Paying for AI’s Data Centers Would Take 250 Million Jobs. America Has 159 Million

To pay for the AI data centers being built today, AI would have to replace the equivalent of 250 million jobs. The US has 159 million nonfarm payroll jobs.

That assumes the companies building AI need about $2.5 trillion a year in AI revenue, a scenario I explain below, and that they collect a dime for every dollar of work their AI does. Each job counts at $100,000, roughly what a full-time US job costs its employer, and making a worker more productive counts the same as replacing one.

In the 2016 edition of The Future of Leadership in the Age of AI, Luka Ivezic and I predicted that AI itself would become a commodity, the way personal computing had before it. Capability would keep climbing while costs fell, and “when a high level of supply meets a high level of demand, you have a commodity.” The PwC study we quoted inthe 2020 edition expected AI costs to decline over ten years. OpenAI’s price for GPT-4-class capability fell from about $30 per million input tokens in early 2023 to below $0.30 by late 2024, a hundredfold drop in under two years.

Everyone who uses AI gains from those falling prices. The companies building its infrastructure are in a harder position. By my estimate they have spent about $1.5 trillion on chips, data centers and power since ChatGPT launched in November 2022, and they are adding $800 billion to $1 trillion a year.

When I ran the numbers on data centers in space earlier this month, readers filled the comments and my inbox with the news that I was trying to hold back the future, and several pointed out that Elon Musk is smarter than I am. He may well be. I have written about AI for fifteen years and still take on AI advisory work on occasion. I expect AI to change how most of us work, and my question here is narrower: can the companies building AI’s infrastructure earn back what they are spending?

Most of the inputs in what follows come from the AI companies’ own filings and statements, from the industry’s more optimistic forecasters, or from US government statistics, and the arithmetic is simple enough to check. If my conclusion is wrong, then one of those inputs, my arithmetic or the way I have framed the question is wrong, and I would like to see the better number. If my numbers and framing are correct, however, and you don’t like the conclusion, your issue is with mathematics, not me.

Whether the builders earn their money back depends on how much of the value AI creates they get to keep. In the long-run US data, the buyers of new technology have kept most of the gains, although the split varies with how competitive a market is. The fiber boom of the late 1990s shows what that means for the builders. The cable laid then ended up carrying the internet, and Global Crossing, one of the companies that laid it, went bankrupt in 2002. Harris Kupperman titled his analysis of AI data centers Global Crossing Is Reborn.

What Others Have Calculated

Aswath Damodaran, who teaches valuation at NYU’s Stern School of Business, made the clearest version of the skeptical case I have heard, in a September episode of the BiggerPockets Money podcast. He called the build-out the largest factory in history, built before anyone knows what it will produce, and estimated that more than $2 trillion has gone into it. On his most optimistic numbers, AI products and services bring in about $250 billion a year.

In an August post on his Musings on Markets blog, Damodaran capped AI’s market at what companies now pay people for the work AI would take over, which he puts at $12.96 trillion of US employee compensation in 2025. That figure matches BEA’s wages and salaries alone; with benefits, US employee compensation was about $15.7 trillion in 2025 and now runs above $16 trillion a year. On the podcast he asked who would be left to buy what AI-run companies sell if AI took a large share of that pay. Luka and I raised the same objection in 2016 against forecasts of mass unemployment: a small elite that automated everyone else’s job would have no market left to sell to.

Sequoia Capital’s David Cahn has run the revenue-gap arithmetic since 2023. In his June 2024 update, AI’s $600B Question, he started from Nvidia’s data-center revenue and doubled it to cover the rest of each data center. He doubled it again for the margin AI companies need and compared the total with what AI actually earned. Bain & Company put the requirement at $2 trillion of annual revenue by 2030. JPMorgan calculated that a 10% return on the build-out through 2030 takes about $650 billion a year in perpetuity, the equivalent of $34.72 a month from every iPhone user.

Four authors have since expressed the requirement in terms of labor or of capital deployed.

Daniels counted only the equity invested in AI companies, not the data centers. Ptarmigan and EY both measured against white-collar pay. EY also argued that companies cannot pay for AI across the economy through layoffs alone, because lower wages mean less spending by the customers those companies depend on.

