US Grid Headroom Turns Negative by 2027, Forcing AI Datacenters to Behind-the-Meter Power

SemiAnalysis published a bottom-up quantitative forecast projecting that US grid headroom turns negative by 2027 as AI datacenter demand grows from ~21 GW of new capacity added in 2026 to 84 GW added annually by 2030 [1][2]. The core mechanism is a firm-capacity gap: solar and storage each add over 20 GW of nameplate generation per year, but reliable peak capacity has barely increased because intermittent sources cannot be counted on at demand peaks [3]. The structural response is behind-the-meter (BTM) generation — datacenters building their own on-site power to bypass the grid — which SemiAnalysis forecasts will supply over half of new US datacenters by 2028, with a BTM equipment market exceeding 50 GW/year by 2029 [1]. Separately, Google reported a 37% electricity increase in 2025 and acknowledged its AI buildout is outrunning grid decarbonization [8].

If the grid-headroom projection holds, utility interconnection cannot support the planned AI infrastructure buildout, and the BTM shift will redirect hundreds of billions in capital from utility-scale renewables toward on-site firm generation — primarily natural gas in the near term. This reshapes grid planning, utility economics, and corporate carbon accounting simultaneously.

SemiAnalysis has published a detailed bottom-up forecast arguing that the US power grid will exhaust usable headroom for new datacenters by 2027. AI datacenter gross power demand is projected to grow from roughly 21 GW of new capacity added in 2026 to 84 GW per year by 2030. Modeling required reserve margins across US regions, SemiAnalysis concludes that available headroom is already approaching zero and crosses into negative territory by 2027 [1][2].

The critical insight is that the problem is not simply demand growth but a mismatch between the kind of generation being added and the kind datacenters need. Solar and battery storage are each adding over 20 GW of nameplate capacity per year to the US grid, but firm capacity — what grid operators can actually commit to delivering at peak demand — has barely moved, because intermittent sources cannot guarantee output when it is needed most [3]. This gap makes grid interconnection queues slow and uncertain even where physical generation exists. The situation is worsened by an asymmetric contractual structure: datacenter developers must post letters of credit, security deposits, or take-or-pay contracts to secure utility commitments, while utilities face no penalties if they deliver interconnection late [1].

The structural response SemiAnalysis identifies is behind-the-meter (BTM) power: on-site generation that bypasses the utility grid entirely. The firm forecasts BTM will power over half of all new US datacenters by 2028 and that the total addressable market for BTM datacenter equipment will exceed 50 GW per year by 2029 [1]. Natural gas is the near-term answer for firm BTM capacity, a path SemiAnalysis explored in earlier work on onsite gas deployments at AI labs [4]. Industry analysts and datacenter operators have separately been treating BTM as a planning requirement rather than an option, evaluating solar, storage, and gas configurations across colocation and hyperscaler builds [5][6][7].

Google's 2025 environmental report, published in July 2026, provided independent demand-side confirmation: Google's electricity consumption rose 37% in 2025, the largest single-year increase in the company's history, with total growth since 2019 exceeding 250% [8]. Google maintained operational carbon neutrality through clean energy purchases but acknowledged explicitly that its AI infrastructure buildout is 'currently accelerating faster than the grid is decarbonizing' [8]. The admission reflects a tension present across the hyperscaler sector: AI expansion is simultaneously straining grid capacity and corporate carbon commitments, and purchasing clean energy certificates does not resolve the physical supply constraint.