The future of artificial intelligence may depend, in part, on whether providers can reduce their appetite for electricity and water
5 min
Artificial intelligence is poised to transform both work and everyday life. But it has a dark underside: AI computer centres consume enormous amounts of electricity and water, to power their processing chips and cool the heat they emit.
Annual U.S. electricity use by data centres of all types will rise from 3% to 4% of the nation’s total today to between 11% and 12% in 2030, with AI being the main driver, according to projections from consulting firm McKinsey.
Meantime, AI’s demand for water globally in 2027 could account for more than the total annual amount withdrawn for use in Denmark or half of that in the U.K., according to researchers at the University of California, Riverside and University of Texas at Arlington.
All of that heavy use is causing logistical and public-image problems for the industry. Some utilities struggle to supply the needs of AI providers, and communities push back, fearing the added use will boost power prices and deplete water supplies.
The biggest AI providers, including Amazon , Alphabet Inc.’s Google, Meta and Microsoft , say they are working to be both carbon-neutral and replenish more water than they use—even as they continue to build massive data centres.
“It will be harder to build data centres, especially where energy already is at a premium or water might be scarce,” says Ed Anderson, research vice president at technology advisory firm Gartner. But, he adds, “the economic opportunity is rich enough that the providers will find a way.”
Below are some of the steps tech companies and researchers are hoping will reduce AI’s appetite for power and water.
One way of addressing power consumption is to make chips less power hungry. Nvidia , the largest maker of AI processors, says its newest ones, called Blackwell, will be about 25 times as energy efficient as its previous high-end version. Meanwhile, Amazon, Google, Meta and Microsoft are designing their own processing chips, in part to cut costs but also to make them use less power.
“Each generation has been significantly more efficient than the prior one,” says Google’s Partha Ranganathan , a vice president and engineering fellow, speaking of his company’s processing units.
Equipment used to cool data centres creates another issue: where to get the vast amount of water these systems consume. Google says its data centers globally used about 6.1 billion gallons of water in 2023, equivalent to the water used to irrigate and maintain 40 golf courses in the Southwest each year.
OpenAI’s GPT-3 model, meantime, consumes the equivalent of a 16.9-ounce bottle of water for every 10 to 50 responses it provides to users’ queries, according to the researchers at UC Riverside and UT Arlington. OpenAI declined to comment on the finding.
Data-centre water typically comes from municipal water systems. But in an era of water shortages, diverting drinking water for an industrial use has created tensions in some locales. That has sent AI companies searching for other sources, including rainwater, treated wastewater or water left over from factory processes.
Amazon, for example, uses recycled wastewater for cooling at its Santa Clara, Calif., data centers. The water comes from the city’s sewage-treatment system after it undergoes a three-step process that removes 99% of impurities.
Some researchers have experimented with carefully controlling what kind and how much information an AI model takes in during training. Usually, training a so-called large language model AI, such as OpenAI’s ChatGPT and Microsoft’s Copilot, involves ingesting hundreds of billions of words from the internet and elsewhere, then learning the relationships among them.
And that is energy and water intensive. Training an AI model called BLOOM over a 3½-month period consumed enough electricity to power the average U.S. home for 41 years, according to a Stanford University report.
As for water, training one of Google’s AI models, known as LaMDA, used about two million liters of it, both to produce the electricity used and keep the computers cool—enough to fill about 5,000 bathtubs, according to Shaolei Ren , a professor of electrical and computer engineering at the University of California, Riverside. Google declined to comment on the research, but said it is “committed to climate-conscious cooling of our data centres.”
One possible solution is to have AIs remove redundancy and low-quality data, instead of just vacuuming up the whole internet. The goal is a much smaller set of data that the AI system can more easily sift through when a user asks it a question.
This can lower electricity consumption, according to some researchers.
AI systems that limit the information they take in are also less likely to “hallucinate”—give false or misleading answers—and can respond in ways that are more on-point because of the higher quality of the data they contain, experts say. Microsoft found that one of its pared-down AIs exceeded that of vastly larger ones in measurements of common sense and logical reasoning .