Paul Kedrosky and Alex Preston of Man Group argue that AI tokens are becoming a commodity whose price keeps falling, and that model makers competing on price are badly placed to earn lasting returns. Kupperman, Ben Thompson and Goldman Sachs have each written about how short the useful life of an AI chip is.

I organize the question around one explicit assumption, the share of the value AI creates that ends up as AI companies’ revenue, and set it beside the evidence William Nordhaus assembled in 2004 on who keeps the gains from innovation.

What the Build-Out Has to Earn

What Has Been Built, and How Fast

Damodaran counts $1.7 trillion of cumulative capital spending at just six companies: Alphabet, Amazon, Meta, Microsoft, Oracle and CoreWeave. Some of that went into ordinary cloud computing, and a good deal of AI spending happened elsewhere, from Chinese cloud providers to government programs such as the EU’s AI gigafactories. My estimate of cumulative AI-related capital spending worldwide since late 2022 is about $1.5 trillion, within a range of $1.1 trillion to $2 trillion, depending on how much of the hyperscalers’ spending one counts as AI. It is not a reported industry total. Goldman Sachs economists, using a different method, expect cumulative AI investment since 2022 to reach about $1.8 trillion by the end of 2026.

Companies are still spending more each year, though the published figures measure different things. PwC expects global data-center capital spending of roughly $800 billion in 2026, rising to $1.1 trillion in 2030. Goldman Sachs’s baseline scenario puts AI capital spending at $765 billion this year and $1.6 trillion in 2031, and the Goldman economists’ broader count of global AI-related investment comes to about $1 trillion this year. Nvidia, the largest single supplier, reported $89 billion of data-center revenue in the quarter to July 26, a measure of what suppliers sell rather than of total AI investment.

The 50-Cent Rule

Under four assumptions, the companies that own AI infrastructure need about 50 cents of AI revenue a year for every dollar they invest. I treat the industry as one chain, from data-center owners to the labs and application companies, and count revenue once, when an end customer pays. In McKinsey’s forecast of $5.2 trillion of AI data-center spending by 2030, 60% goes to chips and computing hardware, which last about five years. The buildings, power and cooling that make up the rest last about 20. Investors expect a return of about 10% a year, which I treat as a simple annual charge on the money invested. And about half of each dollar an AI company earns goes to power, staff, research, sales and tax, which leaves the other half to pay for the infrastructure and the return on it.

$$$\frac{\dfrac{60%}{5\ \text{years}} + \dfrac{40%}{20\ \text{years}} + 10%}{50%} = \frac{12% + 2% + 10%}{50%} = 48%\ \text{of the investment, every year}$$$

The result moves with the assumptions. If 70% of each revenue dollar were left for the infrastructure, the ratio would fall to about 34 cents; at 30%, it would rise to 80 cents. A standard annuity calculation at the same 10% gives about 41 cents. GeometricInvestor independently estimated 0.35 to 0.60 times deployed capital, a range that includes my 0.48.

At 48 cents on the dollar, the $1.5 trillion already spent sets an annual hurdle of about $0.7 trillion of AI revenue. Justifying spending at today’s pace requires more. Applying Sequoia’s second doubling, for the margin, to $800 billion a year of spending gives $1.6 trillion of annual revenue. Applying Bain’s 2030 ratio, capital spending at about a quarter of revenue, to today’s spending gives $3.2 trillion. Forty-eight cents on the $5.2 trillion McKinsey expects by 2030 comes to $2.5 trillion. These methods answer different questions, so I treat $2.5 trillion a year as a working scenario for the expanded build-out, not a forecast.

