High-tech companies, data center developers and utilities are revealing the outlook for US power outrageous demand in the face of unprecedented growth in AI.
Among all hand squeezes, a new paper published this week suggests that the situation may not be so disastrous if data center operators and other heavy power users have slightly reduced their use. It’s there.
By limiting the power drawn from the grid to a maximum of 90% for a few hours at a time, new users can unlock 76 gigawatts of capacity in the US – a total of about one day per year. According to Goldman Sachs, this is more than all data centers use globally. To put that number in perspective, it is about 10% of the US peak demand
If data centers reduce more usage, they can gradually unlock more capacity.
Such programs are nothing new.
For decades, utilities have encouraged large power users, such as shopping malls, universities and factories, to reduce their use during peak demand, like on hot summer days. These users could knock down the air conditioner or turn off the thirsty machine for a few hours, and in return the utility would give credit to their bills.
Data centers are primarily on the sidelines and instead choose to maintain customer uptime and performance levels. This study argues that data centers may be flexible and therefore could be ideal demand response participants.
There are several ways data centers can trim their power usage, the research says. One is temporary flexibility, or shifting computing tasks to low demand times. For example, AI model training can be easily rescheduled to accommodate simple reductions.
The other is spatial flexibility, in which companies migrate computational tasks to other regions that are not experiencing high demand. Datacenters can also allow operators to consolidate loads and shut down parts of the server.
Additionally, if the task is mission critical and is unable to perform delays or shifts, data center operators can always rely on alternative power sources to compensate for the reduction. Batteries are ideally suited to this as even moderately sized installations can provide nearly instantaneous power for several hours.
Some companies have already joined these ad hoc versions.
Google has used a carbon-aware computing platform originally developed to trim emissions to enable it to meet demand. Enel X worked with the data center to stabilize the grid by leveraging a battery from a discrete power source (UPS). PG&E also offers to quickly connect the data center to the grid if the operator agrees to participate in a demand response program.
These tweaks do not completely eliminate the need for new power sources. But they could turn potentially catastrophic situations (a lack of all new AI servers) into something that can be more easily resolved.