There are, by one careful count, about 20 million AI chips now humming inside data centers around the world. That number is on pace to double roughly every nine months, according to the research firm Epoch AI — a trajectory that would put the global fleet at roughly 200 million chips, ten times today's total, by the end of 2028.
The estimate anchors a sweeping New York Times feature published July 29, 2026, which drew on Epoch AI, Cleanview and SemiAnalysis to map the scale of the AI infrastructure boom now under construction from the American Midwest to the Persian Gulf. The Times frames the build-out in historic terms, comparing it to the 19th-century railroad boom, the New Deal and the Manhattan Project. Epoch AI's own published data puts the underlying growth rate even higher in places: the firm's researchers found that global AI computing capacity, measured in H100-equivalent units tied to Nvidia's benchmark 2022 chip, has grown by roughly 3.3x per year since 2022 — a doubling time of about seven months, with a 90 percent confidence interval of six to eight months. Combined with data centers reaching gigawatt scale and hundreds of billions in annual capital spending, the numbers describe an industry still accelerating rather than leveling off.
“This is the largest scale infrastructure build-out in the history of humanity,” said Rob Wachen, a co-founder of the chip startup Etched, which has raised more than USD 1 billion to meet surging demand for AI components, in comments to the Times. Amazon's Peter DeSantis, who leads foundational AI models at the company, told the paper that Amazon has doubled its computing capacity since 2022 and plans to double it again by next year: “It's hard to get your mind around the scale.” The pace shows up concretely in the chip count: the world had about 2.4 million H100-equivalent chips in March 2024; by Epoch AI's tracking, that figure has already grown roughly eightfold in under two and a half years, with millions more advanced chips coming online every month.
The scale-up is inseparable from an energy story. AI data centers consumed about 64 gigawatts of electricity globally last year — comparable to Germany's total consumption — according to SemiAnalysis, and that figure is expected to roughly quadruple by the end of 2030, eclipsing the combined power draw of every country in South America and Africa. Epoch AI's own tally, released in January, put total AI data-center power capacity at approximately 30 gigawatts as of the fourth quarter of 2025, on par with the peak electricity demand of New York State. At the highest end of the build-out, industry estimates cited by the Times put the cost of each additional gigawatt of AI data-center capacity at USD 40 billion to USD 60 billion once servers, land, connectivity and utility hookups are included. Five individual U.S. facilities are on track to reach the 1-gigawatt threshold in 2026 alone. Behind the compute race sits a geographic imbalance: the United States hosts about 5,500 data centers, roughly ten times its nearest rival, and American companies — Amazon, Google, Microsoft and Meta chief among them — control about 80 percent of global AI computing power, per Epoch AI. China, by contrast, held roughly 1.16 million H100-equivalent chips at the end of 2025, up from about 244,000 at the start of 2024, though that figure excludes smuggled or offshore compute Chinese firms may be using.
Why It Matters
The doubling logic matters because so much of the AI industry's roadmap is built on it. The belief that feeding models more data and more compute reliably yields more capability — the so-called Scaling Laws — is the premise behind Anthropic chief executive Dario Amodei's forecasts that AI could soon perform large swaths of white-collar work, and Google DeepMind's Demis Hassabis's description of a coming era of “10x of the Industrial Revolution at 10x the speed.” Google chief scientist Jeff Dean told the Times the payoff of this spending is capability itself: “You see capabilities emerge at larger scale that didn't occur at smaller scale,” he said, adding that the goal is bringing those capabilities “to hundreds of millions or billions of users,” not just a niche research audience. But the same exponential curve that promises faster drug discovery and more capable AI agents also compounds costs and risk. AI infrastructure investment is forecast to top USD 1 trillion globally by 2029, up from USD 318 billion last year, per IDC — comparable to the entire economic output of Switzerland — and Goldman Sachs projects Amazon, Google, Microsoft, Meta and Oracle alone will spend about USD 750 billion this year, up from roughly USD 400 billion in 2025. Economists warn the spending is outrunning proven returns. “Each time you've had a technological revolution, this kind of bubble bursting happened,” said Philippe Aghion, the 2025 Nobel laureate in economic science, who compared the moment to past cycles like the dot-com crash. Oxford economist Carl Benedikt Frey put the corporate logic more bluntly, describing an “AI arms race” in which companies that don't keep investing are “acknowledging defeat.”
What to Watch
Watch whether the doubling rate itself bends — either because chip and grid construction can't keep pace, or because financing tightens on infrastructure bets that haven't yet proven their returns. Watch the energy fight play out locally, as data-center power demand and water use become flashpoints ahead of November's U.S. midterm elections, and globally, as China's chip stockpile (still a fraction of America's) narrows the gap through domestic semiconductor production. And watch labor-market signals, like the Remote Labor Index cited by the Times, where AI models' ability to complete freelance-style tasks reportedly jumped from 2.5 percent in October to 16 percent by July — an early proxy for how fast this compute deluge translates into displaced or transformed work.
"This is the largest scale infrastructure build-out in the history of humanity."- Rob Wachen, Co-founder, Etched (to The New York Times)