bottlenecks.

The issue is not just generation, but the pace of grid construction

The report points out that the market consistently underestimates the complexity of US grid construction.

Currently, the biggest constraints are not merely generation capacity, but run through the entire supply chain.

On one hand, the construction cycle for natural gas power plants typically lasts 4 to 6 years, and in the next two years, new US gas power projects will be very limited. After tracking about 40,000 power assets, SemiAnalysis estimates that from 2026 to 2027, annual new US natural gas installations will be less than 10GW, with improvements only evident after 2028.

On the other hand, delivery times for key equipment such as high-voltage transformers, gas turbines, and circuit breakers have generally extended to 3-4 years, far above historical averages. Additionally, problems like project approval, grid connection queues, financing, and community permits further slow the construction pace.

Many data center developers have encountered similar situations: power companies initially promised to provide hundreds of megawatts of load by 2027, but later notified delays until 2029 or even later, with power companies often not liable for such delays.

For AI companies who trade computing power for revenue, this kind of uncertainty is virtually unacceptable.

Renewables are insufficient to fill the AI power gap

SemiAnalysis emphasizes that although US solar and storage installations will continue to grow rapidly in the next few years, these capacities do not equate to genuinely reliable power supply for large data centers to operate continuously.

Using the ELCC (Effective Load Carrying Capacity) metric commonly used in the power industry, the report finds that due to the obvious intermittency and highly synchronized generation times of solar and wind power, their additional installed capacity contributes far less to reliable system power supply than nominal capacity.

As the proportion of renewables increases, their marginal contribution will continue to decrease.

Although storage systems can alleviate short-term load fluctuations, they also suffer from diminishing marginal benefits. Once large amounts of 4-hour storage are in operation, system risks will gradually shift to those with longer-duration power gaps, and relying on storage alone will be insufficient to meet AI data center 24/7 operation needs.

Therefore, in the coming years, dispatchable sources like natural gas will remain the backbone supporting AI infrastructure expansion.

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

Against the backdrop of the public power grid struggling to meet demand, Behind-the-Meter is quickly becoming the new option for large AI data centers.

BTM refers to data centers directly building or incorporating exclusive power generation facilities within the campus itself, rather than fully relying on the public grid for power supply.

SemiAnalysis believes that compared to the long and uncertain wait for grid access, BTM's greatest advantages are 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 notes that in the overall total cost of ownership (TCO) for AI cloud business, the proportion of power costs is not high, but securing stable power supply could correspond to billions or even tens of billions of dollars in revenue. Therefore, enterprises are more willing to bear the cost of building self-supplied power and are unwilling to wait years for grid connection.

Meanwhile, some AI data centers are also lowering the requirement for "five nines" (99.999%) power reliability as traditionally demanded by data centers, in exchange for faster rollout. For example, some hyperscale AI data centers are accepting less redundancy in power supply, further improving the economics of BTM solutions.

Competition for AI infrastructure is shifting from GPUs to energy

SemiAnalysis believes that in the next few years, the critical constraint for US AI industry competition will no longer be GPU supply, but the ability to secure power resources.

The report predicts that as grid capacity continues to tighten, more and more AI data centers will adopt a "self-generation + public grid" hybrid model, and US power infrastructure will be restructured as a result. Investments in gas power generation equipment, fuel cells, on-site generation systems, and related power equipment are expected to become important beneficiaries of the next phase of AI capital expenditures.

For the entire AI industry, this means that the focus of competition is gradually moving from chips and servers to energy infrastructure. Whoever first secures stable, reliable, and scalable power resources will have a better chance of gaining the upper hand in the next round of AI computing power competition.

Risk Disclaimer and Liability StatementThe market involves risk, and investments should be made with caution. This article does not constitute personal investment advice and does not take into account any user's particular investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions herein are suitable for their specific circumstances. Invest accordingly, at your own risk. ```