Opinion
What if the AI runs out of electricity?
"The hard truth is that every agent runs on a power plant," writes Square One Labs co-founder Amir Fishelov. "The industry has simply been acting as if it doesn't."
Think of a mid-sized software company that let most of its support team go last year. Automated agents now answer most tickets much faster and at a fraction of the cost. Across the hall, R&D no longer hires junior developers. The development team outputs twice as much code with fewer bugs and ships updates at unprecedented speed. Management is excited about AI adoption, the board is pleased, and gross margins look healthier than ever. There is only one problem: nobody accounted for what this newly added AI cost line really depends on. What happens when the price of the compute powering AI agents doubles in a single quarter, or when the provider quietly shifts capacity to larger enterprise accounts?
That vulnerability is not unique to those companies massively adopting AI capabilities. Next in line is the AI vendor itself, whose valuation is dependent on a steep adoption curve where every forecast assumes adequate inference capacity will simply be there. Finally, behind both stands the investor underwriting multiples on that revenue growth. Three distinct balance sheets, all anchored to a single unwritten premise: compute will remain abundant and cheap. Yet competence you rent can always be repriced or rationed while competence you own cannot. That is the fundamental distinction the market has stopped making.
This is not hypothetical. U.S. data centers consumed roughly 76 TWh of electricity in 2018 and 176 TWh in 2023, about 4.4% of the nation’s total, according to Berkeley Lab’s report for the U.S. Department of Energy. Scenarios in that same report project demand could be reaching 325 to 580 TWh by 2028, or up to 12% of total U.S. generation. Compute providers are already feeling the pinch. On its latest earnings call, Microsoft acknowledged that customer demand continues to outpace available capacity. Amazon has stated it will not have enough capacity to satisfy projected demand in 2026.
The electrical grid tells the very same story. PJM Interconnection, which coordinates the wholesale electricity market across 13 eastern states hosting the highest concentration of data centers, saw its 2028/29 capacity auction clear at the price cap. Even then, it fell roughly 6.8 gigawatts short of its own reliability requirements as massive new loads joined the queue. In the real world, physical hardware moves slowly: the Department of Energy points out that large power transformers take up to five years to deliver. Software scales in months; transformers, high-voltage steel, copper, and regulatory permits scale in years.
The immediate consequence will not be a blackout, it will be a price hike. Scarcity expresses itself in the cost per token, per query, or per agent-hour. Like in every commodity market, compute will simply be rationed to the highest bidder, and the first casualties will be the gross margins of products priced on the belief that machine intelligence only gets cheaper. A growth model premised on falling costs falls apart when the marginal cost of intelligence is dictated by the price of a megawatt. How many board decks today actually stress-test for that sensitivity? I believe that very few. Most still treat power the way early internet companies treated bandwidth: as a commodity someone else will handle.
Someone most likely will, but the solutions are physical, and there are only two. The first is building new generation faster and closer to where compute lives. The second is designing data centers that deliver the same compute on drastically less power. The latter path sounds reassuring, but it is far less dependable: historically, cheaper and more efficient computing leads to higher aggregate consumption, not lower. Efficiency gains may buy time but rarely provide structural relief. Until either solution scales, the prudent baseline is that the power crunch will outlast the forecasts. Whoever resolves that bottleneck, whether by erecting new generation infrastructure or developing step-change data center efficiencies, will capture a share of the value created by everything running on top.
The hard truth is that every agent runs on a power plant. The industry has simply been acting as if it doesn't.
Amir Fishelov is Co-founder and Managing Partner at Square One Labs (S1L).
