How Do You Power AI Without Overpowering The Grid?
gettyA question on many minds is: How do you power AI data centers without overwhelming the electrical grid?
To the casual user, AI is easy and effortless. You ask a question or ask a model to summarize a document. Then, like magic, a few seconds later, a full, coherent, logical response appears.
But behind that seemingly simple interaction, there’s an enormous amount of computing infrastructure at work. And that infrastructure, typically housed in data centers, requires increasing amounts of electricity to run.
As AI deployment accelerates, data center power consumption is becoming a top-of-mind issue. We’re already seeing it become a political hot potato, as communities across the country start protesting the building of data centers in their backyards. But beyond the political considerations, there’s a more basic issue: Where is all that power actually going to come from in the first place?
Data center power consumption is already rising exponentially. In 2023, Lawrence Berkeley National Laboratory estimated that 4.4% of U.S. electricity was gobbled up by data centers. Depending on how the market develops, that number could soar to 12% by 2028. The U.S. Department of Energy agrees, estimating that AI could cause data center electricity consumption to double or even triple between 2023 and 2028.
The U.S. power grid needed an update even before AI. So, if AI demand keeps expanding at this rate, a capacity and reliability problem is already expected. Utilities, grid operators and infrastructure planners themselves are starting to voice real concerns about data center growth. Case in point, the North American Electric Reliability Corp., NERC, the grid’s not-for-profit monitor, is sounding the alarm over AI’s unprecedented growth and warning of widespread blackouts.
Today’s data centers are clearly starting to tax regional grids to the point where utilities have to scramble to keep up. Embedded in NERC’s warning is concern for “high impact” risks from data centers oscillating grid load either from safety cutoffs, component fault cutoffs or their own operations. These centers represent such a large single load that the system can be significantly shocked from their changes. This can have a direct impact for keeping the power on for their other customers: hospitals, homes and businesses already on that same grid.
A push for having the largest users generate their own electricity risks adding generation inefficiency and drastically increasing pollution and water consumption in local communities. Getting rid of the heat data centers generate adds to the electrical load required and also water consumption.
I’d argue that to solve the data center power problem, you also need to reconsider how efficiently AI computation itself gets performed. Because the way AI processors are configured today, there is a lot of room for improvement.
The reality is that today’s AI chipsets require massive amounts of power. A big reason for that is their need to constantly move data around. You see, traditional processors keep memory and computation separate. Before a chip can actually perform its task, it must go get the data it needs from memory. That back and forth can happen billions of times over for a single workload, which eats up a lot of energy.
So how do we solve this problem? By making AI chipsets that are designed for both high performance and power efficiency. In short, we still need to power the next generation of AI. We just need to do it using a lot less energy. The typical chip today is trying to push the limits of AI with ever more powerful, and power-hungry, processing.
That’s not the way to go. Yes, we need processors that deliver more raw performance to run increasingly powerful AI models. But that performance can’t keep coming at the cost of ever-higher power draw. What we actually need is more performance per watt, not just more watts. New chipsets specifically designed for both higher performance and higher efficiency must be a major component of AI development going forward. To do that, we will need to look at the use cases and optimize both the hardware and the software, but also look at doing both differently than what it took to get us here.
For technology leaders, the challenge is not simply finding enough electricity to fuel AI growth, but ensuring that growth remains economically and environmentally sustainable. As we plan the next generation of AI, energy efficiency must become a measure of innovation itself, because the future of AI depends not just on what we can build, but on whether we can power it sustainably.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
