The AI Bottleneck Series: Waiting for the Watts
Within the current AI ecosystem, electricity may be the most limiting constraint today.
Artificial intelligence (AI) has emerged as one of the most powerful secular growth themes of our time, transforming how businesses operate and how people live and work. Yet AI is not a single technology. It is a complex ecosystem supported by a long and interconnected value chain, spanning semiconductors, networking equipment, data centers, power infrastructure, and software. As AI adoption accelerates, strain can emerge at key points along this supply chain, creating bottlenecks where demand outpaces available resources. While these constraints can create uncertainty, we believe they can also reveal compelling investment opportunities. In this series, we explore the most significant AI bottlenecks, the forces driving them, and the investment opportunities they may create.
Within the current AI ecosystem, electricity may be the most limiting constraint today. A large AI data center campus can require a gigawatt or more of continuous power, roughly equivalent to the aggregate consumption of a mid-sized American city. While that need may seem massive, the challenge is less about generating that power than connecting to it. An AI data center can typically be built in 18 to 24 months, but connecting a facility of that size to the public grid can take five to seven years, due to required engineering studies, regulatory approvals and physical upgrades.1 In Texas alone, ERCOT, the state grid operator, is reviewing requests from large new customers, roughly 90% of them data centers, for about 474 gigawatts, which is more than five times the most power the state has ever used at once.2
Why Is This Happening?
The cause of today’s bottleneck lies in the two decades leading up to the introduction of ChatGPT in 2022. As Figure 1 shows, U.S. electricity use grew steadily from 1975 to 2005, then stopped. More efficient lighting, appliances and factories, plus the movement of heavy industry overseas, kept electricity consumption essentially flat for nearly twenty years.
With electricity demand flat, utilities and independent power producers commissioned new plants only sparingly, and the manufacturers of turbines, transformers and other grid equipment calibrated their production to replacing aging assets rather than enlarging the fleet. The launch of ChatGPT changed the demand picture abruptly. Data centers consumed roughly 4% of U.S. electricity in 2023, and we estimate that share could reach approximately 30% by 2030, depending on the pace of AI adoption. We project total U.S. electricity demand growth could exceed 6% a year through 2040, a pace not seen since the postwar era. We recognize this is a bold projection, well above most official forecasts and a sharp break from the recent past. In our view, however, it reflects the scale of data center capacity already planned or under construction, and that growth is why the power supply chain is straining: the industry spent twenty years sized for stagnation and is now being asked to grow rapidly. Consider the equipment that new capacity requires: a gas turbine ordered today may not arrive until 2031, and a large transformer takes up to four years for delivery.3
The Solution: Bring Your Own Generation
As we wrote back in April in
The AI Power Problem commentary, data center operators have responded by taking power into their own hands. Bring Your Own Generation (BYOG), the practice of installing gas turbines, engines or fuel cells on site to power the data center directly, has moved from a niche solution to a central strategy. Binding orders for on-site power generation equipment now total about 75 gigawatts, and OpenAI’s flagship Stargate campus in Texas alone will run on more than 500 gas engines.
4 A one-gigawatt campus can produce tens of billions of dollars in revenue each year. Multiply that by a five-to-seven-year wait for a grid connection, and the opportunity cost of delay can exceed $100 billion. Paying a premium for on-site power that can be running in under two years is a rational decision. BYOG also sidesteps a growing political problem. Concerns about higher household electricity bills, driven in part by the cost of grid upgrades to serve new data centers, have made data centers unpopular neighbors; in August, Texas paused approvals of new grid-connected data centers pending a statewide audit. A campus that brings its own power is not asking ratepayers to fund new transmission lines on its behalf.
Looking Ahead
For data center operators, we believe on-site power is a necessity rather than a choice, and the companies that supply it are booked years in advance, which gives them strong pricing power. In our view, select gas turbine and engine manufacturers, makers of the switchgear and transformers that connect generators to servers, and the specialized contractors that install them remain primary beneficiaries of this bottleneck. The grid will catch up eventually. Until it does, the companies that help data center operators bypass the wait for a grid connection will continue to command a premium.
