Over-hyped A.I. data centers are just one reason to beware of an A.I. investment bubble

I enjoy using A.I., artificial intelligence. The A.I. overview that pops up at the top of most of my Google searches seems pleasingly accurate. So is the A.I. summary of the reviews of an Amazon product that I’m considering purchasing. And I’ve played around with asking ChatGPT questions, and making amusing images with Grok Imagine.

But a fascinating article in The New Yorker by Stephen Witt, “Inside the Data Centers that Train A.I. and Drain the Electrical Grid,” clued me in to some scary facts that illustrate the danger of believing that A.I. is going to be an economic gold mine and a cultural success. Here’s a PDF file if that link doesn’t work for you.
Inside the Data Centers That Train A.I. and Drain the Electrical Grid | The New Yorker

• the Nvidia GB300 computing rack is a little taller than a refrigerator and costs a few million dollars
• in a busy year, a typical rack will use more electricity than a hundred homes
• in a Microsoft A.I. data center the area dedicated to computing was the equivalent of twenty football fields
• a data center can use as much electricity as the city of Philadelphia, about one gigawatt
• the former CEO of Google says the U.S. must add 92 gigawatts of power to the national supply to meet data center demand
• that’s ninety-two Philadelphias
• in the near term, new data centers will largely be powered by fossil fuels
• this will speed up global warming, already a massive threat to humanity

The United States stock market is doing well. But much of that growth is coming from tech companies heavily involved with A.I. So there’s worrying signs of an A.I. bubble that could pop, leaving our economy in a situation akin to the 2008 financial meltdown that was spurred by rampant speculation in housing.

Today the Washington Post ran a story, “Is there an AI bubble? Eight charts will help you decide.” (That’s a gift link.)

(1) The largest tech companies are spending unprecedented sums on AI. Their capital expenditures — money spent on assets like buildings and equipment — have soared over the last two years, as they splashed out on data centers stuffed with powerful computer chips to run AI software. Skeptics have questioned how those immense investments will be recouped, pointing out that ChatGPT and rival AI products are racking up billions of dollars in losses each year. JPMorgan recently calculated that the tech industry must draw an extra $650 billion in revenue every year — nearly double Google’s total sales in 2024 — for AI investments forecast through 2030 to earn a modest 10 percent return. ChatGPT maker OpenAI’s revenue is not public but its CEO, Sam Altman, has said it will soon reach a rate that would bring in $20 billion annually. (The Washington Post has a content partnership with OpenAI.)

(2) It’s not unusual for companies to be bound into a web of deals with suppliers and partners in their industry. But some recent AI deals have struck industry watchers as out of the ordinary, with companies like OpenAI and chipmaker Nvidia making large investments in their own customers or suppliers. Some analysts have warned the deals resemble unsustainable patterns seen in previous financial bubbles, where troubles at one company quickly spread to others. Spokespeople for Nvidia and Oracle declined to comment. “OpenAI’s growth comes from real usage and real value. In the three years since ChatGPT launched, we have built the fastest-growing consumer and enterprise platform in history,” a spokesperson for OpenAI said. Nvidia chief executive Jensen Huang has defended his company’s deals, saying it makes sense for it to invest in the broader AI ecosystem. “I fully expect that investment to translate to extraordinary returns,” Huang said on the company’s third-quarter earnings call this week.

(3) For the AI boom to pay off, the technology must work not just in controlled demos or diverting apps but in messy, real workplaces. But one recent study that challenged top AI systems to perform a wide range of work tasks found that they did poorly. The tasks were chosen to represent the kind of work that might be assigned to a remote human worker, including creating a video advertisement and drawing up an interior design plan. The best AI system successfully completed a measly 2.5 percent of the tasks, according to the research team from Scale AI and the Center for AI Safety, a nonprofit that works to understand risks from AI. That suggests it’ll be some time before AI displaces humans from some jobs. A Yale study released last month concluded that AI has not yet disrupted the labor market in a significant way.

(4) Businesses could be big business for AI companies — if they actually use the technology. But several surveys have found that while many companies are experimenting with AI tools, few have widely deployed them. McKinsey reported in November that nearly two-thirds of global businesses adopting AI have not yet scaled the tools across the company. One problem is that when companies give their employees AI tools, many workers don’t want to use them, said Matt Kropp, a Boston Consulting Group executive who works with corporations trying to deploy AI. “We’re humans, we work the way we work,” he said in an interview. “You have to do something to get people to change their habits and behaviors.”