Bain & Company estimated on Sept. 29 that the artificial intelligence industry would need about $6 trillion a year in revenue by 2031 to pay for the computing power it is building. The consulting firm could find less than a third of that sum in businesses that exist today.
The gap is wider still against what the industry actually takes in. OpenAI’s annualized revenue was nearing $70 billion in late September, Axios reported, and Anthropic’s passed $65 billion at the end of July, according to Bloomberg. Together the two biggest model makers are running at about $135 billion a year, a little over 2 percent of Bain’s target. Bain says $4.2 trillion of the total must come from new markets, and roughly $2.7 trillion of that is assigned to products nobody has built yet.
History offers two familiar endings for a build-out this size: the fiber glut that bankrupted much of the telecom industry in 2001 and 2002, or the railways, which remade two economies. The record doesn’t split that neatly. The railways collapsed too, more than once, and the people who paid for them often lost. What separated the episodes was who carried the debt and how long the assets stayed useful. On that second count, a graphics chip that hyperscalers write off over about five years has less in common with either.
How Bain gets to $6 trillion
The arithmetic starts with spending. Bain projects that annual outlays on A.I. infrastructure, counting new data centers and the replacement of chips, memory and networking gear already installed, could reach $1.5 trillion by 2031, according to the report. It then assumes capital spending settles at about a quarter of industry revenue, a ratio it calls “ambitious but reasonable” given cloud providers’ history. Divide $1.5 trillion by 25 percent and the answer is $6 trillion.
The near-term figures are already large. Microsoft, Google, Amazon, Meta and Oracle could spend $780 billion on capital projects this year, nearly five times the level of three years earlier, Bain wrote. The report also projects $5 trillion to $6.5 trillion of data center spending by 2030, adding about 150 gigawatts and nearly tripling global capacity in five years, The Next Web reported. Local opposition blocked or delayed at least 75 projects worth $130 billion in the first quarter alone, the same report found, a fight now playing out in state moratorium bills.
David Crawford, the chairman of Bain’s global technology practice and the report’s lead author, said in a statement that the infrastructure was being built well ahead of demand, and that paying for it sustainably would mean adding about 1 percentage point a year to global economic growth. He framed the problem as one of invention, not efficiency.
What the industry needs is a wave of innovation that will dwarf what mobile and cloud unlocked.
The target has moved fast. In its 2025 report, Bain put the requirement at $2 trillion in annual revenue by 2030 and said the industry would still fall about $800 billion short. One year and one extra year of horizon later, the number has tripled. David Cahn, a partner at Sequoia Capital, ran a similar exercise in 2023 and arrived at $200 billion. By June 2024 it had become “AI’s $600B question.”
The $6 trillion is a requirement, not a forecast, and it moves with the ratio underneath it. In 2025, capital spending at Microsoft and Meta already ran above a third of total sales, Noah Smith, the economist and writer, noted. If companies accept thinner returns for longer, the revenue needed to justify each dollar of hardware falls. If chips must be replaced faster, it rises.
Where the money would come from
Bain divides the $6 trillion into what A.I. already does and what it might. Consumer subscriptions and advertising could bring in $200 billion to $400 billion by 2031. Enterprise uses, chiefly software development, sales, marketing, customer service and IT operations, could add $1 trillion to $1.4 trillion “in gains to providers alone,” according to the report. Together that is $1.2 trillion to $1.8 trillion. Bain’s own survey work has found the gains are slow to show up: only 6 percent of marketers in one recent study said A.I. had delivered large ones.
The remaining $4.2 trillion is spread across four categories. Chatbots that displace search and sell ads could produce $100 billion to $200 billion or more. Autonomous cars, trucks, drones and factory automation are put at $400 billion. Physical A.I., meaning simulations, digital twins and robots, is valued at $900 billion, on the assumption that it cuts research and manufacturing costs by 10 percent. The fourth bucket, new products and uses “that don’t exist today,” gets no dollar figure. By the arithmetic of The Register, it has to carry about $2.7 trillion. Bain’s examples include drug discovery, mental health support and materials science.
The Register wrote that the last category was an answer “your granny could have come up with.” Steven Rattner, the financier who led the Obama administration’s auto task force, put the gap in one line on X.
A.I. companies will need to find at least $4.2 trillion a year in new revenue by 2031 to pay for their data center buildout, per @BainandCompany.
Buyers of A.I. are getting their own version of the bill. In a report on U.S. government agencies, McKinsey found that the price per token is collapsing while total spending rises, because agents, software that carries out multistep tasks on its own, consume far more computing than a chatbot answering a question, Semafor reported. Tim Ward, a McKinsey senior partner and its global public sector leader, told Semafor that for many governments, “usage might have been essentially near-zero cost.” Very soon, he said, “it won’t be.”
Companies see the same pattern. GPT-4’s API cost $60 per million output tokens when it launched in 2023, and models of similar ability now cost a fraction of that, Fortune reported. Lari Hämäläinen, another McKinsey senior partner, told Fortune that a single agent task can cost up to 30 times more from one run to the next. For providers, those rising bills are the revenue in Bain’s model. For a city agency or an insurer, they are a budget line someone has to defend.
