The research firm argues that grid infrastructure, which takes years to build, now limits how fast AI compute can scale
The AI industry spent years worrying about who could get enough chips. A new Delphi Digital report argues that was the wrong question.
The report, titled “The AI Boom Has a Power Problem” and published on October 29, 2026, says electricity supply and grid infrastructure are now the main limits on AI growth. Semiconductors, it argues, have become the easier part of the problem.
The reasoning is about timing. Chip production can ramp quarter by quarter. Power plants, transmission lines and grid connections take years. When one input moves at sprint speed and the other moves at the pace of a zoning board, the slower one sets the speed limit.
A dynamo, a historian and a grid under strain
Delphi opens with an unusual choice for an investment report: a scene from 1900. The historian Henry Adams walked into the Gallery of Machines at the Paris Exposition and stood in front of a forty-foot dynamo.
The report frames Adams as someone who saw early that electricity would become something society organized itself around. It pairs that story with an epigraph from energy scholar Vaclav Smil. Delphi casts its own subject as a race for electricity and an examination of which companies are positioned to profit from it.
The literary framing lands against some very unliterary data. Texas offers the starkest example.
ERCOT, which runs most of the Texas grid, saw its large-load queue climb from 63 GW at the end of 2024 to 474 GW by June 2026. Roughly 90% of that queue is attributed to data centers.
The mid-Atlantic tells a similar story. PJM, the largest grid operator in the eastern US, has had its capacity auctions clear at the price cap for three consecutive rounds.
Those auctions came in 6.6 to 6.8 GW short of reliability targets. They were the first shortfalls in the auction’s history.
Capacity auctions are how PJM pays power producers to promise they will be available years ahead. Clearing at the cap means buyers paid the maximum allowed and still could not secure enough supply.
The numbers behind the demand wave
Delphi is not alone in sounding the alarm. Goldman Sachs has said data center power demand could surge by 170% between 2025 and 2030.
Goldman also expects about $7.6 trillion in capital expenditures through 2031 to meet that demand.
Goldman projects global hyperscaler debt issuance will jump by $420 billion in 2027, a rise of more than 60% from 2026 levels.
Grid regulators are reading the same tea leaves. NERC, the body responsible for North American grid reliability, anticipates a 224 GW increase in summer peak demand over the next decade.
NERC also warns that 13 of its 23 assessment regions risk shortfalls by 2030. The Department of Energy has issued forecasts pointing to a similar mismatch between demand and generation capacity.
Why the bottleneck shifted
For much of the recent AI boom, the conversation centered on chip supply. Access to advanced processors determined who could train models and who had to wait.
Delphi’s argument is that chip supply has a natural advantage: factories can scale output on a relatively short cycle. Energy infrastructure does not work that way. New generation, transmission upgrades and interconnection approvals involve permitting, construction and regulatory review that stretch across years.
What this means for investors and the AI buildout
The most direct implication is a shift in where scarcity, and therefore pricing power, sits in the AI supply chain. If electricity is the binding constraint, then whoever controls reliable power and grid access holds leverage over the companies that need it.
That points attention toward power generation, grid equipment and interconnection services. Delphi explicitly frames its report around the companies positioned to profit from the race for electricity. The PJM auction results suggest the price signal is already flashing: buyers are paying the maximum and still coming up short.
The debt projections add another layer. A $420 billion jump in hyperscaler borrowing in a single year means the AI buildout is increasingly financed with leverage.
The key metrics to watch are straightforward. Future PJM auction results will show whether shortfalls persist or ease. ERCOT’s large-load queue will indicate whether data center demand keeps compounding or starts to plateau. And hyperscaler debt issuance through 2027 will reveal how aggressively the biggest players are betting that the power will eventually show up.
Henry Adams stared at a single dynamo and sensed the world was about to reorganize around electricity. More than a century later, Delphi is making a similar bet: that the next phase of the AI story will be decided less by the cleverness of the models and more by the stubborn physics of the grid.
