As artificial intelligence adoption accelerates across corporate software and development pipelines, gross margins look healthier than ever, yet companies remain acutely vulnerable to a single unwritten variable. According to a report by Berkeley Lab for the U.S. Department of Energy, U.S. data center electricity consumption surged from 76 terawatt-hours in 2018 to 176 terawatt-hours in 2023, exposing a hidden physical bottleneck in digital infrastructure.
The Bottom Line
- Rising Compute Costs: Software efficiency scales in months, but physical grid capacity and large power transformers require up to five years to deliver.
- Grid Constraints: Wholesale electricity markets like PJM Interconnection cleared capacity auctions at the price cap while falling short of regional reliability requirements.
- Margin Pressures: As compute becomes scarce, the marginal cost of machine intelligence will be dictated by megawatt pricing rather than falling software costs.
When Software Scaling Collides with Physical Power Limits
Consider a mid-sized software firm that replaced most of its support staff with automated agents last year, cutting ticket response times while driving margins up. Across the hall, R&D teams ship code at twice the speed with fewer bugs. Management and boards celebrate the immediate cost savings, but the entire operational model rests on an unexamined assumption. What happens when the compute powering those agents doubles in price within a single quarter, or when providers quietly shift capacity to larger enterprise accounts?
This vulnerability extends directly to AI vendors whose valuations rely on steep adoption curves, and to the investors underwriting those revenue multiples. Three distinct balance sheets share a single unwritten premise: that compute will remain perpetually abundant and cheap. Yet rented competence can always be repriced or rationed in ways owned competence cannot. The market has largely failed to price that fundamental distinction into corporate valuations.
Data Center Consumption Outpaces Grid Infrastructure
The scale of the electrical demand is already straining national infrastructure. Berkeley Lab’s report for the U.S. Department of Energy shows data centers consumed roughly 76 TWh in 2018 and 176 TWh in 2023, accounting for roughly 4.4% of total U.S. electricity generation. Projections in that same report indicate demand could scale to between 325 and 580 TWh by 2028, absorbing up to 12% of total generation.
Compute providers are already experiencing the pinch. On its latest earnings call, Microsoft acknowledged that customer demand continues to outpace available capacity. Meanwhile, Amazon stated it will not have enough capacity to satisfy projected demand in 2026. The electrical grid mirrors these supply constraints directly.
PJM Interconnection, which coordinates wholesale electricity markets across 13 eastern states hosting high concentrations of data centers, saw its 2028/29 capacity auction clear at the price cap. Even with that pricing outcome, the auction missed its target reliability figures by roughly 6.8 gigawatts as a surge of new demand flooded the system. Physical hardware moves slowly: the Department of Energy notes that large power transformers require up to five years to deliver. Software scales in months, but transformers, high-voltage steel, copper, and regulatory permits operate on timelines measured in years.
Rationing Compute to the Highest Bidder
The immediate consequence of this supply mismatch will not be a sudden blackout, but a sharp price hike. Scarcity expresses itself directly in the cost per token, per query, or per agent-hour. Like any commodity market, compute will be rationed to the highest bidder. The first casualties of this shift will be the gross margins of products priced on the belief that machine intelligence only gets cheaper over time.
A growth model premised on falling costs disintegrates the moment the marginal cost of intelligence is dictated by the price of a megawatt. Very few corporate board decks currently stress-test for that specific sensitivity. Most organizations continue to treat electrical power the way early internet companies treated bandwidth: as a commodity that someone else will handle.
Infrastructure Realities and the Path Forward
Solutions to the power crunch are physical, and there are only two. The first approach involves building new electricity generation faster and closer to where compute resides. The second involves designing data centers capable of delivering identical compute loads on drastically reduced power.
While that second alternative sounds comforting, it proves far less reliable: past patterns show that making computing cheaper and more efficient consistently drives up overall demand instead of reducing it. Efficiency gains may buy time, but they rarely provide structural relief. Until either generation or efficiency solutions scale effectively, the prudent baseline remains that the power crunch will outlast corporate forecasts. Whoever resolves that bottleneck will capture a share of the value created by everything running on top of it.
