US-based AI software company Empromptu AI has launched a new platform designed to reduce rapid fluctuations in power demand caused by GPU workloads in AI data centers.

Dubbed Grid Guard, the software aims to predict and smooth spikes in electricity demand by coordinating when GPU clusters execute compute-intensive tasks. The company claims that rather than permitting thousands of GPUs to begin processing simultaneously, the platform staggers workloads by between 50 and 200 milliseconds to reduce sudden changes in load.

According to Empromptu, the launch was in response to the growing attention on power quality within AI data centers, where increasingly dense GPU deployments can create highly variable electrical demand.

Empromptu said early deployments of Grid Guard have reduced power volatility by an average of 80 percent without significantly affecting AI workload performance. The company did not disclose where the software has been deployed or provide independent validation of the results.

According to Empromptu, the platform combines four functions. These include forecasting short-term changes in GPU power demand, scheduling workloads to reduce synchronized load spikes, sending demand signals to power infrastructure such as batteries and generators, and analyzing the cost of different workload patterns.

The company argues that managing power demand in software could reduce the need for additional electrical infrastructure deployed to absorb short-duration load spikes.

"AI data centers are being built at a pace the power grid was never designed to support," said Shanea Leven, CEO and co-founder of Empromptu. "Grid Guard is how we make that buildout sustainable, by making the software smarter about when and how workloads fire."

Grid Guard is the first data center-focused product launched by Empromptu. The company, based out of San Francisco, said the software is available immediately.