SemiAnalysis forecasts US grid capacity turns negative by 2027, forcing AI data centers into self-supplied power era.

A SemiAnalysis report warned that the spare capacity of the U.S. power grid could turn negative as early as 2027, far insufficient to meet the surging electricity demand from AI data centers. Due to slow grid expansion and shortages of dispatchable power sources such as natural gas, newly built large-scale AI campuses will find it increasingly difficult to rely on the public grid. The report forecasts that by 2029, the behind-the-meter (BTM) onsite power generation equipment market for data centers will exceed 50 GW annually. Securing reliable power supply is now overtaking GPUs as the new critical bottleneck in the U.S. AI computing race, accelerating the industry's shift into an era of self-supplied power.

The AI computing power race is pushing the U.S. power grid toward a breaking point. According to SemiAnalysis, the U.S. power grid's 'headroom'—the remaining capacity available to accommodate new large-scale loads—could turn negative as early as 2027. This implies that newly built AI data centers will increasingly struggle to rely on the public grid for stable power supply.

As grid expansion lags far behind the growth in AI computing demand, an increasing number of data centers will have to deploy their own generation systems (Behind-the-Meter, or BTM), potentially ushering in an era of 'self-supplied power' for U.S. AI infrastructure. The report forecasts that after 2028, more than half of newly constructed data centers in the U.S. will adopt BTM power solutions; by 2029, the annual market size for BTM equipment serving data centers alone could exceed 50 GW, making it one of the fastest-growing segments within the entire AI infrastructure investment chain.

SemiAnalysis contends that the current pace of data center construction in the U.S. has far outstripped the grid's ability to deliver additional capacity. The report projects that U.S. data centers' incremental electricity demand will surge from 21 GW in 2026 to 84 GW by 2030. However, over the same period, the U.S. grid will only be able to add approximately 15 GW annually of new capacity with verified reliability—measured as Effective Load Carrying Capability (ELCC)—with levels potentially rising above 20 GW per year only toward the end of the century.

Critically, this new capacity is not exclusively allocated to data centers; it must also serve other growing loads, including manufacturing facilities, semiconductor fabs, and residential consumers. Over the next few years, new power supply additions in the U.S. will find it increasingly difficult to keep pace with the explosive growth in AI data center demand.

SemiAnalysis's modeling indicates that after accounting for peak load requirements and reserve margins, the U.S. grid's remaining headroom for accommodating new large-scale loads is nearly exhausted and is projected to become negative around 2027. Continuing to rely on the conventional grid for large AI campus developments will face escalating constraints due to power supply bottlenecks.

The report notes that the market has generally underestimated the complexity of building the U.S. power grid. The current primary constraint stems not just from power generation capacity, but from bottlenecks across the entire supply chain.

Natural gas power plants typically require a construction cycle of four to six years, and the number of new U.S. natural gas-fired generation projects scheduled for the next two years is very limited. After tracking approximately 40,000 generation assets, SemiAnalysis forecasts that annual additions of natural gas capacity in the U.S. will remain below 10 GW in both 2026 and 2027, with significant improvement not expected until after 2028.

Lead times for critical equipment—such as high-voltage transformers, gas turbines, and circuit breakers—have commonly extended to three to four years, far exceeding historical averages. Additionally, project permitting, interconnection queue delays, financing challenges, and community approval issues are further slowing construction timelines.

Many data center developers have already encountered similar situations: utilities initially promised hundreds of megawatts of load availability by 2027, only to later notify them that delivery would be delayed until 2029 or even later—and often without assuming liability for such delays. For AI companies whose revenue depends directly on computing power, this level of uncertainty is nearly unacceptable.

SemiAnalysis specifically emphasizes that although solar and energy storage installations in the U.S. will continue to grow rapidly over the next few years, nameplate capacity does not equate to reliable power supply capable of supporting continuous operation of large-scale data centers.

Using Effective Load Carrying Capability (ELCC)—a metric widely adopted in the power industry—the report finds that due to the pronounced intermittency of solar and wind power and their highly correlated generation profiles, the contribution of newly added renewable capacity to system reliability is significantly lower than its nominal installed capacity. As the share of renewable energy continues to rise, its marginal contribution will keep declining.

While energy storage systems can mitigate short-term load fluctuations, they also face diminishing marginal returns. Once a large volume of 4-hour storage capacity is deployed, system risks will gradually shift toward longer-duration power supply gaps, making it increasingly difficult for storage alone to meet the round-the-clock operational demands of AI data centers. Therefore, dispatchable power sources such as natural gas will remain central to supporting the expansion of AI infrastructure over the next few years.

Against the backdrop of an increasingly strained public grid struggling to meet demand, Behind-the-Meter (BTM) solutions are rapidly becoming the preferred option for large-scale AI data centers. BTM refers to data centers directly building or co-locating dedicated power generation facilities to supply electricity internally within their own campuses, rather than relying entirely on the public grid.

According to SemiAnalysis, compared with waiting for lengthy and highly uncertain grid interconnection processes, BTM's greatest advantage lies in its speed and certainty. For AI labs such as OpenAI and Anthropic, computing power directly determines their capabilities in model training and inference—and thus future revenue growth. While electricity costs represent a relatively small portion of the total cost of ownership (TCO) for AI cloud operations, securing reliable power supply could translate into billions—or even tens of billions—of dollars in potential revenue. Consequently, companies are more willing to bear the upfront costs of building on-site generation than to endure multi-year delays in grid interconnection.