Google, Nvidia, and Emerald AI have founded the AI Energy Management Alliance (AEMA). The organization aims to make it easier for AI data centers to adjust their power consumption to the available capacity on the power grid. This would allow new data centers to be connected more quickly.
The rapid growth of AI infrastructure is driving a sharp increase in electricity demand. At the same time, the computing infrastructure in data centers offers opportunities to manage that demand more flexibly. AEMA aims to develop agreements that allow grid operators to factor this flexibility into their assessments of new connections.
For example, a data center can reschedule certain computing tasks for a later time when the power grid is under heavy load. Battery storage or local energy generation can also be used to temporarily reduce the amount of power drawn from the grid. In some situations, batteries can even feed stored electricity back into the grid.
Faster connection
According to AEMA, this could have implications for how new data centers are connected. Existing procedures generally assume a relatively constant electricity demand. If a data center can demonstrably reduce its power draw as soon as the grid requires it, available grid capacity could be used more efficiently.
The alliance therefore aims to develop technical requirements covering, among other things, the speed at which data centers can adjust their power demand, how long they can sustain that adjustment, and how predictable that flexibility is. Agreements must also be made in advance regarding behavior during outages and other exceptional circumstances on the grid.
AEMA also advocates for faster connection procedures for data centers that can provide verifiable guarantees regarding their flexible power usage. The costs of a connection should better reflect the actual load on the power grid. Any benefits, such as avoiding or delaying grid upgrades, can also be factored into this.
Google already has more than 1 GW of flexible capacity
The three founders are themselves already working on technology for flexible energy use in data centers. Google reported earlier this year that it now has more than 1 GW of so-called demand response capacity. This allows the company to adjust the power consumption of its data centers to the conditions on the power grid.
To this end, Nvidia has developed DSX Flex, among other solutions. This software is designed for data centers with Rubin GPUs and can coordinate the use of grid power, local energy generation, and batteries. It can also take signals from energy providers into account.
Emerald AI and Nvidia are also working on a 100 MW data center in Manassas, Virginia. There, they plan to deploy Emerald Conductor, software that enables server clusters to reduce their energy consumption when the power grid is under strain. The goal is to minimize the impact on running AI workloads, reports SiliconANGLE.
Eighteen other participants
In addition to Google, Nvidia, and Emerald AI, AEMA is supported at launch by eighteen other companies. These include AI firm Anthropic, chipmaker Analog Devices, and various energy companies.
One of the alliance’s first tasks will be to develop best practices for flexible data centers. These should, among other things, define how facilities respond to power supply issues. The goal is for such guidelines to be incorporated into the design of new data centers from the outset.
In addition, AEMA aims to standardize measurement methods that will allow operators to determine how effectively they are adjusting their energy consumption. The organization also seeks to encourage the sharing of technical and operational data.
The idea behind the alliance is that data centers do not have to be solely a fixed load on the power grid. By coordinating computing, storage, and, where possible, local energy generation, part of their electricity demand can become adjustable. Whether this actually leads to faster connections and fewer grid expansions will depend in part on the agreements AEMA is able to develop with energy companies and grid operators.