AI Power Deals Pile Up as US Generation Gap Tops 100 GW
Signed capacity contracts are outrunning what utilities can actually deliver, putting gas and battery buildout economics under pressure.
Bank of America analysts have put a number on the gap between what AI power demand requires and what regulated utilities can supply: more than 100 GW over the next five years. The firm forecast the United States will need more than 230 GW of new generating capacity through 2030, while utilities are expected to add only about 93 GW of accredited supply.7 That divergence is now the central pricing question for gas-fired generation, grid batteries and the equipment manufacturers rushing to fill the space between signed contracts and delivered megawatts.
The demand case behind that gap is large. Goldman Sachs Research projects global data center power use will jump about 50% by 2027 and could surge up to 165% by decade's end compared to 2023 levels.5 The Electric Power Research Institute sees data centers consuming between 9% and 17% of U.S. electricity supply by 2030, or up to 790 terawatt-hours, against roughly 4% as of mid-2026.2 Those are forecasts built on AI adoption curves that have not been stress-tested against a demand slowdown.
Babcock & Wilcox has already moved on the promise. Shares closed at $14.54, up 129.34% year to date, after the company announced a $2.4 billion design-build contract with Base Electron for 1.2 GW of natural gas-fired power, driving backlog up 470% to $2.8 billion.1 Management guided 2026 core adjusted EBITDA to $70 million to $85 million, roughly 80% year-on-year growth, excluding any data center upside, and the global pipeline exceeds $12 billion.1 Base Electron is evaluating an additional 1.2 GW option. The stock is pricing in execution.
Battery storage is chasing the same signal. The U.S. added a record 57.6 GWh of new battery energy storage capacity in 2025, according to the Solar Energy Industries Association, bringing total deployed capacity to 166.1 GWh.2 The group projects annual deployments will reach 110 GWh by 2030, with a significant share driven by data center demand, and Fluence Energy is engaged in over 30 GWh of data center-related projects globally, with a meaningful U.S. portion, CEO Julian Nebreda said.2
But interconnection queues are a harder constraint than capital. Battery storage firms see surging interest from data centers yet face lengthy grid connection waits and a supply chain heavily dependent on China.2 The queue problem is not new; the scale now threatens to turn signed capacity contracts into delay claims. Projects with the strongest sponsors, best locations and clearest utility partnerships will move forward. Others face delays, cost overruns or longer interconnection timelines.4
BofA expects more data center developers to shift toward behind-the-meter generation as utilities struggle to bring capacity online fast enough.7 Manufacturers including Caterpillar, INNIO, Rolls-Royce and Wärtsilä have expanded production lines to meet that demand.7 Behind-the-meter power is rational for a hyperscaler with a hard deadline, but it removes load from grid oversight, making aggregate demand harder to forecast and harder to unwind if AI economics soften.
The equity market appears skeptical about how much contracted revenue translates into value. Total contracted revenue across some of these deals runs approximately $2.6 billion, yet Bitzero trades at a market cap of roughly $130 million.5 Separately, a developer recently proposed a $12 billion data center complex in St. Louis, a scale that assumes demand holds through the decade.5 The Economist has drawn parallels between the data-center investment boom and the 1990s telecoms bubble, noting McKinsey's forecast of $5.2 trillion in worldwide AI-related capital spending.3
The bull case is visible: capital is available, customers are willing to sign long-term capacity agreements, and the demand growth trajectory from both Goldman and EPRI is steep enough that even a partial miss still requires substantial new generation.6 BofA adds a structural qualification: planned generation additions may overstate available supply because intermittent resources contribute less accredited capacity during peak demand than their nameplate ratings, which would widen the effective gap further.7
Base Electron's decision on that second 1.2 GW option will be read as a live confidence indicator for the AI power trade. If it exercises, the market takes it as evidence that near-term demand holds. If it stalls, financing costs across the sector move.1