Emerald AI will deploy its Conductor software this year at a Digital Realty facility in Virginia's Data Center Alley, throttling chip power in real time when the regional grid is under stress. The Washington-based firm, working with Nvidia and Digital Realty, is pitching the site as one of the first power-flexible AI factories on a live grid. The bet: data centers can plug into existing infrastructure years sooner if they agree to dial back during the handful of hours each year when demand peaks.
The numbers behind the pitch are striking. A 2025 Duke University study found US grids could accommodate an additional 76 gigawatts of load — about 5% of total US grid capacity, and roughly enough to cover projected data-center growth through 2030 — if facilities agreed to curtail usage just 0.25% of the time, or about 22 hours a year. A separate Princeton analysis funded by Google found a 500-megawatt facility willing to flex for under 1% of the year could reach full operation three to five years faster than an inflexible one in the PJM territory.
Speed is the core problem Conductor is designed to solve. PJM, the largest US grid operator, currently needs eight years to bring new generation online, according to RMI. Data centers themselves can be built far faster, creating a queue of stalled interconnect requests across the country. Emerald's argument is that flexibility — software-mediated load shedding tied to grid signals — sidesteps the wait entirely by working within the headroom that already exists.
Key facts
- 01Duke University researchers found US grids could absorb 76 GW of new load — roughly 5% of total capacity — if data centers reduce usage just 0.25% of the year, about 22 hours.
- 02A Princeton study funded by Google found a 500 MW facility willing to flex for under 1% of the year could reach full operation three to five years faster.
- 03Emerald AI will deploy its Conductor software this year in Virginia's Data Center Alley with Nvidia and Digital Realty as partners.
- 04Local opposition stalled over $150 billion in data-center projects in 2025, per Data Center Watch, with PJM needing eight years to bring new generation online.
- 05US electricity demand is projected to rise 25% by 2030 compared with 2023 levels, with data centers a major driver.
Emerald demonstrated Conductor in December 2025 by simulating the UK grid's response to a 2020 Euro tournament match between England and Germany, when millions of viewers flipped on electric kettles at halftime. In the simulation, Conductor instructed a London data center to slow its power-hungry chips at the exact moment National Grid faced the demand spike. The Virginia deployment will be the first time the software runs against live grid conditions rather than a recreated load profile.
The political backdrop is unfriendly. Data Center Watch tallied over $150 billion in stalled data-center projects in 2025, with more than a dozen states considering bans and local moratoriums already in effect in Minneapolis and DeKalb County, Georgia. The GRID Act, a bipartisan bill in the US Senate, proposes to sever new data centers from public grids entirely. Hyperscalers including Microsoft and Oracle have responded by proposing off-grid natural gas plants, and xAI's Colossus site outside Memphis brought gas turbines in on flatbed trucks — a buildout now drawing regulatory and resident pushback over emissions.
The underlying inefficiency is real. US grids are built to meet peak demand even when peaks last only a few hours a year, leaving substantial unused capacity the rest of the time. A 2025 Stanford study of western North America transmission lines pegged average utilization near 30%.
Demand response — utilities calling industrial customers to power down during heat waves — has existed for decades but is slow and hard to scale. Virtual power plants emerged in the 2000s with smarter, more granular load control, often paying participants for adjustments they barely notice. Conductor extends that logic into the AI data-center stack: when the grid signals stress, the software preserves the most time-sensitive workloads and defers the rest.
Skeptics argue flexibility distracts from the harder task of actually building more transmission and generation, and that leaning on demand-side curtailment could leave the grid more vulnerable during sustained stress events. Utilities and grid operators, which tend toward operational conservatism, also have to rewrite long-standing practices to integrate flexible loads at scale. And for data-center operators bound by customer SLAs, compromising on power draw — even briefly — is a hard sell internally.
US electricity demand is projected to rise 25% by 2030 compared with 2023 levels, driven by data centers alongside EVs and electrification. That trajectory means flexibility and new buildout are not really alternatives; the grid will need both. The question is which one unlocks AI capacity faster in the next 24 months, and on that horizon, software has a structural advantage over concrete.
If Conductor performs in Virginia the way it did in the London simulation, the economics shift quickly. Saving three to five years on a 500 MW interconnect is worth more than the marginal compute lost to 22 hours of annual curtailment, and Nvidia's involvement signals that the chip vendor sees flex-aware deployment as a way to keep selling GPUs into a constrained grid. The data-center industry has spent two years insisting the only answer to AI's power problem is more power. Emerald is making the case that the cheaper answer, at least for now, is smarter power.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
