- AI workloads generate variable power demand, straining the power grid with sudden peaks
- Peak shaving is one way data centers can help smooth out demand levels for utilities
- Flex is one vendor peddling peak shaving tech, but there could soon be others
Data centers are facing increased demand for artificial intelligence (AI) compute power. The challenge is that AI workloads tend to be spikey in nature, wreaking havoc on the utility grid. With AI forecast to account for more and more workloads over the next five years, finding ways to smooth out AI’s peaks and valleys is a priority.
“Utilities hate [AI's spikey] load profile and don’t really want to support that load profile because they can’t move their power generation up and down to match what a data center service provider needs or what someone with an AI workload needs,” Dell’Oro Group’s Lucas Beran told Fierce. “So, they just have to generate power to support that peak.”
“What becomes really important is not just to optimize the consumption of, but to deal with the volatility of, power that’s driven by this back and forth between inference and training and the spikes that go along with how AI is utilized for computational requirements,” Michael Hartung, CCO at data center supplier Flex told Fierce.
Enter peak shaving, a new technique being used by data centers to smooth out demand levels. In a nutshell, the idea is to use battery power stored on-site at data center campuses to absorb spikes in energy demand rather than passing that demand along to the electrical grid.
According to Beran, most data centers have plenty of battery and generator power available on standby as backup power in case the electrical grid goes down. With peak shaving, data centers can just put those existing assets to more active use. Not only can this help the utility grid handle AI demand better, but it also provides good PR for data centers facing backlash focused on power consumption.
While there aren’t too many vendors out there actively peddling peak shaving solutions, Flex recently ramped up its efforts in this area. Last month, the company teamed with Musashi Energy Solutions to deliver Capacitor-based Energy Storage Systems (CESS) for data centers. These, Hartung said, are battery and switch solutions specifically designed to mitigate the demand surges that GPU compute generates while processing AI workloads.
A company called Enchanted Rock is doing something similar, albeit using natural gas-powered microgrids that can be built on data center campuses.
Beran noted that peak shaving isn’t necessarily common today, it is “trending towards being common.” You better believe we’ll be watching this space.