Unlock 76 GW of US Data Center Power Capacity

AI and Electricity Demand: A Potential Solution

Concerns have been escalating among tech companies, data center developers, and power utilities regarding the rapidly increasing demand for electricity in the U.S., driven by the expansion of artificial intelligence.

A New Perspective on Capacity

However, a recently published study proposes that the situation may be less critical than initially feared, provided data center operators and other substantial electricity consumers implement minor adjustments to their usage.

By voluntarily limiting their power intake from the grid to 90% of their maximum capacity for short periods – totaling approximately one day annually – new users could potentially unlock 76 gigawatts of electrical capacity within the United States.

Significant Capacity Unlocked

This figure, as noted by Goldman Sachs, surpasses the total power consumption of all data centers worldwide. For context, this represents roughly 10% of peak electricity demand across the U.S.

Further reductions in power usage by data centers could progressively release even greater capacity.

Demand Response Programs: A Familiar Approach

These types of programs are not entirely novel.

For decades, utilities have incentivized large electricity users, including shopping malls, universities, and factories, to reduce consumption during peak demand periods, such as hot summer days.

These users might temporarily adjust air conditioning settings or pause energy-intensive machinery, receiving a credit on their electricity bills in return.

Data Centers and Demand Response

Historically, data centers have largely refrained from participating in such programs, prioritizing consistent uptime and performance for their clients.

The study suggests that data centers are uniquely positioned to become effective participants in demand response initiatives due to their inherent flexibility.

Strategies for Reducing Power Consumption

The research identifies several methods for data centers to decrease their power usage.

One approach is temporal flexibility, involving the rescheduling of computing tasks to periods of lower demand. For instance, AI model training could be adjusted to accommodate brief curtailments.

Another is spatial flexibility, where computational workloads are shifted to regions experiencing lower demand.

Even within data centers, operators can consolidate loads and temporarily deactivate a portion of their servers.

Alternative Power Sources

For critical tasks that cannot be delayed or relocated, data center operators can utilize alternative power sources to compensate for any reductions.

Batteries are particularly well-suited for this purpose, as even relatively small installations can provide several hours of power with minimal delay.

Early Adoption and Initiatives

Several companies are already exploring these concepts.

Google has leveraged its carbon-aware computing platform, initially designed to reduce emissions, to facilitate demand response.

Enel X is collaborating with data centers to utilize the batteries within their uninterruptible power supplies (UPS) to enhance grid stability.

PG&E is offering expedited grid connection for data centers that agree to participate in demand response programs.

A More Manageable Future

While these adjustments won't entirely negate the need for additional power generation, they could transform a potentially critical situation – where a significant portion of new AI servers are underpowered – into one that is more readily addressed.