Power Access for AI: The Flexibility Compact

The Access-to-Power Bottleneck

“The cost of intelligence should eventually converge to near the cost of electricity,” writes Sam Altman, forecasting a future where intelligence becomes as cheap and abundant as electricity. The promise of affordability in both intelligence and electricity, however, remains encumbered by the bottlenecks of power grid access caused by static assumptions about AI data center consumption and risk-aversion on the part of grid managers.

Power utilities still plan for worst-case scenarios in both grid forecasting and interconnection studies. But AI data centers do not have to be built—or operated—in ways that trigger those worst-case outcomes. In fact, they can help relieve them. As the U.S. fleet of data centers expands quickly, state governments, regulators, and industry leaders can act together to enable that buildout with new models of grid planning.

Before gigawatt-scale loads became common, grid planners assumed large power consumers were inflexible: always drawing their full demand, even during grid peak hours. This assumption still drives two key drags on grid power access for AI.

First, at the regional level, planners solve for regional grid stability by assuming every large load user will pull full power at the same time. That triggers requirements for long-term upgrades to the distribution and transmission system—new lines, substations, and transformers—to handle a total peak that rarely occurs.

Second, when generators and batteries are sited with loads, today’s grid study scenarios treat all of these pieces as cumulative, worst-case grid draw contributions, not as flexible offsets to each other. This means higher site-specific upgrade costs that assume everything draws power at once. The resulting delayed access is why so many sites are exploring grid-isolated, behind-the-fence configurations for data centers, generators, and batteries.

Magnifying these unhelpful outcomes, regulators are increasingly assuming AI compute sites are inflexible, peak-heavy stressors on the grid. That assumption is fueling reactive policy conversation: transmission cost-allocation policies, for example, assume the only answer is to overbuild infrastructure and get someone (ratepayers, data center customers, or both) to pay for the overbuild.

There is a deep asymmetry to this problem. Markets already can motivate power producers to curtail their output at moments of oversupply; there is no framework to incentivize large power consumers to cut demand when power is scarce. Even though data center sites are already deploying advanced onsite storage, compute workload flexibility, and grid-responsive site controls or energy-efficient power profiles, the grid planning process treats them as static loads.

We need a new grid planning model where large loads earn expedited interconnection by offering verifiable, real-time agility—and where planners, in return, treat that flexibility as real, measurable, and valuable in their reliability studies. And moreover, can combine them so that the datacenter fleet is a tool for grid stability. It is a two-way bargain that is the likely, tenable path to keep AI growing in the United States without overwhelming the grid.

Verifiable Operating Profiles

The missing piece is verifiable models of how large data centers will adjust load in practice. Today, grid planners will not grant transmission service based on load flexibility profiles which they cannot verify, and operators lack the data to prove how a data center will actually run—neither seem to have a complete set of tools to enforce this behavior to match the modeled behavior, even if they agree on it.

Grid flexibility and system planning, forecasting, and operator experts are getting together in research and deployment collaboratives like the Electric Power Research Institute (EPRI)’s Data Center Flexible Load Initiative to outfit data center sites with flexibility tools to make them adaptive to the needs of their local power grids. EPRI recently reported that its initiative has brought together Google, Duke Energy Corporation, Oracle, NVIDIA, Emerald AI, Salt River Project, Data4 Group, Schneider Electric and Réseau de Transport d'Électricité to testbed data center flexibility, utilizing AI workload choreography to delay and defer power usage that is coincident with grid stress hours: all of while respecting compute contract commitments, building grid-connected or site co-located batteries as longer duration buffers and creating uninterruptible power supply solutions for real-time grid disturbances. In a separate partnership with the Department of Energy’s National Renewable Energy Laboratory, Verrus has demonstrated how non-GPU IT infrastructure can provide grid-interactive flexibility, using a 70 megawatt data center testbed to showcase fast load curtailment, seamless islanding, and dynamic auxiliary control systems.

This innovative work portends a future where (1) AI compute centers can be studied with guaranteed operational profiles capable of removing incremental power loads from the grid to a margin that results in zero peak-load contribution, and (2) grid planners attain confidence in using these profiles to solve smarter planning cases to bring sites on faster with less risk of infrastructure overbuild and commissioning delay.

Price Signals for Flexibility in Grid Policy

Grid flexibility at data center campuses will only happen if policy incentives create a market for them and prompt stakeholders to arrive, data in hand, at the first utility-initiated grid study meeting and provide those inputs in a rigorous review during steady state and dynamic interconnection analysis.

Here is what those incentives should look like:

1. Utilities should develop rigorous, standardized, and transparent large load interconnection and commissioning standards that are willing to accept data center customers with defined operational criteria for flexibility.
2. Contractual commitments between sites and grid providers should firm up the obligation for sites to operate in the way they claim they will in interconnection applications.
3. In some, but not all cases, these requirements may be incorporated into tariffs approved by a regulator—but in all cases, the requirements should be publicly posted so that large load customers come to the table with an action plan to meet utility standards and deliver on them with preferred or reference technology selections.
4. As a rule of thumb, the flexibility compact is that grid customers should get faster, or more capacity in interconnection, in exchange for compliance with a mandatory operations plan for curtailment in peak hours and emergency conditions. This would occur through site controls which give grid operators the confidence that there is a documented and enforceable method for the grid operator and grid user to agree on a standard operating procedure for how, and on what timescale and magnitude, a data center load will gracefully remove peak load and bring it back to the grid without generating disturbances or imbalances.
5. While the grid operator can be agnostic on which technology stack serves the back end of the site’s curtailment function (i.e., software for IT load, compute workload, batteries, and generators), the outcome should be a predictable and verifiable, telemetry-based handshake between the customer and the grid.

Early Activation of Flexibility Solutions for Better Grid Study Outcomes and Capital Allocation

Flexibility solutions described above will only be engineered into planned campus sites if there is a market for those first principles solutions. Enabling a flexibility compact means that data center builders and customers at these sites can coinvest private capital in flexibility-enabling solutions, and utilities can focus on right-sizing the investments that are actually needed to serve these customers without overbuilding for a rare coincident peak event that can be avoided entirely.

This approach also mitigates the risk of stranded costs tied to infrastructure sized for a peak that never arrives. By aligning capital deployment with realistic, dynamic load profiles, the system cost of AI buildout can be significantly reduced—ultimately protecting all ratepayers from bearing the price of unnecessary grid expansion driven by outdated assumptions. Technology advocates must strive to make these solutions common parlance in the AI industry and investor C-suites as much as they must do so with legislators, regulators, and data center builders. Until then, these solutions will feel like the edge of possibility while we remain trapped in the circle of too much demand and too little supply.

Power Access for AI is an Infrastructure Issue

The abundance of power for the AI era will be defined by the speed and creativity of our choices to ensure AI compute sites are customer-friendly, grid-aware, and technically durable. The flexibility compact converts uncertainty into a shared asset, aligning incentives so that every party—developer, utility, regulator, and customer—wins. Planned, enforceable, and priced, this approach will enable collaborative, cocreated success for AI infrastructure in the United States.

Arushi Sharma Frank is a senior associate (non-resident) with the Energy Security and Climate Change Program at the Center for Strategic and International Studies in Washington, D.C.

The author also advises Emerald AI and industrial data center developers involved in battery storage, flexible site controls, and retail energy solutions.