Emerald AI's Conductor: flexible data centers to ease grid strain
Emerald AI will deploy Conductor this year in Virginia to throttle facility power during grid spikes while keeping urgent server work.
TL;DR
- 01Emerald AI will deploy Conductor this year in Virginia to throttle facility power during grid spikes while keeping urgent server work.
- 02Conductor is designed to let a data center lower its immediate electricity consumption when the local grid needs relief, while prioritizing the timeliest and most important computing jobs.
- 03In Emerald’s simulation, throttling chips prevented a potential supply shortfall.
Emerald AI will deploy its Conductor software this year in a Virginia data center in Data Center Alley, using real-time controls to reduce facility power draw during grid stress while preserving the most urgent server tasks. The company ran a December 2025 simulation that recreated the electricity demand spike from a 2020 Euro soccer match, during which an AI program instructed a London data center to slow some high-power chips to help avoid blackouts.
How Conductor and flexible connections work
Conductor is designed to let a data center lower its immediate electricity consumption when the local grid needs relief, while prioritizing the timeliest and most important computing jobs. In Emerald’s simulation, throttling chips prevented a potential supply shortfall. Emerald’s announced deployment will connect Conductor to the live grid in Virginia and operate with partners that include Nvidia and Digital Realty.
Facilities that accept flexible connections have three main options when grid power is throttled, experts in the piece describe. One is on-site backup power or generation, paid for and run by the data center. A second is joining a virtual power plant, or VPP, where utilities coordinate reductions from many enrolled customers and the data center pays those participants for their flexibility. The third is simply running with lower power at peak times, shifting or delaying nonurgent workloads.
The article notes utilities already use a slow, manual version of this idea called demand response, where large customers are asked to curtail usage during heat waves or other peaks. Conductor and other modern systems aim to automate and refine that practice, using faster, more granular controls and digital models.
Prior context: why operators are looking for flexibility
Grid constraints are a core barrier to bringing new data centers online. The grid operator PJM, which covers Virginia and is the largest in the US, needs eight years to bring new generation online, according to RMI. Developers face local pushback and regulatory limits: organizers stalled over $150 billion worth of projects in 2025, according to Data Center Watch, and more than a dozen states are considering bans. Local moratoriums are in place in Minneapolis and DeKalb County in Georgia. At the federal level, the GRID Act would propose separating new data centers from public grids.
Several studies cited show flexibility could free substantial capacity. A February 2025 Duke University report found the US grid could offer an additional 76 gigawatts, about 5% of total capacity, to facilities that reduce draw just 0.25% of the time, roughly 22 hours a year. A Princeton-led study funded by Google examined PJM locations and concluded a 500-megawatt facility capable of flexing for less than 1% of the year could reach full operation three to five years faster than an inflexible site. The Duke report also estimated flexibility could lower electricity rates by 0.5% to 2.8%.
Those numbers matter because analysts expect the combination of data centers, electric vehicles, and air-conditioning to drive about a 25% increase in US electricity demand by 2030 compared with 2023 levels.
Deployment and industry response
Some hyperscalers are choosing less flexible routes, building off-grid generation. The article cites xAI’s Colossus site outside Memphis as an example, where gas turbines were trucked in on flatbeds during buildout; the facility is now operating and facing regulatory and resident pushback. Advocates of flexibility argue that if data centers can cut load for a small number of high-demand hours, they can avoid building or waiting for new generation and instead use existing grid headroom.
Nvidia’s head of sustainability, Josh Parker, framed the trade-off plainly: “We need to solve the energy equation,” he said, arguing that AI-factory flexibility bridges demand for AI and grid limits. GridCare and other startups are building tools that model the grid under many conditions, producing digital twins and recommendations for where flexibility can safely unlock capacity.
Why it matters
If Conductor and similar systems work at scale, they could shorten the time it takes to bring data centers online by avoiding long waits for new power plants. That would shift some of the infrastructure debate away from building more generation and toward smarter use of what already exists. It would also offer a political argument for data centers: reducing peak draws can limit the need for peaker plants and ease local concerns about pollution and high electricity prices.
At the same time, flexibility raises trade-offs. Data centers must accept occasional throttling, utilities must adopt faster operational practices, and some critics worry the approach could distract from necessary grid buildout. The success of Conductor will depend on those operational and political shifts as much as on the software itself.
What to watch
Monitor Emerald’s Virginia deployment this year and whether Conductor is integrated with local VPP programs or on-site backup systems. Also watch PJM and other grid operators for changes to interconnection practices or incentives that reward flexible data-center behavior.
Written by The Brieftide · Source: MIT Technology Review
The Brieftide Daily · 06:00
Briefs like this one, in your inbox every morning.
Continue reading
More in AI InfrastructureIEEE launches virtual training course on large language models
IEEE is offering a virtual training course that teaches engineers to use large language models as reasoning engines in development.
AI4SE and SE4AI: A decade review of AI in systems engineering
H. Sinan Bank, Daniel R. Herber and Thomas Bradley map three research phases and assess 1.
Amazon's AWS may sell Trainium chips to challenge Nvidia
AWS executives say selling Trainium to third parties is possible, with Andy Jassy estimating a potential ~$50 billion annual run rate.
Hyperscalers AI spending to outpace cash flow by Q3 2026
Epoch AI data shows infrastructure spending growing ~70% annually versus operating cash flow at ~23%, with a crossover around Q3 2026.