Our Take
Flexibility is real and grid operators want it, but data centers are only testing it now because building new power plants takes eight years—this solves a timeline problem, not an energy problem.
Why it matters
Governments are blocking data center projects over power demand; a working flex model could unblock billions in stalled infrastructure approvals. But it requires data centers to accept lower compute during peak hours, which remains unproven at scale.
Do this week
Grid operators: audit your peak-demand hours (typically 22 hours per year per Duke study) and identify which data center workloads can defer non-critical tasks during those windows before 2026 interconnection deadlines arrive.
Emerald AI is deploying its Conductor software in Virginia this year
Conductor is designed to reduce data center power draw when the grid is under stress, while keeping critical servers running. The company simulated the system in December 2025 using demand from a 2020 UK soccer match—a moment when millions of kettles caused a sudden electricity spike. Had Conductor been live, it would have automatically throttled non-essential chips at the moment National Grid needed the headroom.
The Virginia facility, called a "power-flexible AI factory," is a partnership between Emerald AI, Nvidia, and Digital Realty. It will connect to the live PJM grid, the largest in the US. The test matters because PJM takes eight years to bring new generation online (per RMI, an energy research group), making grid approval the actual bottleneck—not construction.
Two academic studies support the feasibility. Duke University researchers found the US grid could offer 76 gigawatts of additional capacity (about 5% of total US capacity) to data centers willing to reduce usage just 0.25% of the time, or roughly 22 hours per year (published February 2025). Princeton and grid-modernization companies, in a Google-funded study, found that a 500-megawatt facility capable of flexing less than 1% of the year could reach full operation three to five years faster than an inflexible one.
The real problem is political, not physical
Data center operators have stalled over $150 billion in projects in 2025 (per Data Center Watch), and more than a dozen states are considering bans. Local moratoriums are active in Minneapolis and DeKalb County, Georgia. The GRID Act, a bipartisan Senate bill, would sever new data centers from public grids entirely.
Flex capacity addresses two separate grievances. First, it shortens approval timelines by using existing infrastructure instead of waiting for new power plants. Second, it theoretically reduces pressure on local electricity prices and avoids stranding power from renewables during low-demand hours. A Duke study found flexibility could reduce rates by 0.5% to 2.8%.
But this assumes data centers will accept the trade: faster grid interconnection in exchange for accepting power limits during peak hours. Hyperscalers like Microsoft, Oracle, and xAI are currently choosing the opposite path, rolling out off-grid natural-gas generation. Emerald's bet is that the value of speed outweighs the cost of lost compute.
The parallel problem is that grids operate at roughly 30% utilization most of the year, according to a 2025 Stanford study of transmission lines in western North America. Flexibility exploits that headroom without requiring infrastructure upgrades.
Three mechanisms exist for data centers to flex
On-site backup power or generation is the simplest but most expensive option. The facility absorbs the full cost of hardware that sits idle most of the year.
A second path uses a Virtual Power Plant (VPP), where the utility throttles power to willing residential or commercial customers (often with smart thermostats, solar, or batteries) and the data center compensates them. VPPs operate in nearly 40 states, but rules vary widely and their capacity is limited by adoption.
The third option is direct demand reduction: the data center simply uses less power at peak times. Conventional wisdom says operators won't accept this. But Emerald's chief scientist, Ayse Coskun, reports "a clear and growing trend" of operators willing to trade flexibility for faster grid interconnection.
For grid operators, the payoff is control. Instead of managing spikes reactively, they can orchestrate demand across an ensemble of flexible loads, making the system more resilient to renewable variability and unexpected outages.