By my count, AI companies earn somewhere between $150 billion and $250 billion a year from AI products and services at current run rates, a range that mixes reported run rates with estimates and leaves out the rental of computing capacity between AI companies. Anthropic’s run rate reportedly passed $65 billion at the end of July, and OpenAI’s reached $40 billion, while Damodaran puts the whole market at $250 billion at most. Run rates annualize a recent month and are not audited revenue. They also move fast. In September, Axios, citing The New York Times, reported that Anthropic’s pace had passed $100 billion, up 50% in two months, so my range is an August snapshot. The range can also count a sale twice when one AI company resells another’s models. It leaves out AI’s returns inside existing businesses, such as better ad targeting, which help pay for the same data centers. On the August range, AI revenue has to grow roughly three to five times to clear the hurdle on what is already built, and 10 to 17 times to support today’s pace.

The Bill in Jobs

Asked on Bloomberg TV in late 2025 whether the AI investment could pay off without destroying jobs, Geoffrey Hinton said that “to make money you’re going to have to replace human labor.”

I convert revenue into jobs using a $100,000 job as the unit. US civilian employers paid an average of $49.46 per hour worked in wages and benefits in June 2026, about $99,000 for a 2,000-hour year, which I round to $100,000. The United States has 159 million nonfarm payroll jobs, and employers pay their employees more than $16 trillion a year, so each job count below is also, roughly, a share of the entire US payroll.

The counts are job-equivalents, not layoffs. An AI system that lets existing staff produce an extra $100,000 of net value is worth as much to the AI company as one that replaces a $100,000 employee. What changes the count is how much of that $100,000 the employer hands to the AI company and how much it keeps.

In a pure cost-saving deal, no employer would accept the top row, because it would keep none of the savings. At a 50/50 split, the hurdle on what is already built equals the pay of 14 million jobs, more than the 12.64 million jobs in US manufacturing.

The bottom row comes from Nvidia’s chief executive. On the BG2 podcast, Jensen Huang said he would give a $100,000 employee a $10,000 AI that doubles or triples their productivity “in a heartbeat.” If a worker’s pay measures the value of their output, and the AI doubles that output, the AI company collects a tenth of the value it creates; if the AI triples output, a twentieth. At a price of a tenth of each worker’s pay, AI companies would need their tools in the hands of workers paid a combined $25 trillion a year to justify today’s pace of building, one and a half times what every US employer pays every employee.

The United States is only the yardstick. AI revenue comes from every country, but pay is lower in most of them, so the same revenue requires more jobs elsewhere.

Goldman Sachs economists wrote in August 2025 that AI could displace 6% to 7% of the US workforce if it is widely adopted, within a range of 3% to 14%, and they doubt it will cause large employment reductions over the next decade. Challenger, Gray & Christmas counted 116,175 announced US job cuts for which employers cited AI in the first eight months of 2026, and about 188,000 since it began tracking the reason in 2023. Those are employers’ stated reasons for announced cuts, not measured displacement, and layoffs are only one channel, but the total comes to roughly 1% of the 14 million job-equivalents in the 50/50 row. Andy Challenger’s summary of July was that “AI is shifting the labor market, it is not dismantling it.”

The Share AI Companies Have to Keep

The rows of the jobs table differ in one assumption only, the share of the value that ends up as AI companies’ revenue. I can turn the calculation around and ask how large that share has to be, using published estimates of the value AI will create.

McKinsey estimated in 2023 that generative AI could add $2.6 trillion to $4.4 trillion a year across the 63 uses it analyzed, and $6.1 trillion to $7.9 trillion once its effect on knowledge workers across the economy is included. Goldman Sachs economists put the gain at about 7% of global GDP, almost $7 trillion, reached over a 10-year period. The three figures measure different things, from economic benefit to GDP, and none of them is cash available to pay AI companies, so I use them only to show the scale of the share.

At today’s pace of building, AI companies’ revenue would have to equal between 32% and 96% of the value these estimates put on generative AI, depending on whose estimate you use. McKinsey and Goldman both estimated the value at full adoption, which Goldman spreads over a decade, so the share needed in the next few years is higher. Clearing the hurdle on what has already been built requires a more modest 9% to 27%.

If those estimates measured the value AI creates for its buyers, a company paying prices high enough to fund today’s build-out would keep between 4% and 68% of it, before its own costs of putting AI to work. At the low end of McKinsey’s core estimate, the buyer would keep 4 cents of every dollar.

