Data-center operators ready to negotiate grid flexibility in exchange for faster interconnection timelines.

As the U.S. faces an electricity affordability crisis alongside surging demand for the first time in decades, the Federal Energy Regulatory Commission's June 18 order directing system operators to provide transmission for flexible large loads has underscored an urgent imperative: data centers must adopt load flexibility.

The benefits are substantial. A 2026 Duke University Nicholas Institute study found that even a 1% to 2% reduction in data center peak demand can lower electricity rates by 0.5% to 2.8% while protecting reliability. Research by the Electric Power Research Institute (EPRI) demonstrates that such flexibility is achievable and can accelerate data center interconnections to the grid.

The pressure is intense. Cloud computing demand is driving this urgency globally. From Q1 2025 to Q1 2026, Amazon Web Services' cloud business grew 28%, Microsoft Azure grew 40%, and Google Cloud revenues surged 63%, according to Halcyon's chief strategy officer Nat Bullard. "Growth rates this high in already-mature businesses mean total revenue doubles in two years or less," Bullard noted. "That revenue can only be serviced with compute, and that compute can only serve when energized."

Projections underline the scale of the challenge. Goldman Sachs reported in May that U.S. data center power demand will reach 66 GW in 2027, up from 31 GW in 2025, with summer peaks growing from 4.1% to 8.5%. The EPRI estimates that data centers could consume as much as 17% of U.S. electricity usage by 2030.

Flexibility offers a solution to multiple stakeholders. The North American Electric Reliability Corporation acknowledged in its 2026 large load risk mitigation guidelines that data center flexibility can reduce electricity demand and, consequently, the ratepayer costs associated with maintaining reliability. Public opposition to data centers—chiefly centered on local electricity costs—can be addressed through flexibility. Utilities, researchers, and analysts agree that flexibility can accelerate interconnections, benefit system reliability, and increase system utilization, thereby lowering rates.

EPRI's FlexMosaic framework operationalizes this concept. Flexibility adds "headroom" to power systems by enabling operators to add large loads while maintaining reliability and minimizing infrastructure investment. The framework identifies five flexibility "classes": Class A supports power systems during infrequent extreme stresses; Class B manages daily or weekly demand peaks; Class C addresses long-lasting energy shortages; Class D protects against sudden supply or demand swings; and Class E provides frequency stabilization. The flexibility class depends on notification time needed, response duration and frequency, and the depth and speed of response. Data centers with Class D and E flexibilities that mitigate local thermal overloads or voltage drops unlock the most system value.

To achieve flexibility, data centers can leverage three pillars: managed workloads, reduced AI plant energy consumption, and back-up power, according to Anuja Ratnayake, EPRI's emerging technologies executive leading the DCFlex Initiative. FlexMosaic aims to standardize both data center designs and utility programs that incorporate flexibility.

Workload flexibility is the innovation that changes the calculus. Boston University researchers reported in June that today's AI data centers can "be designed with flexibility as a core operational principle." Modeling found that training and inference workloads "can offer between 18% and 55% flexibility relative to their average power consumption" while still meeting quality of service requirements.

Emerald AI demonstrates how this works in practice. The platform operates as an orchestration and optimization layer between utilities and data centers, translating grid requirements into operationally feasible dispatch targets while preserving utility dispatch authority. "Every utility-data center interaction is governed by operational parameters agreed upon by both parties in advance," said Mansi Shah, Emerald AI's head of product. Utilities define event parameters including maximum magnitude, minimum notice period, frequency limits, and event duration. Data centers set "hard floors that protect critical workloads and infrastructure under all circumstances." This arrangement gives utilities confidence that contracted flexibility will perform reliably while assuring data centers that operational and technical boundaries will never be violated.

The approach has been validated in demonstration projects. EPRI, NVIDIA, and utility demonstrations using Emerald AI software have shown that training workloads can be paused or slowed and inference queries redirected to less stressed data centers. In a peer-reviewed test, Emerald AI's platform ramped an AI workload at Salt River Project 25% during a three-hour peak. Another test shifted a workload from Virginia to Illinois without service degradation. A UK test reduced an AI load over a third in under a minute while protecting critical compute. Portland General Electric's tests showed 20% power reductions in simulated weather emergency scenarios.

These demonstrations have achieved flexibility rates of up to 40%, with future infrastructure designed for flexibility promising material improvements. The company has proven both "temporal flexibility" by slowing or pausing AI workloads and "spatial flexibility" by rerouting latency-sensitive loads across the country to where power is available.

In late 2026, EPRI, NVIDIA, Emerald AI, and partners plan to bring the 96-MW Aurora AI Factory online in Manassas, Virginia, to validate workload flexibility at scale in a data center designed for this purpose.

The remaining challenge is standardizing agreements between utilities and data centers. An upcoming EPRI-led pilot will address this gap, with Emerald AI working with Silicon Valley Power (SVP), which serves 58 data centers in its 20-square-mile territory. In phase one, Emerald AI is developing a bidirectional communication platform with an NVIDIA data center so that SVP can monitor real-time load, send signals to reduce load in response to simulated events such as heat waves or lightning strikes, and verify the reduction. SVP must also maintain full control over the contracted flexibility.