Big tech’s great AI power grab

The comparison with the famously capex-happy energy industry is apt not just because of the sums involved. AI needs vast amounts of processing power. And that processing power needs vast amounts of electricity. On May 2nd Bob Blue, chief executive of Dominion Energy, one of America’s biggest utilities, said that data-centre developers now regularly ask him for “several gigawatts” (GW). Dominion’s total installed capacity is 34GW.

Graphic: The Economist

View Full Image

Graphic: The Economist

JPMorgan Chase, a bank, calculates that Microsoft, Amazon’s cloud arm (AWS), Alphabet, Meta and Microsoft consumed 90 terawatt-hours (TWh) of electricity in 2022, as much as Colombia. And that was mostly before ChatGPT set off the AI revolution in November that year. The ensuing hoopla led the the International Energy Agency (IEA), an official forecaster, to predict that data centres (including those dedicated to AI and equally energy-hungry cryptocurrencies) will gobble up more than 800TWh globally in 2026, more than double the amount in 2022 (see chart). BCG, a consultancy, reckons that data processing could triple its share of American power consumption by 2030, to 7.5%.

And not just any power will do. The technology titans want theirs to be clean. In April their industry association warned Georgia Power, which had managed to fast-track the approval of 1.4GW of new fossil-fuelled generation by pointing to rising demand from data centres, that its members would build fewer of these in the southern American state if the utility spewed extra carbon. Combined with rising demand from increasingly electrified transport, heating and parts of heavy industry, digital technology’s power needs are putting enormous strain on the businesses that generate and distribute electricity.

BloombergNEF, an information firm, reckons that annual grid investment needed to fully decarbonise global electricity by 2050 will need to rise from about $300bn in 2022 to $600bn in 2030 and well over $800bn in 2050. Risk-averse utilities, which would normally undertake grid-expansion projects under the watchful eye of cost-minded regulators, have neither the money nor the appetite to do so.

Enter big tech itself. The deep-pocketed giants have already been the biggest force behind green “power-purchase agreements”, which helped kickstart America’s renewables boom by persuading utilities and other investors to build wind and solar farms. They are now getting in on the green-energy action more directly.

On May 1st Microsoft and Brookfield, one of the world’s biggest infrastructure investors, announced a deal to build 10.5GW of renewables capacity in America and Europe by 2030. The arrangement is meant to enable the software giant to meet its pledge to have 100% of its electricity use, 100% of the time, come from zero-carbon sources by 2030. Microsoft and Brookfield have not revealed the price tag, but adding a gigawatt of wind or solar capacity can cost about $1bn.

One problem is that data centres tend to consume power at a steady rate, including when the sun is not shining nor the wind blowing. So technology firms are also thinking of ways to make data-processing more flexible. In March Sidewalk Infrastructure Partners, an infrastructure-technology company co-created by Alphabet, presented a detailed plan for how this could be achieved. It involves a combination of microgrids (which can run independently but also exchange energy with others nearby), batteries and advanced software in order to enable shifting less time-sensitive tasks, such as training AI models, to periods of fallow demand. Jonathan Winer, one of Sidewalk’s founders, expects such data centres to pop up first in energy-constrained places like Arizona, California and Massachusetts.

Renewables are not the only area of big tech’s power interest. In March AWS paid $650m for a 960-megawatt (MW) data centre in Pennsylvania powered by a nuclear reactor located next door. Microsoft has struck a deal with Constellation Energy, America’s biggest nuclear operator, for supply of nuclear power for its data centre in Virginia, as a backstop when wind and solar are unavailable. Both firms have also been looking at “small modular reactors”, a promising though unproven nuclear technology.

Google, meanwhile, is dabbling in geothermal energy. The search giant has signed the first-ever corporate deal to develop “enhanced” geothermal power with Fervo, a startup that has raised $430m in venture capital. Inspired by the shale industry, the hot-rocks hotshot has developed horizontal wells, monitored using fibre-optic cables. Its site in Nevada produces round-the-clock, carbon-free power for the local grid—which Google then acquires. Tim Latimer, Fervo’s boss, says that every drilling rig his firm operates can add 100MW of power. The firm is developing a 400MW commercial plant in Utah that will start feeding the grid in 2026. The Department of Energy reckons that innovations like Fervo’s could expand geothermal output in America around 20-fold, to more than 90GW, by 2050.

Google and Microsoft have also teamed up with Nucor, a giant American operator of steel mini-mills, which consume lots of electricity. In March the trio announced that they will aggregate demand and jointly offer contracts to clean-energy projects, both early-stage commercial ones and entirely novel “first-of-a-kind” ventures. The idea is to guarantee custom for developers of promising technologies like long-duration energy storage, clean hydrogen, next-generation geothermal and nuclear energy.

The AI industry’s most exotic power plays come courtesy of Sam Altman, the techno-optimistic boss of OpenAI, maker of ChatGPT and Microsoft’s main model-making partner. In a quest to power the AI revolution, he has invested in Helion, a nuclear-fusion startup, and Exowatt, a startup developing solar modules that can act as both electricity generators and thermal-storage batteries. Mr Altman is now looking to raise $500m for Oklo, which is working on nuclear micro-reactors that run on spent fuel from larger ones and that could power individual factories, corporate campuses and, of course, AI server farms. These wagers may seem fanciful. Then again, 18 months ago so did the idea that an AI could write essays or paint like a human.

To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.

© 2024, The Economist Newspaper Limited. All rights reserved. From The Economist, published under licence. The original content can be found on www.economist.com

Leave a Comment