The PJM Interconnection, which runs the largest power grid in the United States, set a record capacity price of $269.92 per megawatt-day for the 2025-2026 planning year, up 22 percent from the prior record. PJM has since agreed to cap capacity prices near $325 per megawatt-day through 2030, following a deal meant to shield customers from extreme spikes. The driver is not weather or fuel. It is data centers.
PJM's own market data shows data centers caused a $9.30 billion price increase across the 2025-26 capacity auction. For residential customers in the thirteen states PJM serves, that translates to roughly $18 more per month on electric bills, according to the Avid Solutions projections cited in industry analysis. For a mid-sized cloud provider leasing colocation space in Ashburn or Columbus, the math is worse, because capacity charges now make up more than a quarter of the total bill for many commercial accounts.
This matters right now because the same grid that cloud providers relied on for cheap, abundant power is becoming the binding constraint on where new capacity can be built. A Ropes & Gray analysis of data center investment in 2026 notes U.S. data center power demand could reach 35 to 45 gigawatts by 2030, roughly double 2024 levels. The mismatch between that demand curve and the pace of new generation is not a planning problem that resolves itself. It is a physical problem, and it is showing up in site selection decisions today.
Grid scarcity is rewriting where clouds get built
Cloud providers used to chase cheap land and tax incentives. Now they chase megawatts. The Cushman & Wakefield 2026 Data Center Development Cost Guide puts average greenfield costs at $17.6 million per megawatt for the most modern facilities, with wide variation by market and end use. That figure does not include the interconnection queue, the transformer lead times, or the community opposition that now accompanies most large projects.
According to ABI Research, active AI-dedicated data center capacity in 2026 stands at 3.1 gigawatts, or 44 percent of total installed capacity, with 741 megawatts of new AI capacity being added this year alone. That is capacity that has to be energized, cooled, and connected to fiber in markets where the local utility is already telling developers to wait. The cloud providers with the deepest pockets, Amazon, Microsoft, and Google, can afford to build their own substations and buy power purchase agreements years in advance. Smaller colocation operators and enterprise private clouds cannot.
The result is a two-tier market. Hyperscalers secure power first and lease space second. Everyone else competes for whatever capacity is left, at prices that reflect the scarcity. DataCenterHawk's 2026 colocation pricing guide shows retail colocation, at 250 kilowatts or less, carrying the highest per-unit pricing of any deployment tier, while wholesale and hyperscale deals get better rates per megawatt. The gap between retail and wholesale pricing is widening, not narrowing.
Edge computing is where the growth is going instead
If the grid cannot absorb another 500-megawatt campus in Northern Virginia, the workloads that do not need that much power move closer to the user. Persistence Market Research values the global edge data center market at $19 billion in 2026, projected to reach $60 billion by 2033. That is a 17.3 percent compound annual growth rate, and it is being driven by inference workloads that need low latency more than they need raw compute.
Edge sites are smaller, typically one to five megawatts, and they are often built on existing utility infrastructure that has spare capacity. A retail chain running computer vision at its stores does not need a hyperscale campus. It needs a cabinet in a regional facility with a 50-millisecond round trip to its point of sale. That is a different business than training a foundation model, and it is growing on a different timeline.
The S&P Global analysis of 2026 data center trends points out that leased data centers can offer a secure place to store data at lower cost than public cloud storage, with automated interconnection. That is the pitch edge operators are making to mid-market customers who are tired of egress fees and capacity charges from the big three. It is not a replacement for public cloud. It is a release valve for workloads that never needed hyperscale in the first place.
Liquid cooling is no longer a niche choice
The AI clusters going into these facilities generate heat densities that air cooling cannot handle. Liquid cooling, both direct-to-chip and immersion, has moved from pilot projects to standard specification for new builds above 30 kilowatts per rack. The tradeoff is water. A facility that uses evaporative cooling in a drought-prone region is not a good neighbor, and regulators in Arizona, Texas, and parts of California are starting to ask questions about water rights before they ask about power.
This is where the physical constraints start to interact. Power, water, land, and fiber are all finite. A site that has three of the four is not buildable. Operators are now running site selection as a multi-constraint optimization problem, and the winning sites are the ones that already have the infrastructure, not the ones with the cheapest land.
What this means for buyers
If you are buying cloud capacity in 2026, the price you see on the rate card is not the price you will pay. Capacity charges, egress fees, and minimum commitments have turned cloud pricing into a negotiation. The Fortune Business Insights projection that the global data center market grows from $319.70 billion in 2026 to $775.83 billion by 2034 tells you where the money is going. It does not tell you whether your bill will go up or down.
My read is that it goes up for most buyers. The grid is not going to get cheaper, cooling is not going to get simpler, and the demand from AI training is not going to slow down. The edge market will absorb some of that pressure, but edge capacity is a complement, not a substitute. The buyers who win are the ones who lock in capacity now, negotiate egress caps into their contracts, and stop assuming that cloud prices only go down.
The next twelve months will show whether the PJM price cap holds and whether new generation actually gets built. If it does not, the capacity crunch moves from a regional problem to a national one, and the cloud pricing model that has defined the last decade stops working.
For more on data center investment trends, see the Ropes & Gray 2026 analysis. Related coverage of cloud infrastructure is available at our cloud and edge category page.