AI model training and serving require vast quantities of power, but not necessarily all at once. With the first of several gigawatt-scale datacenters due to come online next year, Meta is looking at ways to offload excess energy capacity by selling it on the wholesale market.

This week, Atem Energy LLC, a subsidiary of Zuckercorp, filed an application with the Federal Energy Regulatory Commission requesting permission to sell energy, capacity, and other ancillary services at market-based rates.

Access to adequate power has become a major bottleneck for AI infrastructure providers, many of which have been forced to commission large natural gas generator plants or even help reignite dormant nuclear reactors to support their datacenter buildouts.

According to the filing, Meta’s Atem division hopes to begin selling excess capacity as soon as November 16.

In theory, Atem should allow Meta to lock in large power commitments under favorable terms and then resell any excess capacity it doesn’t need or can’t yet fully utilize, allowing it to get ahead of the company’s datacenter power demands.

Meta’s power demands, we’ll note, are substantial. Back in July, Zuck revealed plans to build several multi-gigawatt datacenters with the first, dubbed Prometheus, slated to come online in 2026.

“We’re building multiple more titan clusters as well. Just one of these covers a significant part of the footprint of Manhattan,” Zuckerberg touted in a July Facebook post.

Meta’s Hyperion datacenter campus, located in Richland Parish, Louisiana,  meanwhile, is expected to scale up to five gigawatts of capacity over the next few years. To power it, Meta has already commissioned Entergy to build three combined-cycle combustion turbine plants totaling 2.26 gigawatts of capacity.

In addition to gas, it also issued a request for proposals to supply the company’s datacenters with between 1-4 gigawatts of nuclear energy by early 2030s.  

And it’s not just Meta either. Earlier this year, investment bank Goldman Sachs predicted global datacenter power demand would more than double from about 55 gigawatts by the end of the decade, with AI being the main driver. ®