Post: Could Virtual Power Plants help power the AI Boom?

Could Virtual Power Plants help power the AI Boom?

AI’s growth curve and the grid’s build-out curve are no longer moving at the same speed and that gap is now the one of the biggest constraints on how fast the AI industry can scale.

Between 2020 and 2025, US electricity demand grew about 1.7% annually compared with 0.1% annual growth between 2005 and 2019. But demand has been rising steadily since 2020, driven by data center growth, the U.S. Energy Information Administration (EIA)’s Short-Term Energy Outlook (STEO) has found.

Published in March 2026, the report notes that development of large computing facilities and growth from expanded industrial use of electricity are likely to continue driving growth in US electricity demand in the near term. It forecasts that electricity load will increase by 1.9% in 2026 and 2.5% in 2027, with the highest load growth being in the ERCOT and PJM regions where forecast growth of annual electricity load averages 10% and 3%, respectively, between 2025 and 2027. Midcontinent Independent System Operator (MISO) and the Southwest Power Pool (SPP), will also likely experience high growth in power demand from data centers.

In a recent report Gartner said it expects global data center electricity consumption to hit 565TWh in 2026, up from 447TWh in 2025, with worldwide data center power demand climbing 27% this year alone to reach 132GW. By 2030, Gartner puts that figure at 290GW.

In the US, Goldman Sachs Research projects data center power demand will climb from 31GW in 2025 to 41GW in 2026, then to 66GW in 2027. That pushes data centers’ share of total US peak summer power demand from 4.1% in 2025 to 8.5% by 2027. But regional pressure is uneven: Mid-Atlantic, Mid-Continent and Northwest power markets face the sharpest reliability risk because planned generation additions are not keeping pace with the load entering into their interconnection queues. This means that power availability has become a bottleneck on how fast hyper-scalers and enterprises can bring AI capacity online.

Utilities and grid operators are increasingly writing curtailment obligations directly into interconnection agreements for new large loads. Instead of guaranteeing every megawatt of contracted capacity around the clock, operators are asking data centers to accept dispatchable, on-call reductions during system stress events in exchange for getting connected to the grid.

Texas’s SB6 is the clearest example: loads above 75MW face mandatory remote curtailment requirements, with a compliance threshold that took effect at the end of 2025. Other markets are moving in the same direction. Data centers get speed to power, and in return, they take on a share of the flexibility burden.

That shift changes what “reliable power” means for data centers; it’s now about being able to prove, in real time, that curtailment calls can be met in full, every time, without breaching customer Service Level Agreements (SLAs).

The VPP story making headlines

One response to the grid squeeze has come from the demand side of the market. Distributed, behind-the-meter capacity that already exists is faster to activate than anything that has to be built from scratch.

In mid-2026, a coalition of home battery, EV and smart-thermostat providers announced a framework to pool more than 16GW of flexible capacity and offer it to data centers and utilities as a “capacity-as-a-solution” resource. A similar approach has appeared elsewhere in the market, with “bring your own capacity” style deals pairing large cloud and technology buyers directly with Virtual Power Plant (VPP) aggregators to unlock incremental capacity on constrained grids by pooling flexible assets.

But residential VPPs depend on millions of individual households choosing to opt in and stay opted in. Research on managed EV charging programs has found participation rates as low as 1% without a meaningful financial incentive, rising to under 5% even with payments worth 15% of a household’s power bill. Analysts have also questioned how quickly framework announcements convert into contracted, dispatchable megawatts on the ground. Aggregating enough committed, reliable capacity from a fragmented demand base to matter to a single hyperscale campus is a different engineering and commercial problem.

A commercial and industrial answer

There’s a second, less publicized version of this story playing out directly at the data center campus level. Most large data centers already have, or are adding, behind-the-meter batteries, fuel cells, onsite solar, generators and capacitors as backup and resiliency infrastructure. Separately, compute operators are starting to explore workload flexibility: training jobs that can be deferred, batch jobs that can pause and resume, GPU and cooling power that can be scaled down, all without touching latency-critical inference traffic. The problem is that neither lever works well enough on its own.

