When the Grid Becomes the Scheduler: Power Is Deciding Where AI Lives

By Bill Kleyman

When I first started in this industry, placing an IT workload was straightforward. Generally, it was an architecture conversation. We asked about latency, security, data sovereignty, cost, and access to compute. Power was somewhere below all of that, assumed to be readily available at roughly the same cost wherever we put the workload.

Oh, how we’ve flipped that hierarchy.

Flexential's 2026 State of AI Infrastructure report found that reliable grid power influences deployment decisions for 89% of organizations. When asked the top reasons they place an AI workload in a given region, organizations rank power cost or pricing differences first, cited by 55%, more often than data residency, security, performance, and even AI compute availability.

This is bigger than deciding where a hyperscaler builds its next campus. Power is now helping determine where enterprises run inference, which cloud region they select, whether they use colocation, and when a workload moves closer to its data. The grid has quietly become a workload scheduler. Kubernetes would like a word.

Electrons Are the New Availability Zone

The AFCOM State of the Data Center 2026 report tells a similar story from inside the facility. Seventy-four percent of respondents plan to deploy AI-capable infrastructure, 72% expect AI to increase capacity requirements, and average rack density jumped from 16 kW to 27 kW in one year. Among organizations bringing workloads back from the cloud, 81% reported higher power demand and 36% called the impact significant.

Those numbers connect IT placement directly to electrical infrastructure. An application team may prefer a particular location for latency or governance, but if that market can’t provide capacity at a workable price and timeline, the architecture changes. Increasingly, the most important AI availability zone may be the one with available, affordable electrons.

From Power Customer to Power Industry

The speed and size of data center electricity growth is forcing uncomfortable planning questions for some operators, utilities, regulators, and customers. Gartner forecasts that global data center electricity use will reach 565 TWh in 2026, up 26% in one year. It attributes 175 TWh, or 31% of data center electricity use, to AI-optimized servers; cooling and other infrastructure are separate at 195 TWh, or 35%. Gartner predicts AI server’s electricity use will rise 48% in 2027 and shoot past conventional servers’ use. In the United States, the Department of Energy and Lawrence Berkeley National Laboratory estimate data centers could rise from 4.4% of national electricity use in 2023 to between 6.7% and 12% by 2028.

Energy consumption is only half the story. The other half is who generates the power and how they do it. AFCOM found that 65% of operators are using or considering on-site generation. Twenty-five percent already generate on site, 23% plan to within a year, and another 17% are exploring it. As data centers implement microgrids, batteries, fuel cells, turbines, power purchase agreements, and potentially even nuclear generation, they’re no longer simply buying energy. They are procuring, storing, managing, and increasingly producing it.

Put more provocatively, data centers may be becoming one of the world's largest quasi-energy industries without a matching utility-style framework. They are not unregulated: air permits, zoning, interconnection rules, environmental reviews, utility commissions, and reliability standards all apply. The problem is fragmented oversight and new large-load rules. This is not a case for a blanket moratorium. Targeted pauses may be reasonable where basic protections are missing, but the better answer is transparent policy that makes projects pay their way and meet clear milestones.

The Ratepayer and Permission Test

It’s important we really acknowledge that the gap between electricity demand and supply risks landing on real people. PJM's July capacity auction cleared at $325 per megawatt-day, near its approved cap. Data center growth is not the only cause. Other industries and uses are increasing demand as well, as we electrify more of the economy. Generation retirements, supply-chain delays, interconnection backlogs, and market design all matter as well. Still, PJM's independent market monitor identifies data center load growth as a major driver of recent capacity-market pressure. Capacity prices are only one component of a retail bill, but citizens understandably ask who will pay for the generation and transmission that’s built to serve enormous new loads.

While many data center developers are explicitly promising to pay their way for new infrastructure and power, people are wary. A Gallup poll released in May 2026 found that seven in 10 Americans oppose an AI data center in their local area, including 48% who strongly oppose one. Energy, water, utility bills, pollution, and quality of life were all cited. In July, New York paused certain new data center permitting while it develops statewide standards and a community-benefits framework. Moratoria are blunt instruments, but they’re also a warning light on the dashboard.

Public trust is now one of the biggest threats to digital infrastructure deployment, possibly the biggest. The scarce resource isn’t always land, chips, or transformers. Sometimes it is permission.