How far short today
The fastest-growing numbers in the industry belong to the two labs. Anthropic’s run rate, a projection of recent monthly sales over a full year, was about $9 billion at the end of 2025 and passed $65 billion seven months later, Fortune wrote, citing Bloomberg. OpenAI’s was nearing $70 billion by late September, up more than 70 percent since the start of the third quarter, with business sales more than doubling, Axios reported, citing people familiar with the financials. OpenAI has not confirmed the figure.
Run rates flatter. Anthropic’s revenue for all of 2025 was about $4.6 billion, according to a draft prospectus that Reuters reviewed, Yahoo Finance reported. Neither company has published audited results for this year.
Even at face value, about $135 billion is not the whole A.I. market, which also takes in cloud rentals and A.I. features sold inside existing software. It is the clearest measure of what customers pay for A.I. itself. Getting from that base to $6 trillion by 2031 would mean more than doubling every year for five straight years. Henrik Zeberg, a macroeconomist known for bearish market calls, ran the same sum from a higher base on X after the report came out.
Today AI industry generates 175 Bn USD
(sources: ChatGPT and Claude)So - to achieve 6 Trillion by 2031 - it needs to see CAGR of 103% - every year until 2031.
And.... this is BEFORE the effects of Commoditization set in 😆 ... and before further AI capacity is build.
This is…
Growth like that has happened in this industry, briefly. Anthropic’s run rate rose about sevenfold in seven months. The four largest cloud providers held more than $2.3 trillion in contracted backlog as of their latest quarters, Yahoo Finance reported in August, citing Bank of America: $678 billion at Microsoft, $638 billion at Oracle, about $514 billion at Google Cloud and $496 billion at Amazon Web Services. “Compute remains mostly supply constrained today,” wrote Vivek Arya, the bank’s semiconductor analyst. Only 12 percent of Oracle’s backlog is due within a year.
The fiber precedent
In the five years after the Telecommunications Act of 1996, carriers laid fiber across the United States on the belief that internet traffic was doubling every 100 days. That claim came from WorldCom and was false, according to a 2003 article in the Yale Journal on Regulation that cited The Wall Street Journal. U.S. telecom capital spending reached about $120 billion in 2000, Mr. Smith wrote, drawing on a Federal Reserve Bank of San Francisco study, or around 1.2 percent of gross domestic product.
By mid-2001, carriers were using about 2.7 percent of their lit fiber capacity on average, according to an estimate by Michael Ching, a Merrill Lynch analyst. Executives at Qwest and Corvis disputed his method, Light Reading reported that July, and Mr. Ching said the fair comparison was with average use of about 15 percent in the 1980s. The glut was real either way. Global Crossing filed for bankruptcy in January 2002. WorldCom followed that July with $107 billion in assets, then the largest bankruptcy in American history.
The fiber did not go to waste. Ben Thompson, who writes the Stratechery newsletter, argued in a November 2025 essay that the network failed carriers left behind became the base of today’s internet, and that its cheapness followed from the bankruptcies. “All of this ran over fiber laid by bankrupt companies,” he wrote. He also called A.I. a bubble and said a pop would “almost certainly happen.”
His own caveat cuts against the comparison. Fiber, once in the ground, works for decades. Most hyperscalers depreciate their chips over five years, Mr. Thompson wrote, “and that may be generous.” The buildings and power connections last longer, 15 to 20 years by the estimate of Microsoft’s chief financial officer, whom he quoted.
The railway precedent
In 1845, British railways took in about £6 million. Share prices at the height of the mania that followed implied revenue of about £60 million by 1850 or 1852, according to a study by Andrew Odlyzko, a University of Minnesota mathematician who studies financial manias. The actual figure in 1852 was £15 million. Investors who included Charles Darwin, John Stuart Mill and the Brontë sisters had reliable numbers showing demand would fall short, Dr. Odlyzko argued, and bought anyway.
Traffic caught up eventually. Railway revenue in Britain and Ireland later reached £109.4 million a year, he wrote, far beyond what the mania had priced in. The proportions are worth setting beside Bain’s. British investors priced in a tenfold rise in about six years and got two and a half times. Bain’s target is roughly 44 times the two leading labs’ current run rates, in about five.
American railroads ran the cycle repeatedly. Railroad investment peaked at about 6 percent of U.S. G.D.P. in the late 19th century, Paul Kedrosky, an investor and research fellow at M.I.T.’s Initiative on the Digital Economy, estimated in a July 2025 essay, against about 1.2 percent for A.I. data centers then. Advisor Perspectives found the average across the 1870s and 1880s was lower, about 2.4 percent. After the Panic of 1893, a quarter of the capitalization of American railroads was in the hands of receivers, according to a Federal Reserve working paper by Mark Carlson.