Researchers at several universities have found that capping the amount of electricity used by AI computers has only a minor effect on the outcome, such as slightly more processing time.
Experts at the Massachusetts Institute of Technology and Northeastern University say that reducing the power to one of Meta’s AIs by 22% to 24% slowed the speed at which the AI responded to a query by only 5% to 8%. “These techniques can lead to significant reduction in energy consumption,” the researchers say. They add that the method also caused the processors to run at a lower temperature—which could trim the need for cooling.
Meta declined to comment on the research, but said it has had efforts to boost data-centre energy efficiency “since we started designing our first data center over a decade ago.”
Meantime, a team at the University of Michigan, University of Washington and University of California, San Diego devised an algorithm to modulate the use of power during training. The technique could cut power use by up to 30% , they say.
Some researchers believe companies should give users more context about the environmental impact of AI, to let them make more-informed decisions about the technology. Ren, of UC Riverside, proposes that AI providers disclose the approximate amount of electricity and water each query consumes—akin to how Google tells people searching for flights the amount of carbon emissions each trip would create.
Another proposal is to devise a rating system for the power efficiency of AI systems, akin to the government’s Energy Star ratings for home appliances and other products. Such a system could help people choose AI models for differing tasks based on their energy consumption, according to Sasha Luccioni , an AI researcher at Hugging Face, a company that makes machine-learning tools.
Academics and others have come up with other proposals to minimise AI’s environmental impact by tapping into green energy. For instance, companies might build more data centers in countries with abundant, low-emission power, such as hydropower in Norway or geothermal in Iceland. Or companies might do AI calculations at different locations at different times of the day, such as deploying computer centers with high use of solar power during the daytime or wind-powered ones when wind is more reliable at night.
Data-centre computers put out tremendous amounts of heat, and their temperature must be kept in a certain range, often 64 to 72 degrees, to prevent damaging the electronics. Traditionally, this has been done by high-power air conditioning. But air conditioning uses up to 40% of all the electricity consumed by a typical data centre, while devices called cooling towers that expel the heat to the outside air use a lot of water.
In response, the data-centre industry is moving to liquid cooling, which circulates a special liquid or cold water to “cold plates” that sit on top of the processor chips and keep them at a safe and efficient temperature range. The system, called direct-to-chip liquid cooling, uses less power than the traditional method—about 30% less, Nvidia says—because liquid is vastly better at removing heat from the electronics than blowing cold air over them.
Another method under development, called immersion cooling, involves placing the computers themselves inside big tanks of cooling liquid. While showing early promise, there are environmental concerns about the chemicals often used in the setup, says Mark Russinovich , chief technology officer of Microsoft’s Azure cloud-computing unit.
Some companies, meanwhile, are using computing gear that can withstand higher temperatures and doesn’t need as much cooling. Google says its data centres already are 1.8 times as energy efficient as the typical data centre, which it achieved in part by raising the inside temperature to 80 degrees. For every one-degree boost in their temperature, data centres can save 4% to 5% in energy costs, according to the Energy Star program.
Copyright 2020, Dow Jones & Company, Inc. All Rights Reserved Worldwide. LEARN MORE
The Australian leather house has opened an immersive four-day pop-up in Manhattan, unveiling its Bloom Collection and redefining what a product launch can look like.
Following the successful launch of its Palais Collection, MAISON de SABRÉ has unveiled a new modular handbag system offering more than 720 styling combinations.
Wall Street’s hottest momentum trade has reversed sharply, as former winners tumble and heavily shorted stocks surge.
3 min
Wall Street’s hottest trade has gone ice cold.
For years, it paid off to buy stocks that were rising in price—and bet against struggling shares. The momentum trade was especially profitable this year, as investors piled into hot stocks including Micron Technology, Nvidia, Advanced Micro Devices and other artificial-intelligence darlings while wagering against those likely to be hurt by the embrace of AI.