What Nordhaus Found About Who Keeps the Gains

William Nordhaus, the Yale economist who won the 2018 Nobel Prize, measured a related share for the US economy in a 2004 working paper, Schumpeterian Profits in the American Economy. Using data for the nonfarm business sector from 1948 to 2001, he estimated that innovators captured about 2.2% of the total surplus their innovations created. Their customers got the rest, mostly through lower prices.

He found some evidence that the share was even lower in the new-economy industries, where entry was easy, imitators followed quickly, and information cost a lot to produce and little to copy. He then applied his numbers to the dot-com boom. If new-economy entrepreneurs could keep 90% of the surplus their innovations created, their companies would be worth about $6 trillion more, close to the actual gain in their value between 1995 and 2000. At the capture rates he had measured, the figure was $400 billion. Nordhaus suggested that part of the bubble came from investors overestimating how much of the value innovators could keep.

Nordhaus measured something narrower than my share: profits above a normal return on capital. My share is the revenue AI companies need to earn a normal return in the first place, so the two numbers cannot be compared directly, and 2.2% does not belong in my table. A supplier can collect a large bill, spend most of it producing the service, earn a normal return and still leave most of the gains with its customers. Nordhaus showed which way competing suppliers push prices: toward the buyer. In the dot-com case, investors bet on the opposite.

Damodaran and Bradford Cornell described the investor side in The Big Market Delusion, a 2020 paper on how overconfident entrepreneurs and their financiers collectively overprice companies chasing a very large market, and on the correction that follows. In AI, I see the same delusion as a capture assumption. Each lab raises money as if it will price like one of the two big winners Damodaran expects, and together the labs are cutting prices like commodity suppliers.

Token Prices and Buyer Behavior

Kedrosky, Preston and their Man Group colleague Sumant Wahi, using data from Epoch AI and Artificial Analysis, calculate that the cost of AI tokens is falling by more than 70% a year. At that rate, the number of tokens sold has to more than triple every year just to keep revenue flat. These are list prices for comparable capability, not the revenue labs collect per token, which also depends on model mix and discounts. In their September paper on tokens, Preston and Kedrosky report that open-weight models trail the leading closed models by about six months. Open-weight and Chinese models approach 25% of usage in some segments, according to OpenRouter data that Man Group cites.

Enterprise buyers already treat tokens as interchangeable across a growing share of workloads, choosing on price and speed rather than benchmark scores, according to Preston and Kedrosky’s tokens paper. That matches my experience as a CTO. Once a second supplier can do most of the same job, the buyer runs a bake-off and keeps most of the savings.

Enterprises are still buying the tools they rate best in large volumes, as Anthropic’s run rate shows. Kedrosky and Preston expect the companies that own the orchestration layer, tools such as Claude Code and OpenAI’s Codex that direct tokens through multi-step work, to keep their margins as tokens get cheaper. If the labs own that layer, they can charge for the workflow even while token prices fall. If orchestration becomes a commodity as well, they lose that pricing power.

Backlogs and Lenders

Cloud companies report their contracted future revenue as remaining performance obligations or similar backlog measures, with different definitions and reporting dates. These are not bills already owed, and they include more than AI. Microsoft’s commercial obligations were $678 billion at the end of June; in January, when the total was $625 billion, the company said 45% was tied to OpenAI. Oracle’s reached $664 billion in the quarter to August, a quarter in which Oracle’s free cash flow was minus $5 billion and the company sold $20 billion of stock to fund its build-out. Google Cloud reported $514 billion, AWS $496 billion and CoreWeave $104 billion, for a combined $2.46 trillion. The companies do not disclose how much of that total the AI labs account for.

The labs’ biggest suppliers are also among their biggest investors. Amazon committed $50 billion and Nvidia $30 billion to OpenAI’s $122 billion funding round this year, and the Amazon deal came with an agreement for OpenAI to use two gigawatts of Amazon’s Trainium capacity. Microsoft and Nvidia committed to invest up to $5 billion and up to $10 billion respectively in Anthropic, which agreed to buy $30 billion of Azure capacity..