The operators and regulators doing this well are making the compact clear: growth should pay for growth. Georgia created special terms for new loads above 100 MW to reduce cost shifting. Virginia approved a separate GS-5 rate class for customers at or above 25 MW so their unique costs can be recovered more directly. Long contracts, minimum bills, collateral, exit fees, upfront infrastructure contributions, credible load forecasts, and demand-response commitments can protect households and small businesses while reinforcing the grid for everyone.

Doing it poorly means opaque negotiations, speculative capacity requests, late engagement, vague promises, and assuming upstream risk belongs in the general rate base. Some opposition is about local impacts, and some is about AI itself. Neither can be dismissed as simple “NIMBYism.”

Rebuilding the Electrical Backbone

The response can’t be generation alone, whether it’s by utilities or on-site at data centers. We also need to move power differently. Emerging 800 VDC designs can reduce conversion stages and support megawatt-class racks. Schneider Electric is developing 800 VDC sidecars for racks up to 1.2 MW, while Eaton is advancing a similar architecture with NVIDIA. Solid-state transformers may eventually convert medium-voltage power directly to high-voltage DC, with faster control and a smaller footprint.

But this is not a magic power strip. DC protection, fault isolation, standards, technician safety, interoperability, and maintainability still need work. AFCOM's survey found no dominant distribution model for future builds, and 34% of respondents remain unsure. That uncertainty is healthy. At this stage, confidence without extensive testing would be a very expensive outage waiting to happen.

Smarter controls matter just as much. Digital twins, electrical power-management systems, automation, and grid-interactive batteries can help operators forecast demand, shift flexible workloads, smooth peaks, and support grid services. Microsoft has already used data center batteries in Ireland to help stabilize grid frequency, while Google says it uses demand response and workload flexibility when possible. Not every inference workload can wait, but not every compute job is equally urgent either.

Data center demand is moving new energy technologies from pilots to projects. Google and Cypress Creek just broke ground on Arkansas' Steel River Energy Center, planned for 2.5 GWdc of solar and 2.9 GWh of battery storage and billed as the nation's largest solar project to date. Meta has contracted nearly 1 GW of new solar and wind in India, partnered with Sage Geosystems on up to 150 MW of advanced geothermal, and reserved up to 1 GW/100 GWh of long-duration storage using reversible solid-oxide fuel cells. Bankable demand can finance projects that might otherwise wait. But annual renewable matching is not hourly availability, and transmission, storage duration, cost, and permitting still matter. Nuclear is advancing too: Google's Kairos agreement targets up to 500 MW by 2035, while Meta's Constellation agreement supports the 1,121 MW Clinton, Ill., plant and a 30 MW uprate without ratepayer support. Promising, yes. Instant electrons, no.

Inference Changes the Map

Jensen Huang has called this the "age of inference". That matters in terms of the power discussion because AI training tends to concentrate enormous compute into fewer locations, while inference follows users, factories, hospitals, devices, and enterprise data. It is continuous, latency-sensitive, and increasingly distributed. Organizations will need power to run AI in data centers located in communities across the globe.

The result will not be one universal AI factory design. Training may gravitate toward regions with large blocks of affordable power. Inference may spread across colocation, enterprise facilities, regional clouds, and edge environments. Efficiency will improve, but lower cost per token can lead to more token use, more agents, and more total demand. The future is not only bigger infrastructure. It is power-aware infrastructure placed much more intentionally.

More Questions Than Answers

The hardest truth is that we still have more questions than answers about the power that’ll be required by AI and other digital workloads. How much speculative load should utilities build for? Which compute jobs can be delayed or shifted across regions to relieve grid stress without harming customers? Who bears the cost of substations, transmission, or generation built for a data center project that is delayed, downsized, or canceled? When does behind-the-meter generation become utility-like? How do we add capacity without sacrificing affordability, reliability, or community trust?

Those are exactly the conversations we will take on at Data Center World POWER, September 21-23 in Dallas. The event will go beyond how we generate and distribute electricity. We will examine the future of how humans interact with energy: how communities share in its value, how operators and utilities make decisions together, how software turns facilities into grid partners, and much more.

Because the next era of digital infrastructure will not be powered by electrons alone. It will be powered by trust, intelligence, and people.