SemiAnalysis forecasts 40GW+ of behind-the-meter data center power generation by 2028 as hyperscalers bypass grid constraints with on-site generation.
The US grid is currently serving most datacenter load, but we're reaching a critical juncture. As AI labs and hyperscalers drive insatiable demand for power, the grid simply cannot add capacity fast enough. Behind-the-meter solutions have emerged as the only viable path for the largest players to secure the power they need. Nearly a year ago, SemiAnalysis predicted the rapid rise of new entrants in the BTM gas equipment market. Since then, companies like Bloom Energy, Bergen Engines, Wärtsilä and others have seen remarkable success, overcoming capacity constraints from traditional suppliers like GEV and Siemens that many feared would persist.
US datacenter buildout is accelerating dramatically, growing from 21GW in 2026 to 84GW by 2030. Our research indicates that BTM will power well over half of new US datacenters by 2028 and beyond, with the total addressable market for datacenter BTM equipment projected to cross 50GW per year by 2029. The underlying challenge is straightforward: new grid capacity isn't growing fast enough to serve both datacenter load growth and non-datacenter demand.
Our forecast rests on three core building blocks. First, datacenter demand derives from a bottom-up, building-by-building model, supported by chip-by-chip AI demand forecasting via our Accelerator Model and validated through our Tokenomics Model. Second, grid headroom is analyzed through supply and demand dynamics across each major US grid region, following ISO and RTO methodologies including UCAP/ICAP reserves, supply and demand growth, and reliability risks. Third, new grid supply is forecasted bottom-up across 40,000 generation assets, tracking quarterly commercial operation dates and estimating true capacity value via our proprietary ELCC model tailored to each ISO and major non-ISO region.
Our forecast shows barely 15GW of net-new ELCC capacity being added annually, with a rising trend toward 20GW+ by decade's end—effectively all the firm capacity the grid can recognize to serve firm datacenter load and other firm loads like industrial plants and semiconductor fabs. When this accredited supply is netted against peak demand and required reserve margins, available headroom is already approaching zero and turns negative by 2027 across the country. This makes power generation a major bottleneck to grid-connected datacenter load growth, with transmission presenting another critical constraint that will be addressed in forthcoming analysis.
Datacenter operators already see this firsthand. Utilities initially commit to meeting 2027 load schedules (such as 500MW) only to later declare they can deliver only in 2029 due to long-lead equipment requirements for main power transformers, high-voltage breakers, and network upgrades. The burden increasingly falls on developers: securing grid-connected power now often requires substantial letters of credit, security deposits, or take-or-pay commitments to fund generation serving their load. Switch Datacenter, for instance, closed a multi-billion-dollar performance letter-of-credit facility in 2026 to back such obligations. Despite billion-dollar commitments, utilities often face no penalties for missing delivery timelines.
These generation and transmission constraints, combined with inadequate market incentives, make behind-the-meter often the most attractive solution for GW-scale new developments. Many top-tier developers are planning 5GW+ behind-the-meter facilities in Texas, where permitting onsite gas is simpler. These generation and transmission constraints reshape the winners and losers: beneficiaries are not always the usual suspects. While temporary peak gas turbine orders represent a consideration, the longer-term market positioning depends on manufacturing capacity, equipment and installation lead times, and major deal announcements.
Our analysis unfolds in three steps. First, we establish the supply constraint using our new Energy Model to explain power-plant addition forecasts, true capacity value estimation via ELCC, and resulting grid headroom. Second, we examine buyer decision-making: the relative attractiveness of BTM versus grid for the largest power and compute buyers, broken down by use case. Finally, we explore the practical bridge emerging in ERCOT through the Batch Zero process—hybrid co-location structures that blend on-site generation with continued grid access, where early winners are taking shape.
Analysis of 40,000 generators reveals a core challenge: the lack of firm, dispatchable capacity being added over the next two years. Our forecast shows the US industry will add less than 10GW of gas per year in 2026 and 2027, with additions accelerating only in 2028 and beyond. This forecast replicates our Datacenter Model methodology, analyzing construction timelines against hundreds of thousands of empirical datapoints backed by real-time satellite imagery, revealing insufficient power plants in advanced construction to enable 2026–27 delivery.
The 2026–27 shortfall stems from multiple bottlenecks. Institutional and queue friction from utilities and overcrowded generation interconnection queues present the first challenge. In PJM, the queue itself is no longer the binding constraint—conversion is. Though roughly 57GW has cleared studies with executed or offered interconnection agreements, since 2020 about 24GW of projects with fully executed agreements (including 13.5GW of gas) terminated before reaching commercial operation, largely due to permitting denials, supply-chain delays, and financing failures. Permitting alone accounted for 29% of project milestone changes between January 2023 and January 2026, versus 23% for supply-chain delays. Second, the technology mix works against speed: the bulk of utility-ordered gas capacity consists of combined-cycle gas turbines and combustion turbines.