Google, Nvidia, and startup Emerald AI have formed a coalition proposing a new approach to connecting data centers to the power system. Instead of consuming a fixed amount of electricity around the clock, AI data centers could become flexible consumers: during power shortages, they would temporarily reduce their load, shift computing tasks, or use stored energy.

Supporters of the initiative hope to simultaneously reduce strain on power grids, curb rising electricity costs, and make it easier to build new data centers.

Data centers are being recast as “good citizens of the power system”

The new organization is called the AI Energy Management Alliance (AEMA). It includes 20 companies and organizations from the artificial intelligence and energy sectors. Members include Anthropic, energy companies National Grid and AES, as well as power producers Constellation, NRG, and RWE.

Formally, this is not an entirely new organization. Its predecessor, the Advanced Energy Management Alliance, was created back in 2014 to promote electricity demand-management technologies, but over time it all but ceased operations.

Now the initiative has been relaunched with a much narrower goal: to ensure that AI data centers can flexibly adjust their energy consumption, and that regulators and utilities take this into account when connecting them to the grid.

How a data center could reduce its load

Right now, a large data center is typically viewed by the power system as a massive consumer that requires stable power around the clock. But AI workloads in some cases make it possible to change operating modes.

When the power system is short on capacity, a data center can temporarily cut consumption by shifting or pausing part of its computing tasks. Another option is to use battery storage systems or local generation.

This does not necessarily mean completely shutting down the data center. Its load can vary depending on the state of the power grid.

Google has already signed agreements with energy companies allowing it to reduce up to 1 GW of consumption across different regions of the United States when necessary, according to Tyler Norris, the company’s head of energy market innovation.

In exchange for flexibility, data centers want faster grid connections

The initiative’s backers believe this approach could also change the process of connecting new facilities to the power grid itself.

Usually, a large data center requires a substantial increase in available capacity, which can mean building new transmission lines, substations, and other equipment. If a facility can reliably reduce consumption during periods of shortage, the grid requirements may be lower.

That is why AEMA wants to persuade regulators and grid operators to create a kind of “fast lane” for flexible consumers: data centers capable of controllably reducing their load could connect more quickly and in some cases require fewer costly grid upgrades.

In June, U.S. federal regulators had already instructed regional operators to study new options for connecting large consumers that can flexibly manage their electricity demand. This creates room for AEMA’s proposed approach.

Against a backdrop of growing dissatisfaction with data centers

The initiative comes at a time when the buildout of AI infrastructure is facing growing criticism over its resource consumption.

According to a new survey by AP-NORC and the University of Chicago Energy Policy Institute, 84% of Americans are concerned about the impact of data centers on local electricity prices. More than half of respondents also said they were highly or very highly concerned about the environmental consequences of AI development — compared with 41% a year earlier.

At the same time, about 79% of Democrats and 76% of Republicans support requiring data center developers to pay for the grid upgrades needed to connect their facilities.

For the industry, this means that energy consumption is becoming not only a technical issue, but also an economic and social one.

The key question is whether the promised flexibility can be trusted

The new model has an obvious condition: a data center must actually reduce its load precisely when the power system needs it.

Emerald AI CEO Varun Sivaram believes facilities should receive connection advantages only if their ability to reduce consumption can be verified and enforced.

Otherwise, there is a risk that a data center will receive more favorable connection terms but fail to provide the promised flexibility at a critical moment.

So the coalition must now convince not only technology companies, but also regulators, grid operators, and local communities that managed consumption really can offset part of the burden created by new AI facilities.

If this approach becomes widespread, data centers may gradually stop being seen by the power system solely as large and constantly growing consumers. Instead, they could also serve as a controllable resource that temporarily reduces its own consumption during periods of capacity shortage.