The influx of AI and its growing compute demands are revealing major constraints in our power system and causing dislocations in power markets[1]. The power bottleneck was, until recently, expected to be felt towards the end of the decade[2]. However, it appears to be arriving faster than most expected[3]. Satya Nadella, CEO of Microsoft, recently remarked that the company has “chips sitting in inventory”[4] because of limited power availability. It is a striking signal: the bottleneck in AI build-out is no longer GPUs, it is electrons.
For the past four decades, U.S. electricity demand has grown at roughly 1% a year[5], a slow and predictable pace that the system has been designed to accommodate. But data center expansion, driven by AI workloads, is now on track to consume nearly all that growth on its own. Estimates put 2024 data center energy use at 4.4% of all US electricity, a 2x+ increase in the last five years[6]. While the go-forward estimates range widely, the next three years are expected to see another 1.5-3x growth in energy use by AI data centers.[7]. Layer on accelerating electrification and industrial reshoring, and suddenly the grid faces an entirely new challenge: produce and sustain 5%+ annual growth for the first time in decades. That is an enormous, resource-straining leap for an industry optimized for stability, not rapid scale.
If you look closely at the data, the picture gets even tougher. Net capacity growth from traditional firm resources (gas, nuclear, and coal) is negative over the next year[8]. Aging plants are being retired faster than replacements come online[9], and new gas development faces mounting constraints from supply chain bottlenecks, permitting, and pipeline opposition[10]. Most of the headline growth is coming from renewables and storage, which together account for more than 100% of net additions. In many ways that is encouraging; solar paired with storage is increasingly capable of acting as a firm resource. But not all of it meets the uptime and stability standards that AI and industrial loads require.
Geographic mismatches between renewables generation, in the sunny or windy parts of the country, and data centers, clustering near fiber and water, compounds the problem. Connecting compute with power through the buildout of transmission and distribution lines is a notorious challenge[11] fraught with NIMBYism and politics[12]. Therefore, pockets of constraint are already forming in regions with heavy AI development.
Investors are now fully aware of the reality that energy is inextricably linked to artificial intelligence. Public markets have gone gangbusters across the core components of the electricity system (utilities, IPPs, developers, grid and transmission infrastructure, and equipment providers). Companies across the electricity value chain have outperformed the market significantly[13].
Private markets investors have also rushed into energy and power opportunities. Electrons are a commodity, but right now they are being differentiated by speed to availability and firmness (the ability to deliver 24/7). That scarcity premium will not last forever. Durable price advantages and quality (uptime, maintenance, etc.) will be required to create sustainable businesses. Still, we expect this supply constraint to persist for years. Even accounting for significant efficiency gains, the forward compute requirements for AI are staggering[14].
Opportunities exist across the power ecosystem. However, it has been surprising to watch as capital chases every developer with access to a site that happens to have grid, fiber, and water in the hope that an AI datacenter will come knocking. These are projects, not businesses. The focus should remain on key bottlenecks to unlocking capacity and system performance, where technology is used to build repeatable businesses with a durable moats.
[1] Bloomberg (2025) – AI Data Centers Are Sending Power Bills Soaring.
[2] Utility Dive (2025) – PJM capacity-price cap settlement shows faster-than-expected power-demand surge.
[3] RCR Wireless (2025) – EIA: U.S. power use to hit record highs in 2025-26 due to data-center demand.
[4] Data Center Dynamics (2025) – Microsoft delaying AI GPU installs for lack of grid capacity.
[5] Our World in Data (2025) – U.S. energy generation.
[6] LBNL (Dec 2024) – United States Data Center Energy Usage Report (2024).
[7] S&P Global Commodity Insights (Oct 2024) – Data-center grid power demand +22% in 2025, nearly triple by 2030.
[8] EIA Electric Power Monthly (Table 6.01) – U.S. net generation by energy source.
[9] IEEFA (2025) – U.S. on track to close half of coal capacity by 2026.
[10] S&P Global Commodity Insights (May 2025) – Gas-turbine lead times up to 7 years, costs sharply higher.
[11] ACEG (July 2024) – Fewer New Miles: Transmission Investment Trends Report.
[12] U.S. Department of Energy (July 2025) – DOE terminates financial assistance for Grain Belt Express.
[13] NASDAQ (2025) – ZAP ETF: Advanced charting and YTD performance.
[14] Goldman Sachs (2025) – AI to drive 165% increase in data-center power demand by 2030.
h/t Roxanne Tully Baine