Tech and Energy Firms Launch Alliance to Make AI Data Centers Grid-Flexible

Emerald AI, Google and NVIDIA on Wednesday launched the AI Energy Management Alliance, a coalition aimed at making AI data centers more flexible electricity users at a moment when the industry’s rapid expansion is colliding with limits on grid access.

The new group, called AEMA, says it will push for data centers that can reduce or shift electricity use when power systems are under stress, instead of operating as fixed, always-on loads. The pitch is straightforward: if large AI facilities can become more responsive to grid conditions, they may be able to connect faster without adding as much strain to reliability.

That matters because AI data centers are becoming unusually large power users, and developers are increasingly running into interconnection queues, grid upgrade delays and what the industry calls “speed-to-power” problems — the challenge of getting big new loads connected quickly enough to meet demand. AEMA says flexible demand can help address those bottlenecks while supporting grid operations.

The alliance said it will focus on advancing data centers that can dynamically manage electricity use in response to grid conditions, while promoting performance-based, technology-neutral requirements for facilities that commit to that flexibility. It also plans to work on standardizing measurable performance metrics and data sharing, including response speed, duration, predictability and how facilities behave during curtailments, contingencies and other grid events.

Launch partners span both the AI and power industries. In addition to Emerald AI, Google and NVIDIA, NVIDIA’s press materials describe collaboration with energy companies including AES, Constellation, Invenergy and Vistra, among others, to advance what it calls “power-flexible AI factories.”

The companies are also pairing the policy push with a technology blueprint. NVIDIA identified its Vera Rubin DSX reference architecture and DSX Flex software as part of the reference design, while Emerald AI said its Conductor platform is meant to coordinate computing workloads with on-site generation and battery storage. In plain terms, the systems are intended to help data centers dial power use up or down in a controlled way without simply shutting off operations.

To support the case, the companies point to a field demonstration in Phoenix under the Electric Power Research Institute’s DCFlex initiative, involving Oracle Cloud Infrastructure and Salt River Project, the Arizona utility. In that test, a production AI cluster cut power use by about 25% for three hours during a grid-stress event while maintaining acceptable AI performance, according to the companies. The results were documented in a peer-reviewed paper in Nature Energy. NVIDIA and Emerald AI also say they have run demonstrations across five commercial data centers.

AEMA’s website says flexible data centers could unlock up to about 100 gigawatts of existing grid capacity, citing a Duke University Nicholas Institute study. The alliance presents that figure as an estimate of the potential scale if large loads can be made more responsive, not as an established outcome.

The launch also lands as federal regulators are paying closer attention to how large new loads connect to the grid. On June 18, the Federal Energy Regulatory Commission, which oversees interstate electricity markets and transmission, issued Section 206 show-cause orders to the six jurisdictional regional transmission organizations and independent system operators, directing them to justify or reform their interconnection and large-load tariffs. That put data-center growth and the rules for connecting major new electricity users squarely on the policy agenda.

“AI factories are the engines of the intelligence era, and like any great engine, every system must be designed together - energy, compute, networking and cooling as one architecture,” Jensen Huang, founder and CEO of NVIDIA, said in the company’s press materials.

Varun Sivaram, founder and CEO of Emerald AI, framed the issue more bluntly in those materials: “AI factories are too valuable to be treated as either passive loads or permanent islands.”