Key takeaways

  • Google, NVIDIA and Emerald AI launched the AI Energy Management Alliance (AEMA) on September 16, 2026.
  • The group is proposing measurable rules for flexible data centers; it has not announced a certification program or a guaranteed fast-track interconnection process.
  • Google says it has signed 1 GW of demand-response capacity into long-term utility contracts in the United States.
  • AEMA's headline 100 GW opportunity comes from a Duke study that modeled 98 GW of new load at 0.5% average annual curtailment. It is a national scenario, not installed or reserved capacity.
  • NVIDIA says Emerald AI's software has operated at commercial scale, while the first dedicated DSX Flex deployment is planned as a 96 MW facility in Manassas, Virginia.

Google, NVIDIA and Emerald AI launched the AI Energy Management Alliance on September 16. Its goal is to make grid flexibility a defined part of how large AI data centers connect: a facility could reduce grid draw by shifting lower-priority computing, using storage or pairing its load with on-site generation when the power system is stressed.

That is narrower than the draft's original promise. AEMA has published policy principles and performance goals, not a certification scheme, an entitlement to faster interconnection or evidence that 100 GW is already available to projects.

What AEMA is proposing

The alliance says flexible facilities should be judged by the service they can actually deliver rather than by a prescribed technology. Its initial framework calls for utilities and developers to define several obligations before connection:

  • how quickly and for how long a site can reduce grid demand;
  • how predictable that response is during an emergency;
  • ride-through, curtailment and contingency behavior;
  • common technical requirements, performance metrics and operational data; and
  • interconnection costs that reflect a project's system impacts and benefits.

The three founders are joined by 18 launch partners across AI, power production, utilities and grid software. The list includes Anthropic, Analog Devices, AES, Calibrant Energy, Camus, ClearPath, Constellation, Encoord, Fluence, Generate Capital, GridUnity, National Grid, NRG, PassKey, RWE, Splight, Verrus and Voltus.

Those members give AEMA a broad policy constituency. They do not, by themselves, prove that a common operating standard has been adopted by a utility or grid operator.

What is operating today

Google offers the clearest disclosed commercial number. In March, it said it had integrated 1 GW of demand response into long-term contracts with multiple U.S. utilities. The contracts include Entergy Arkansas, Minnesota Power and DTE Energy, following earlier agreements with Indiana Michigan Power and the Tennessee Valley Authority. Google says the capability can limit or shift some machine-learning workloads during selected hours.

That 1 GW is contracted demand-response capacity, not 1 GW of continuous load reduction. How often it can be called, for how long and at which sites depends on the individual utility agreements.

NVIDIA has also described a smaller operating example. Its Eos AI factory in Santa Clara uses Emerald Conductor in Silicon Valley Power's Flexible Load Interconnect Program. NVIDIA says more than 200 utility signals were met and describes one event in which the site's draw fell from 4 MW to 3 MW while priority work continued. The company is explicit that this is not yet a DSX Flex installation, but an earlier commercial-scale demonstration of the concept.

NVIDIA says the first dedicated DSX Flex commercial deployment is planned for its AI Factory Research Center in Manassas, Virginia: a 96 MW Vera Rubin facility. That corrects two easy-to-repeat overstatements—the project is planned, and NVIDIA calls it one of the world's first power-flexible AI factories rather than an uncontested world first.

Where the 100 GW claim comes from

AEMA markets flexibility as a way to tap roughly 100 GW of grid headroom. The underlying Duke University study is more precise. Across 22 balancing authorities representing 95% of U.S. peak load, researchers estimated that the system could integrate:

  • 76 GW of new load with average annual curtailment capped at 0.25%;
  • 98 GW with curtailment capped at 0.5%; and
  • 126 GW with curtailment capped at 1.0%.

The result is a first-order national estimate of "curtailment-enabled headroom." It is not a project queue, a transmission study or a promise that a particular campus can connect. Local generation, transmission and distribution constraints still decide what is feasible.

The study also assumes that planners can rely on the promised flexibility. That makes measurement, dispatch tests and enforceable operating limits central to the case—not paperwork after a site is energized.

The affordability claim needs context

AEMA also cites a Brattle Group analysis in which a 10% improvement in annual system utilization reduced average rates by 3.4% from current conditions. That result came from an illustrative 3,000 MW utility adding 1,000 MW of load under the report's assumptions. Brattle says the exercise is a proof of concept, not a forecast, and that jurisdiction-specific analysis is required.

The useful conclusion is not that every flexible data center cuts rates by 3.4%. It is that new load can spread fixed grid costs across more electricity sales if it connects where spare capacity exists, avoids the system peak and costs less to accommodate than conventional upgrades.

What to watch next

AEMA's launch matters because the bottleneck is moving from whether AI workloads can flex to whether utilities can contract for, measure and rely on that behavior. The next evidence should be concrete: utility tariffs or contracts, operating envelopes, dispatch performance and interconnection decisions tied to verified flexibility.

Until those appear, the alliance's 100 GW number should be read as modeled national potential. Google's 1 GW contract portfolio and Emerald AI's smaller demonstrations show that the mechanism is real; they do not yet show that it can be standardized across markets. Our earlier look at OpenAI and Google demand-response projects covers two adjacent commercial approaches.

Sources: AEMA, NVIDIA, Google, Duke University Nicholas Institute and The Brattle Group.

Sources

This article was researched and fact-checked against the following sources: