A coalition founded by Nvidia, Google and grid-software startup Emerald AI launched Wednesday with a pitch that sounds simple and is anything but: make AI data centers flexible power customers that shrink their demand when the grid is stressed, instead of adding to it.
The AI Energy Management Alliance, or AEMA, counts Anthropic, Analog Devices, National Grid, AES, Constellation, NRG and RWE among its 18 launch partners, alongside the three founders. It spans the whole stack, from the chipmakers whose hardware sets data-center power profiles to the utilities that have to serve them. Veteran retail energy executive Frank Lacey will serve as executive director.
The problem the alliance is naming
Data centers have become the most contentious load growth in the US power system. Utilities warn that large AI campuses drive up costs for other ratepayers, require expensive grid upgrades, and in some markets strain local opposition into outright bans. Polling released alongside the launch shows the public-opinion challenge getting steeper: 76% of Republicans support requiring data-center developers to pay for grid upgrades, according to Axios reporting on the coalition.
AEMA’s answer is demand flexibility. Instead of treating a data center as a fixed, always-on load, the coalition wants operators to shift workloads in time, discharge onsite storage, and use paired generation during peak grid stress. The argument is that the same megawatt, scheduled intelligently, costs the grid less and upsets neighbors less. Josh Parker, head of sustainability at Nvidia, called it “a win-win-win solution” on a call with reporters before the launch.
“By championing policies that recognize and value flexible AI demand, AEMA supports the nation’s competitiveness in AI innovation and deployment while strengthening grid reliability and protecting energy affordability for ratepayers,” the alliance said in its launch statement.
The technology behind the pitch
The alliance pairs Nvidia’s Vera Rubin DSX Flex platform with Emerald AI’s Emerald Conductor software. DSX is Nvidia’s reference design for AI factories, covering site, power, cooling, computing and facility decisions. Emerald Conductor connects utilities and data centers in real time, letting facilities respond to grid signals and orchestrating compute alongside onsite energy resources.
The technical claim is that AI training and inference workloads have slack that traditional data-center loads do not. A batch training job can pause or migrate. Inference latency tolerances vary by application. If even part of that slack is harvestable, a gigawatt-scale campus becomes a grid asset rather than a grid liability, and the costly transmission upgrades that dominate interconnection queues become less necessary.
Emerald AI raised a $90 million Series A in July and has positioned itself as the software layer between hyperscalers and utilities. Its platform verifies that a data center actually delivered the flexibility it promised, which is the piece regulators and grid operators will care about. Measurement is where demand-response programs have historically lived or died, and the company’s entire product is built around that audit trail.
| Role | Members |
|---|---|
| Founders | Nvidia, Google, Emerald AI |
| AI and semiconductors | Anthropic, Analog Devices |
| Utilities and power | National Grid, AES, Constellation, NRG, RWE |
| Executive director | Frank Lacey |
Why now
The regulatory window is open. In June, federal regulators directed regional grid operators to examine new options for connecting large, flexible power users, an opening that AEMA’s approach is designed to fill. Meanwhile the interconnection backlog keeps growing, and data-center power demand forecasts have become the dominant variable in utility resource planning across several regions.
The coalition also arrives amid a pricing squeeze. Nvidia, Google and Anthropic are all racing to secure power for AI expansion, and each has made infrastructure bets this year: Nvidia took stakes in SB Energy and Cloverleaf Infrastructure, and the industry’s land-and-power race has become as competitive as the model race. Making existing grid connections go further is cheaper than building new ones, and it is faster, which matters when campuses are planned in gigawatts.
For utilities, the pitch is ratepayer protection. If AI campuses flex rather than sit flat, the upgrades that get socialized to all customers shrink. For the AI industry, the pitch is political: a data center that visibly backs off during a heat wave is a much easier neighbor than one that does not, and the backlash against data centers has become a real constraint on siting. Communities from Virginia to Wisconsin have blocked or delayed projects over power and water concerns.
What to watch
Three things will tell whether the alliance is more than a press release. First, whether grid operators adopt tariff structures that actually pay for flexibility, since voluntary goodwill does not scale. Second, whether hyperscalers accept the latency and scheduling tradeoffs in production workloads rather than only in demonstration projects. Third, whether utilities like Constellation and NRG, which serve both sides of the market, push the model into their interconnection agreements.
The alliance says it will promote accelerated interconnection for flexible, grid-enhancing data centers using colocated generation, storage, computational flexibility and other demand-management approaches. That is a policy agenda as much as a technology one, and the founding roster, which includes both the companies building AI and the companies powering it, is designed to give it weight in regulatory proceedings.
The skeptical reading is also available. Every member of this coalition has an interest in more data centers being approved, and flexibility is a cheaper concession than paying for grid upgrades outright. The polling cited at launch suggests the public wants developers to bear upgrade costs, not just schedule around peaks. If AEMA’s flexibility model becomes a substitute for direct payment rather than a complement to it, the political problem returns in a different form.
Still, the composition of the launch list is the news. Getting Nvidia, Google, Anthropic and five major power companies to sign the same statement about demand flexibility is the kind of alignment that was unlikely a year ago, and it signals that the industry treats grid access, not chips, as the binding constraint on AI growth.