Richard White, the historian and author of Railroaded, described in a 2025 interview with the writer Derek Thompson what Eastern railroad men concluded when they first looked at the Western lines.
there’s no way in the world these roads are going to pay for themselves. There isn’t the traffic for them.
Mr. Thompson concluded that the railroads, electricity and late-1990s broadband were all bubbles that transformed America, and that with so much debt flowing into data centers, A.I. was unlikely to be the first such technology to avoid overbuilding and a painful correction. He had made the scale point months earlier on X, after Mr. Kedrosky’s essay.
This is insane.
AI capex might account for a larger share of GDP than basically any technology since the railroad.
Basically it’s a mini-wartime economy, but the guns are chips and the tanks are databases
Why chips are not track
Dr. Odlyzko has said the closer analogy is neither. In a 2024 interview with The Economist, as Futurism summarized it, he said the railway and telecom builders knew from the start how they would make money and erred only on how much. A.I.’s business model was still unclear, he said, and he pointed instead to the telegraph and electricity booms, adding that the big technology companies were “insanely rich” and could absorb losses. Bain’s unnamed $2.7 trillion is the kind of gap he described.
Mr. Cahn of Sequoia accepted the railroad comparison in 2024 but listed where it breaks. Track between two cities can carry monopoly pricing, he wrote, while computing behaves more like airline seats. Older chips lose value as newer ones arrive, and he wrote that “this parallel doesn’t exist for physical infrastructure.”
GPU computing is increasingly turning into a commodity, metered per hour.
Mr. Kedrosky made the same point about the buildings, writing that “we aren’t building century-long infrastructure.” Mr. Thompson’s answer is that two things will outlast any crash: chip factories and new power generation. “It’s hard to think of a more useful and productive example of a Perez-style infrastructure buildout than power,” he wrote, referring to the economist Carlota Perez.
Who holds the debt
Telecom’s losses landed hard because they were financed with debt. Mr. Thompson noted that most dot-com losses were equity, while telecom’s borrowing produced a wave of bankruptcies. The hyperscalers spent the early years of the A.I. boom paying from cash flow. That has changed.
Hyperscalers and related companies such as Nvidia issued $225 billion in bonds through midyear, up 973.7 percent, according to S&P Global, Fortune reported. A Nikkei study found off-balance-sheet obligations at five U.S. tech giants, including data center leases and long-term chip purchase commitments, had grown eightfold in four years to $1.65 trillion, more than the $1.35 trillion of debt on their books. Moody’s put such deals at $1.2 trillion, with over $820 billion tied to data centers still under construction. “Market participants are growing leery of quickly rising leverage,” S&P wrote. Mr. Arya of Bank of America expects the hyperscalers’ free cash flow margins to bottom out at around negative 5 to 6 percent in 2027 and 2028, raising the question Goldman Sachs asked last month about who pays.
Howard Marks, a co-founder of Oaktree Capital, wrote in a December 2025 memo that lending to an uncertain venture can be fine, but not where the outcome is “purely a matter of conjecture.” He drew the fiber comparison directly, writing that in the telecom boom, swaps and vendor loans produced “profits that were illusory.”
If that enthusiasm doesn’t produce a bubble conforming to the historical pattern, that will be a first.
Weighing the case
The case for the build-out rests on demand that can be counted today: Mr. Arya’s backlog figures, the labs’ growth and providers’ statements that they cannot supply enough computing. Mr. Thompson adds that even a bust would leave power plants and fabs with decades of use. At the far end, the credit rating firm Egan-Jones wrote in a note to clients on Sept. 30, titled “It’s Over,” that “the complete disruption of the economy is all but certain,” Semafor reported.
The case against rests on the shape of the arithmetic. Bain’s required revenue tripled in a year. The largest piece of its $4.2 trillion gap has no name and no estimate. And the financing has moved into bond markets and off balance sheets, the pattern that preceded the telecom failures.
Against the record, each side is right about a different thing. The demand evidence is real, but it describes 2026, and part of the cloud backlog is the labs’ own commitments to rent computing, to be paid with the revenue in question. The railway comparison holds for the economy: the lines were eventually used far beyond what the mania priced in. It fails for the people who paid first, which is where fiber fits better. Dark fiber sat idle for years and still worked when the traffic came. A chip written off over five years has much less time to wait.
What comes next
OpenAI has put its listing off. Sam Altman, its chief executive, has ruled out going public this year, and the company is in early talks to raise at least $30 billion at a valuation of about $1.4 trillion, Bloomberg reported on Sept. 29, according to TechCrunch. In March it raised $122 billion at an $852 billion valuation.
Anthropic is moving first. It could begin marketing shares the week of Nov. 9 and is aiming for a valuation of up to $2 trillion, Bloomberg reported, about 31 times its July run rate. The company confidentially filed a draft registration statement on June 1. Under S.E.C. practice, that filing must become public at least 15 days before the roadshow, putting the company’s revenue and spending on the record. Meetings with prospective investors are scheduled to begin Oct. 14.