The AI boom is moving from server racks into household and business power bills. PJM's latest market data shows electricity is becoming one of the industry's hardest constraints.
The cost of running the biggest U.S. power grid just jumped in a way that should get every AI startup's attention. PJM Interconnection, which manages wholesale electricity across 13 states and the District of Columbia, saw its total wholesale power cost average $136.53 per megawatt-hour in the first quarter of 2026, up from $77.78 a year earlier.
That is a 75.5 percent increase, rounded to 76 percent, and it did not happen in a vacuum. PJM sits under one of the most important physical layers of the AI economy: Northern Virginia, the country's dominant data center market and the place where hyperscalers, cloud providers and colocation firms have spent years building the infrastructure behind modern computing.
As Bloomberg reported from Monitoring Analytics' latest market monitor report, the pressure is no longer just about whether the energy market is competitive. The more important question is whether the grid can absorb a rush of large, always-on power demand without pushing costs onto everyone else.
For the last two years, the AI conversation has been dominated by GPUs, foundation models and cloud capacity. That made sense. Nvidia chips were scarce, frontier models were expensive to train and startups lived or died by access to compute.
Now the harder problem is showing up one layer below that. Data centers need land, substations, transmission lines, backup power, interconnection approvals and enough firm electricity to run dense server halls at industrial scale. A clever model architecture helps with inference efficiency, but it does not remove the need for physical power delivery.
PJM's report shows how quickly that physical constraint can become financial. In the first quarter, real-time load-weighted average locational marginal prices rose 67.8 percent, from $52.20 per megawatt-hour to $87.57. The total cost increase was broader than energy alone. Capacity costs rose sharply too, up $14.21 per megawatt-hour, or 398.1 percent, while transmission costs also moved higher.
That matters because capacity prices are about preparedness. They reflect what the market must pay to make sure generation is available when demand peaks. If forecasts assume large new data center loads are coming, the market starts pricing that future need before all those facilities are even fully operating.
This is why the issue is bigger than a one-quarter price spike. Monitoring Analytics has already described the impact of data center growth on PJM prices as significant and irreversible in its recent market work. Once transmission projects are planned, capacity obligations are set and new load forecasts enter the market, the cost structure changes for years.
Startups will feel this through cloud bills
Most startups are not buying wholesale electricity directly from PJM. They feel this through AWS, Microsoft Azure, Google Cloud, Oracle, CoreWeave, colocation contracts and inference providers. Those companies are the ones negotiating power deals, siting data centers and managing utilization, but their costs eventually flow through pricing.
That does not mean every AI API call becomes expensive overnight. Cloud pricing is sticky, competitive and often protected by long-term contracts. But the direction is hard to ignore. If power, capacity and grid connection costs rise in the regions where AI infrastructure is concentrated, hosting providers will need to recover that money somewhere.
For an AI company serving millions of inference requests, training specialized models or offering low-margin automation tools, electricity becomes part of unit economics in a much more direct way. The cost of tokens is not just the cost of chips. It is also the cost of keeping those chips powered, cooled and connected to the grid.
That is why some AI companies are now thinking more like industrial operators than software teams. They care about where workloads run, when jobs are scheduled and whether a cloud region has room to grow. The cheapest compute may not be in the most familiar region anymore. It may be wherever power is available, interconnection queues are manageable and regulators are less likely to impose new cost rules after the fact.
The fight is shifting to who pays
The political problem is becoming just as important as the technical one. PJM's territory includes states that want data center investment, but also ratepayers who do not want to subsidize infrastructure built mainly for hyperscale growth.
Maryland's Office of People's Counsel recently filed a complaint at the Federal Energy Regulatory Commission challenging PJM cost allocation rules, arguing that Maryland customers could be assigned $2 billion in transmission capital costs tied to data center-driven upgrades. The complaint says those costs could raise Maryland customer bills by $1.6 billion over the next decade.
This is the next practical test for the AI buildout. Hyperscalers can sign power purchase agreements and promise clean energy procurement, but regional grids still need wires, substations, dispatchable generation and rules for who carries the risk if projected demand does not arrive exactly as forecast.
For the AI market, the lesson is simple. Compute is no longer just a cloud line item. It is an energy market exposure. The startups that understand that early will design with more discipline, choose infrastructure partners more carefully and watch power policy as closely as they watch model releases.
The next phase of AI will still be shaped by better chips and smarter models. But the companies that win may also be the ones that secure reliable electricity without turning their cost base into a moving target.
Also read: A missing Microsoft Hugging Face model raises questions for AI builders • Ex-Meta researcher Tian Yuandong launches a $4.65 billion AI bet • OpenMOSS gets a C++ port as local voice AI chases easier deployment
