The Grid Fights Back: Emerald AI, Google and NVIDIA Launch the AI Energy Management Alliance

For two years the AI industry has treated the electric grid as an obstacle to be overcome — something to queue for, to lobby around, to price into the model of infinite scaling. On September 16, 2026, the grid got a formal seat at the table. Emerald AI, Google and NVIDIA announced the launch of the AI Energy Management Alliance (AEMA), a first-of-its-kind coalition that brings the full AI and power value chains into one organization with a single proposition: data centers that can verifiably flex their electricity demand should get connected faster, at greater scale, and on better terms than data centers that cannot.

The founding trio is joined by eighteen launch partners spanning the computing and energy worlds. From the AI side: Anthropic, the frontier lab that has made grid-impacts mitigation a signature issue. From the energy side: utilities and power producers including National Grid, AES, RWE, Constellation, and NRG. The remainder of the roster — Calibrant Energy, Camus, ClearPath, Encoord, Fluence, Generate Capital, GridUnity, PassKey, Splight, Verrus, and Voltus — covers the flexibility technology stack: storage, grid software, financing, and interconnection planning. GridUnity, which makes grid connection and planning software, will sit on the alliance’s board. Veteran energy executive Frank Lacey will serve as AEMA’s executive director.

The Deal Being Offered

AEMA’s core argument is that power — not capital, and no longer chips — is the binding constraint on American AI infrastructure. Interconnection queues for new large loads stretch five to ten years in many markets, a backlog the alliance says flexible AI could help clear. The problem, as the group frames it, is that traditional interconnection processes were designed around facilities with flat, static demand. They were never built for computing infrastructure capable of responding intelligently when the power system is constrained.

The alliance proposes a trade. A flexible data center can adjust its electricity draw from the grid in several ways: shifting or pausing computing workloads, discharging on-site batteries, using paired generation, or responding to system contingencies. In exchange for committing to that behavior — verifiably and enforceably — it should qualify for faster, risk-adjusted interconnection pathways and cost allocations that reflect its actual system impact, rather than the worst-case impact of an inflexible load.

“The most effective way to accelerate American AI is to make every data center a good citizen of the grid,” said Varun Sivaram, CEO of Emerald AI, in the launch announcement. On a call with reporters he was more pointed about the conditions: data centers should get faster connections only if their ability to cut demand is verifiable and enforceable. “Power-flexible data centers are a wonky topic,” he acknowledged — but the underlying politics are not.

Google has already put capacity on the table. Tyler Norris, the company’s head of advanced energy market innovation, said Google has committed 1 gigawatt of power demand it can reduce when needed, through utility agreements across the United States, Axios reported. Josh Parker, NVIDIA’s head of sustainability, framed the alliance’s method as “technology-neutral, performance-based standards” — judging facilities on measurable grid-reliability metrics rather than on the specific hardware or software used.

Four Principles, One Framework

AEMA’s launch principles, detailed in NVIDIA’s announcement, are more concrete than most coalition manifestos:

• Obligations before connection. Ride-through, curtailment, and contingency-response duties would be defined before a facility ever connects — setting clear rules for staying online during brief grid disturbances, reducing power when asked, and responding to emergencies.
• Standardization. Common technical requirements, performance metrics, and operational data sharing, so that a flexibility commitment in one jurisdiction means the same thing in another.
• Faster lanes for credible commitments. Risk-adjusted interconnection pathways for customers whose flexibility is verifiable rather than aspirational.
• Cost allocation by actual impact. Interconnection costs that reflect the system benefits a facility provides — avoided upgrades, improved ramping capability — rather than a flat charge socialized across ratepayers.

The group plans to press this framework with federal and state regulators, regional grid operators, and utilities. The timing is not accidental.

The Political and Economic Backdrop

The launch lands amid escalating public backlash over AI’s electricity footprint. New polling from AP-NORC and the University of Chicago’s Energy Policy Institute released this week found that 84 percent of Americans worry about data centers’ impact on local electricity prices. Around 79 percent of Democrats and 76 percent of Republicans — a rare bipartisan supermajority — support requiring data center developers to pay for the grid upgrades their facilities require.

The economics reinforce the politics. Sivaram wrote in a Fortune commentary accompanying the launch that the U.S. grid runs at only about 50 percent of its capacity on average, and cited a Brattle Group estimate that each 10 percent gain in utilization lowers electricity rates by roughly 3.4 percent. AEMA’s website claims flexible AI could unlock an extra 100 gigawatts from the existing grid — capacity available years before new transmission or generation could be built.

Regulators are already moving in the same direction. In June, the Federal Energy Regulatory Commission directed regional grid operators to develop options for connecting large, flexible power users. Texas, staring at 474 gigawatts of data center interconnection requests, is finalizing rules to let controllable data centers connect sooner. Silicon Valley Power, a California municipal utility, has launched the country’s first flexible-load interconnection program. AEMA’s bet is that this scattered momentum needs a shared technical framework — and a unified industry voice — to become the default rather than the exception.

The Track Record Behind the Alliance

This is not a coalition of promises without precedent. NVIDIA and Emerald AI have completed six demonstrations of flexible data centers around the world, including a March 2026 trial in which Emerald’s Conductor platform responded to grid operator signals with a roughly 30 percent reduction in a cluster’s power draw on demand — without degrading latency-sensitive workloads. Later this year, NVIDIA, Digital Realty and Emerald AI plan to switch on a power-flexible AI factory in Virginia drawing nearly 100 megawatts, tested in collaboration with EPRI, Dominion, and PJM.

There is also an institutional footnote worth noting: AEMA is technically a relaunch. The Advanced Energy Management Alliance, founded in 2014 to work on demand flexibility and electricity policy, had gone largely inactive before being revived for this effort — a sign that the demand-response playbook developed for industrial loads over the past decade is now being aimed squarely at AI.

What to Watch

The caveats deserve attention. “Up to 100 gigawatts” of unlocked capacity is a theoretical ceiling across the entire build-out, not a delivered figure. Deep flexibility demands increasingly sophisticated workload orchestration — training jobs can checkpoint and pause, batch inference can shift by hours, but critical services cannot. And the alliance’s membership, while broad across energy, is notably thin on the AI side beyond Anthropic and Google: OpenAI, Microsoft, Meta, and Amazon are absent from the launch roster, leaving the industry’s largest power consumers uncommitted to the framework — for now.

Still, the direction is clear. The rules governing power for AI are being written now, and AEMA is the most credible attempt yet to write them from inside the industry. If performance-based flexibility becomes the price of fast interconnection, the data center of the near future may be judged less by how much power it consumes than by how intelligently it can yield it.