The S&P 500 Momentum Index soared 44% in the second quarter, its best quarterly performance on record, and it surged 133% over the past five years, nearly double the broad market’s performance.
Mega funds and rookie investors alike piled into the trade, some using leverage and options contracts in an effort to amplify their returns, propelling the underlying shares higher.
“It is a self-fulfilling prophecy,” said Matthew Tym, managing director at Cantor Fitzgerald, of the trade.
Suddenly, the trade is a loser. The momentum index has tumbled more than 9% since July 1, lagging behind the S&P 500’s 2.8% gain. The index—which tracks stocks in the S&P 500 based on a “momentum score”—is on track for the biggest quarterly underperformance in 25 years. July was the second-worst month for the momentum trade in around 40 years, according to Bank of America estimates; the only month worse was April 2009, in the teeth of the global financial crisis.
Hedge funds that bought momentum shares while shorting low-momentum stocks suffered even more. At the same time, a basket of the most popular stocks held by hedge funds tracked by Goldman Sachs recorded its biggest one-month underperformance in July relative to the S&P 500 in more than 20 years, according to the bank’s analysts.
Momentum trading is based on a rather simple observation: Investments that go up tend to keep outperforming; those that underperform often remain laggards. This kind of trading might seem too simple a stock-picking strategy to work. Yet it often has.
“For decades, it didn’t take a lot of sophistication to run a momentum strategy and make a decent living at it,” says Agustin Lebron, senior researcher at EquiLibre, a trading firm.
Part of the reason: It takes a while for corporate and other information to spread to various investors, so they slowly build positions, producing buying momentum.
“A huge pension fund can’t flip around its positions in a day,” says Lebron. “Behavioral biases also account for some of the effect, as well—people tend to sell their winners too early and hold losers too long.”
Fans of the strategy point to the human tendency to extrapolate from past results—and chase investment returns—noting that momentum patterns have been evident in markets for decades, even centuries. They also say that some of the worst months for momentum strategies are during longer periods of outperformance.
Some have been doing the trade by buying the strongest investments in a sector while shorting the weakest; others lean in to rising markets or asset classes. Still others use a quantitative approach or turn to banks or others who sell ways to make distinct wagers on momentum as a “tradable factor” or a “thematic basket.”
The fans remain believers. “Any strategy has disappointing periods,” says Antti Ilmanen, global co-head of the portfolio solutions group at AQR Capital Management.
The surge in Moderna and other biotech stocks helped crush the momentum trade. These shares were among the most heavily shorted in recent years, but positive news on a cancer vaccine from Moderna and Merck sent those stocks flying, crushing some quant and other hedge funds. Moderna is up around 150% so far this month.
These traders had an especially rough day on Aug. 19, which Goldman Sachs told its clients was the worst day for “systematic long-short managers” in more than two years. About half of the losses were because of momentum trades, the bank said.
Some traders have begun to short, or bet against, the very stocks that propelled the momentum trade earlier this year. Net short positions in futures tied to the Nasdaq-100 index among speculators recently climbed to some of the highest levels of the past two decades, according to data from the Commodity Futures Trading Commission.
The about-face is a sign of how markets have become more treacherous for investors, even as indexes keep climbing. Part of the issue: the recent meltdown of Situational Awareness, a hedge fund that had piled into some of the most popular momentum shares, including chip stocks. After a period of market tumult, Nvidia shares rocketed almost 9% after its earnings, showing how quickly sentiment can shift.
Some investors say the run-up in share prices driving tech stocks higher reminds them at times of the dot-com frenzy decades ago.
Mike Ogborne, the founder of San Francisco-based Ogborne Capital Management, said he has grown more cautious on technology stocks and is keeping more of his portfolio in cash than he typically does.
And he is nervous about the surge in spending by technology giants and quarterly capital expenditures that keep rising.
“It is a little bit like Cinderella and the clock striking midnight. You don’t know when midnight is going to come around,” Ogborne said. “They don’t send a memo around telling you when the capex cycle is over.”
