News in English

What If Tech Execs Don’t Really Need All These Data Centers?

A “fiber glut” might sound like a phrase uttered by a wellness influencer extolling the virtues of ancient grains. In reality, it’s the name for an early 2000s phenomenon that drove some of the world’s fastest-growing companies into the ground. Amid the pre-Y2K dot-com boom, corporations, industry analysts, and even the U.S. Department of Commerce began to circulate a shocking statistic: that the internet doubled in size every one hundred days, implying an annual growth rate of 1,000 percent. Millions of miles of fiber-optic cables were buried below streets and oceans in the late 1990s to keep up with supposedly exponential demand that never materialized. By 2002, just 2.7 percent of the fiber installed in the preceding few years was being used.

Companies that drove the fiber glut were saddled with financial ruin and accounting scandals. “The actual traffic growth was never close to 1,000% per year,” Vint Cerf, senior vice president at the fiber optics firm WorldCom, admitted to The Wall Street Journal. “But I don’t think it was an attempt to misstate anything—it was an honest characterization of what kind of demand we were seeing from these companies.”

Today, a new generation of tech executives are making similarly audacious claims about demand for their products, justifying yet another infrastructure-building binge. The industry this time is artificial intelligence. OpenAI founder Sam Altman has been pitching everyone from the Biden administration to the United Arab Emirates for help in building out a fleet of data centers that would each require five gigawatts of electricity, about a thousand times more than the average data center. OpenAI’s vice president of global affairs—Clinton administration alum Chris Lehane—recently predicted that the U.S. would need to invest $175 billion in AI infrastructure: funds, he contends, that will “reindustrialize the country” and “revitalize the American Dream.” Banks and consultancies have bolstered those claims. Goldman Sachs, for instance, projects that data centers will require an additional 47 GW of power through 2030. The consulting firm Grid Strategies has forecast that U.S. electricity demand will grow by 456 percent over the next five years, by 128 GW. After meeting with OpenAI, Meta, Microsoft, and other firms pushing a data center build-out, the White House is now reportedly mulling an executive order that would fast-track data center construction, potentially allowing new facilities to exceed pollution limits, build on federal lands, and receive preferential access to available power supplies.

Given that most of the country’s power is still generated from fossil fuels, this raises some obvious climate concerns. Major tech companies have started investing in next-generation energy technologies like nuclear fusion, carbon capture, and small modular nuclear reactors. But those require lead time and in some cases luck, and the same companies want new data centers now. Data center growth is already helping to extend the life of coal-fired power plants and fueling a boomlet for the gas power providers furnishing Silicon Valley titans with new turbines. There is nothing inevitable, though, about the tremendous energy demand that AI boosters say they’ll need.

“Nobody has any idea what AI electricity usage in data centers is going to be in three to four years,” says Jonathan Koomey, a researcher and consultant who studies the energy impact of internet and information technology. Electricity demand is indeed growing for the first time in over a decade, he says. Not all of that is from data centers, and larger spikes in demand are generally concentrated in places with new factories and data centers, like Virginia and Georgia. The modest overall load growth happening now, moreover, doesn’t indicate that there’s some looming crisis in which the U.S. will run out of electricity as data centers proliferate.

The bigger risk may well be that these fantastical demand projections are used to bring new fossil fueled power plants online and keep existing ones running—regardless of whether they’re actually needed. Once built, new coal or gas plants are likely to operate for decades. “When people think it’s a crisis they make big mistakes,” Koomey says. “It’s absolutely not a crisis.”

There are economic reasons to push for such a massive infrastructure binge. In many states, building new infrastructure is one of the few ways that electric utility companies can raise rates, decisions that require approval from the public service commissions that regulate what they charge and the profits they make. Big new sources of electricity demand—met with new infrastructure—can mean higher profits. Those utility companies are also political juggernauts in state capitols, where lobbyists can persuade lawmakers to roll out generous incentives to AI developers promising jobs and economic development. All of the above are understandably eager to avoid energy shortages at all costs, of course, and view having too much power as preferable to having too little.

Energy companies like Chevron and ExxonMobil, meanwhile—which recently announced it’ll start providing power to data centers—have a fairly straightforward interest in finding new sources of demand for the fossil fuels they drill and refine. For tech companies themselves, long-term power supply contracts foster path dependency: If new plants get built then there’s a strong incentive to use them, however much data center capacity ends up being needed. That could help cement the capital and resource-intensive AI development that big tech firms, especially, have opted to pursue.

“They’ve gotten themselves into a business model that’s very cost-intensive up front, and very unproven in terms of what profits they’re going to have on the back-end or what the ultimate utility will be,” says Sarah Myers West, co-executive director of AINow and a former senior adviser on AI at the Federal Trade Commission. “There are many different ways of building AI,” she adds, including methods that are more efficient and less power-hungry. “It’s not for nothing that the industry has locked into this large-scale AI paradigm at a moment when the labs that are producing AI have been dominated for the last decade by the companies that run cloud infrastructure businesses,” which are better equipped than smaller competitors to shoulder the massive costs that larger-scale models entail.

The business case for this kind of development isn’t exactly self-evident. The New York Times reported in late September that OpenAI was on track to lose more than $5 billion this year, citing fundraising materials that indicated the company “would need to continue raising money over the next year because its expenses grew in tandem with the number of people using its products.”

Amid that hunt for new cash and revenue streams, leading AI firms have been pitching the White House for public investment in AI as a national security priority. By building hundreds of billions of dollars in new AI infrastructure, OpenAI’s Chris Lehane urged in a speech at the Center for Strategic & International Studies last month, “we can ensure that democratic AI prevails over autocratic AI. That the U.S. version prevails over the People’s Republic of China version.” OpenAI quietly rolled back its own ban on using ChatGPT for “military and warfare” at the beginning of this year, and has already started to rack up defense contracts, joining Amazon and Google in the competition for Pentagon dollars. As Donald Trump prepares to take office, Meta, Amazon, and Altman have donated generously to his inaugural committee.

As these companies and their CEOs look to secure steady returns, they’ve been relatively oblique about how their allegedly exponential growth—entailing potentially massive new fossil fuel emissions—would actually benefit the public. “You could make the argument that these new data centers are doing things that are high value. Society needs to have a discussion, though. We should have that discussion about crypto, about AI, and about computing more generally,” Koomey says. “The answer will likely be different for those different buckets, but part of the problem with the hype is that people are being diverted away from having the value discussion.”

Читайте на 123ru.net