Kedrosky and Preston describe the cloud companies as pipeline operators whose returns do not depend on token prices. Fixed contracts protect a pipeline’s revenue only while its shippers can pay, and some of the largest shippers are AI labs whose own prices are falling.

If those shippers cannot pay, lenders may lose money alongside the cloud companies’ shareholders. Private-credit lenders underwrite data-center loans on 10-to-20-year asset lives while GPU technology turns over every 12 to 18 months, according to Man Group’s AI bubble paper, which expects the first wave of defaults in 2027 and 2028. The Bank of England flagged AI companies’ growing use of private credit and structured finance in its July Financial Stability Report.

Chip Lives Are Shorter Than AI’s Adoption Timeline

In Tracking Trillions, Goldman Sachs analysts call the useful life of AI chips, typically estimated at four to six years, “the single most influential variable” in the build-out’s economics. Ben Thompson argued in The Benefits of Bubbles that chips break down and are superseded, so they don’t stay in service for years as cheap, fully depreciated assets the way the infrastructure of earlier bubbles did. Kupperman estimated that the data centers built in 2025 alone would incur $40 billion of depreciation a year against $15 billion to $20 billion of revenue. Five years is an accounting life, not a physical one. New chip generations arrive every 12 to 18 months, which can cut what an older chip earns long before it stops working, although older chips can keep earning on less demanding work.

Companies adopt new technology on a slower schedule. Goldman’s economists spread the productivity gain from generative AI over 10 years, while the chips bought this year will be fully depreciated around 2031 on a five-year schedule. Reaching $2.5 trillion a year by 2030 from August’s $150 billion to $250 billion requires AI product revenue to grow by 78% to 102% a year for four years. Even from $300 billion, the growth needed would be about 70% a year. Jared Bernstein and Ryan Cummings have described the AI companies’ position as a race against time.

What Would Change My Conclusion

Five things would make a payback more likely.

• Pricing power in the orchestration layer. If enterprises lock into one lab’s agent tools, that lab can hold its prices while token prices fall. The sign to watch is lab gross margins rising while per-token prices keep dropping.
• Consolidation. Two winners with pricing discipline could collect a large share of the value. They would also take in a large share of what employers now pay workers, the political problem Damodaran warns about.
• Far larger value, sooner. Epoch AI argues that worldwide labor compensation, on the order of $50 trillion, is the prize, and that full automation would capture much of it. If AI companies automate that much work within one chip generation, the capture shares above are too pessimistic.
• Cheaper, longer-lived capacity. Longer chip lives, higher utilization and a slower build-out would all lower the annual hurdle.
• Returns inside existing businesses. Some of the spending pays for better ad targeting and recommendations at Meta and Google rather than for AI sold as a product. Those gains are hard to separate out, and they reduce what AI products have to earn on their own.

What It Means for Investors, Buyers and Policymakers

Investors should separate the companies that build AI infrastructure from the companies that use AI. In Nordhaus’s data, users kept almost all the value of new technology. Builders can still earn a normal return in that world if prices cover their full costs. Prices falling by more than 70% a year make that harder for chips already bought, whose cost is fixed, unless usage grows fast enough to make up the difference.

Technology buyers are on the other side of the same calculation. Each procurement team that switches models on price and speed lowers the share AI companies keep. A long commitment at today’s per-token prices is a bet against price declines of more than 70% a year.

Policymakers should plan for two risks that can arrive together. One is a handful of AI companies collecting a large share of what employers now pay workers. The other is losses in private credit, insurance and pension portfolios if part of the build-out is written down.

Who Gets Paid

Damodaran is right that the factory went up before anyone knew what it would produce. In 2016, Luka and I expected AI itself to become a commodity. The price of a given level of AI capability has since fallen a hundredfold in under two years, although frontier models still sell at a premium. What we did not work through is who pays for the factory once its output is cheap. Damodaran tests a story by asking whether it is possible, plausible and probable. On my numbers, a payback is possible, plausible if enough of the five developments above arrive in time, and not yet probable. We will know more in 2027 and 2028, when, by Man Group’s reckoning, the first data-center leases come up for renewal and the chips inside them are two generations old.