Cloud providers spent the last decade racing to build more data centers. Now the race has hit a wall that concrete and capital cannot solve: electricity.
Across the US and parts of Europe, utilities are telling hyperscalers that new capacity simply cannot be connected in time. Dominion Energy in Virginia, the country's largest data center market, has queued requests stretching years into the future. Ireland's grid operator has paused new data center approvals in Dublin until at least 2028. The constraint is not land, not capital, and not chips. It is megawatts.
This matters now because the same AI workloads driving record cloud revenue also draw far more power per rack than traditional compute. A single AI training cluster can pull tens of megawatts, and inference at scale is not far behind. The result is a squeeze: customers who expected cheap, abundant compute are finding prices rising and regions sold out.
Utilities, Not Chipmakers, Now Set the Pace
For years the limiting factor in cloud expansion was semiconductors. TSMC's capacity and Nvidia's allocation dominated every infrastructure conversation. That has not disappeared, but it has been joined by a harder problem. Power generation and transmission take longer to build than data centers, and in many jurisdictions the permitting process for new substations stretches past five years.
Microsoft said in its most recent earnings call that it is capacity constrained in several regions despite owning the land and having the hardware on order. Amazon Web Services has been buying power purchase agreements at a record pace, locking in future renewable output years ahead of delivery. Google has gone further, signing deals for advanced nuclear and geothermal power that will not come online until the 2030s.
These are not abstract planning exercises. They are direct responses to a physical ceiling. When a utility cannot deliver power, a completed data center sits idle, and the capital invested in it earns nothing.
Edge AI Was Supposed to Be the Answer
The edge computing pitch has always been about distance: move compute closer to users and devices, cut latency, reduce backhaul. Edge AI extended that pitch to inference, promising that models could run on devices and small regional nodes instead of central clouds.
The power crunch complicates that story. Edge nodes still need electricity, and they need it in places where the grid is often weaker than in major cloud regions. A retail chain deploying AI vision systems across hundreds of stores is not just buying cameras and accelerators. It is adding sustained electrical load to buildings that were never designed for it, often in areas with limited transformer capacity.
I think the gap between the edge AI marketing and the electrical reality is wider than most vendors admit. The hardware gets cheaper every quarter. The power to run it does not.
Prices Are Already Moving
Cloud pricing has been remarkably stable for a decade, with providers cutting rates as scale improved. That era is ending in capacity constrained regions. AWS raised prices on certain EC2 instance types in 2023 for the first time in years. Azure and Google Cloud have followed with targeted increases on high demand instance families.
The increases are not dramatic yet, but the direction matters. For the first time, the cost of compute is being driven by energy cost and grid scarcity rather than by hardware efficiency gains. In Europe, where natural gas prices remain volatile and carbon pricing adds cost, the effect is more pronounced.
Customers are responding by changing architecture. Workloads that once ran continuously in the cloud are being restructured to run during off peak hours or shifted to regions with spare capacity. Some firms are moving inference back to on premises hardware not because it is cheaper, but because they can control the power supply.
What the Industry Is Doing About It
The responses fall into three broad categories. First, efficiency: better cooling, higher density racks, and more efficient accelerators. Liquid cooling is moving from niche to standard in new AI data centers because air cooling cannot handle the thermal load of dense GPU racks. Second, location: building in places with surplus power, including rural areas and regions with stranded renewable energy. Third, generation: on site power, from natural gas turbines to small modular reactors.
Each has limits. Efficiency gains are real but diminishing. Location shifts add latency and complicate data sovereignty. On site generation raises emissions questions that conflict with the net zero commitments every major cloud provider has made.
The Edge Question Nobody Is Answering
The real test will be whether edge AI deployments can secure power at all. A hyperscaler can negotiate with a utility. A mid sized company deploying AI to two hundred retail locations cannot. Those deployments depend on the local grid, and in many markets that grid is already strained by electrification of transport and heating.
There is also a regulatory dimension. Some jurisdictions are beginning to treat data center power demand as a public policy issue rather than a commercial one. Ireland's moratorium is the most visible example, but similar discussions are underway in the Netherlands, Singapore, and parts of the US.
The irony is that edge AI was meant to reduce dependence on central cloud capacity. Instead it may just move the bottleneck from hyperscaler campuses to thousands of smaller sites that are even harder to power.
What This Means for Buyers
If you are planning AI infrastructure in the next two years, power availability should be a first line item, not an afterthought. Ask your cloud provider which regions have spare capacity and which are constrained. Ask your edge vendor how their hardware handles power fluctuations and whether it can throttle gracefully.
The companies that treat electricity as a procurement problem rather than an engineering problem will be the ones that miss their deployment targets. The ones that secure power early, even at higher cost, will have a durable advantage.
The cloud industry built its reputation on abundance. The next phase will be defined by scarcity, and the organizations that plan for it now will be the ones still shipping AI products in 2027. For more on how infrastructure constraints shape deployment decisions, see our cloud coverage.
According to Reuters, data center power demand has become one of the most closely watched variables in the global energy transition, with utilities and regulators scrambling to keep pace with AI driven load growth.