A battery-only response is capital-intensive to scale and tends to deplete before the utility releases the curtailment signal. A compute-only response hits a hard ceiling the moment SLA headroom is exhausted, since inference workloads serving live traffic can’t be paused without breaching commitments, and compute operators typically have no visibility into the tariffs, dispatch instructions or settlement requirements driving the utility’s signal in the first place.

This is the gap GridBeyond’s Virtual Power Plant and EMS platform is built to close, by co-optimizing both levers from a single orchestration layer rather than treating them as separate systems.

When a utility or ISO curtailment signal arrives, whether it’s a primary emergency dispatch instruction or a secondary market price, frequency or carbon signal, the platform is asset-agnostic across the full Bridge to Power stack: FTM and BTM batteries, fuel cells, onsite solar, gas peakers, capacitors and other DERs. It splits each curtailment call optimally across physical asset dispatch and compute-side action (throttle, pause, shift, cross-region migration), hitting the utility’s full MW target at the lowest cost while protecting site resiliency and SLA compliance throughout.

The difference shows up clearly against real event data. Modeled against PJM’s June 2025 demand response event (a 161,770MW system peak, the highest since 2011, with -4,000MW dispatched across Mid-Atlantic and Dominion zones) a 100MW campus running compute flexibility alone could sustain a 33MW reduction for the full seven-hour window, but only by drawing the rest from the grid. The same campus co-optimized with a 470MWh battery system held a full 0MW net grid draw for the entire event: 33MW from compute flex, 67MW from battery discharge, working in tandem rather than in isolation.

Across curtailment types, the pattern repeats. In an emergency response scenario with a ten-minute utility notice window, the battery bridges the gap in under a millisecond while compute flexibility ramps to a 30% load reduction within 30 seconds — together hitting target well inside the notice window. In a sustained multi-hour peak, compute flexibility carries the duration at up to 40% sustained reduction while the battery handles the ramps and transitions, avoiding the three-to-five-hour depletion point that defines the ceiling of a battery-only stack. And between events, the same orchestrated infrastructure keeps earning from capacity markets, ancillary services and demand response. This means compliance is the floor, not the whole story.

Two models, one direction of travel

Residential VPPs and site-level co-optimization are solving different parts of the same problem from opposite ends of the grid. Aggregated home batteries and thermostats add new, dispatchable capacity to the wider system and can ease regional peak pressure at scale, provided enough households stay enrolled and the capacity converts from framework to firm contract. Site-level orchestration, by contrast, doesn’t add capacity to the grid so much as it makes each individual data centers own existing infrastructure reliable enough to meet a curtailment obligation without threatening uptime, using assets that are already paid for and already on site.

For a data center operator facing an interconnection agreement with curtailment terms attached, both matter. But only one of them is directly within the operator’s control: what happens behind their own meter, the moment a signal arrives.

What this means for data center operators

Curtailment compliance is quickly becoming a condition of grid access itself, not an optional cost-saving exercise. GridBeyond’s approach delivers full compliance on curtailment events, reaching the utility’s MW target in under 60 seconds from signal to confirmed reduction, with zero SLA breaches because critical workloads sit behind a dedicated orchestration layer that never touches them.

That translates into different value for different stakeholders on the same site. For the data center operator, it means the interconnection agreement and permit are protected, SLAs hold, blended energy costs come down, and the same on-site assets that exist for compliance also generate new market revenue between events. For the enterprise or AI customer running workloads on that infrastructure, it means SLA stability drawn from genuine headroom rather than critical capacity, continued access to a site where curtailment is now a condition of connection, and confidence that compliance risk sits with the orchestration layer rather than with them.

The AI boom isn’t going to wait for the grid to catch up on its own timeline. Whether that gap gets closed by millions of home batteries, gigawatts of on-site flexibility, the data centers that treat curtailment compliance as core infrastructure now will be the ones still running when the next emergency dispatch signal comes through.

USA | Curtailment Compliance

GridBeyond‘s Virtual Power Plant (VPP) co-optimized with leading compute curtailment platforms, giving data centers a holistic flexibility stack built to hit every utility signal, every time.

Learn more