Powering Intelligence: The Hidden Electricity Bill of the AI Boom

When the Grid Hits Its Ceiling

On July 14, PJM Interconnection published the results of its 2028-2029 capacity auction, and the outcome left little room for interpretation. The clearing price landed at $325 per megawatt-day, the regulatory maximum. PJM manages electricity reliability for roughly 67 million people across a territory stretching from Illinois through Virginia to the Washington, D.C. metro area, making it North America’s largest grid operator. When its auction hits the ceiling, the market is communicating something structural.

What it is communicating, according to a Fortune investigation published on July 24, is that data centers have become the dominant force reshaping electricity demand across PJM territory, with AI compute loads at the center of that expansion. The auction cleared, but with a supply shortfall of approximately 6.8 gigawatts compared to what PJM’s own planners believe the grid needs to remain reliably powered. That gap is not a temporary anomaly. It reflects a basic mismatch between the speed at which AI-driven demand is being added to the grid and the speed at which new generation can be financed and built.

The Numbers Behind the Auction

The dollar figures embedded in that auction tell a specific story. PJM’s independent market monitor, Monitoring Analytics, calculated that of the $16.4 billion in total capacity charges generated by the 2028-2029 auction, approximately $6.3 billion is directly attributable to data center demand. That is close to 40 percent of one auction’s total burden driven by a single customer category. Zoom out across the last four capacity auctions combined, and the figure attributable to data centers reaches $29.4 billion.

The projection that follows, drawn from Reuters reporting and analysis by consulting firm ICF and cited in Fortune’s investigation, is that households and small businesses within PJM’s territory could see their electricity rates increase by as much as 60 percent over the next five years. This is not a forecast tied to crude oil prices or seasonal weather patterns. It is a structural consequence of the data center construction wave that AI has set in motion, and it reflects how capacity markets distribute costs across every customer connected to a shared grid.

Who Actually Pays, and Why That Matters

The core issue here is not technical. It is about cost allocation. When a hyperscale operator activates a new AI campus consuming 200 megawatts or more, the grid must be reinforced to serve that load. Under current U.S. regulatory structures, the capital required for that reinforcement is socialized: spread across every customer on the grid through capacity charges, from large industrial users to residential accounts.

Capacity markets have always functioned this way, and for decades the mechanism worked without visible distortion because no single category of demand grew fast enough to shift the aggregate cost materially. AI data centers are different in both scale and pace. TechCrunch’s preview of the Smart Systems Stage at Disrupt 2026, published on July 27, framed “AI’s power problem” as a central theme for the near term, noting that as compute demand rises sharply, data center operators and energy companies are racing to secure adequate power before electricity becomes the bottleneck slowing AI’s next phase of growth.

For business leaders outside the technology sector, especially in manufacturing and logistics, a 40 or 60 percent rise in electricity costs over five years is a material change to operating economics. It affects capital investment decisions, facility location choices, and competitive positioning in ways that have nothing to do with AI adoption. The cost is not limited to the companies building data centers; it radiates across the broader industrial base of a region.

The Geography of the Problem

PJM’s footprint is not a peripheral market. It covers much of the U.S. industrial heartland and the northeastern seaboard, including the Northern Virginia data center corridor, which has become the densest concentration of hyperscale compute anywhere in the world. Load growth is concentrated in relatively small geographies, but the cost is spread across a far larger population.

That dynamic is already influencing where AI infrastructure gets built. Site selection teams for hyperscale operators are increasingly focused on regions with surplus low-carbon power, whether from hydroelectric, geothermal, or emerging nuclear programs, that can serve large loads without burdening existing grid customers. The ability to offer reliable and cost-stable electricity is becoming a genuine competitive advantage for jurisdictions competing for AI investment.

What the Industry Is Doing About It

Responses are forming on multiple fronts. Some large technology companies are signing long-term power purchase agreements for dedicated renewable or nuclear generation, aiming to add supply rather than draw from shared capacity markets. Others are pursuing on-site generation or small modular reactor partnerships to reduce their exposure to auction-driven cost swings altogether.

On the hardware side, startups developing specialized AI chips argue that their architectures can deliver meaningfully better energy efficiency than general-purpose GPUs, reducing the power drawn per unit of compute. Whether those gains will offset continued growth in total AI workloads remains an open question, and one that capacity markets will ultimately price in regardless of the answer.

What is clear from the July 14 auction is that the cost of AI infrastructure has already moved from technology balance sheets toward everyday utility bills. The next auction is unlikely to produce a different result.

As regulators, utilities, and technology companies negotiate how to share the cost of powering the next era of computing, tens of millions of households inside PJM’s territory have already provided their answer in advance: they are paying for it, whether they know it or not. The regulatory question of who should bear that cost, and in what proportion, will only grow louder as more data centers come online. What happens next will depend largely on whether policymakers choose to redesign those cost structures before the full bills arrive, or after.