The Hundredth EyeFR
AI & Jobs · Essay 2 / 5

Who Will Buy?

AI can do the work, but it won't reduce inequality. And its rapid rise is fueled by debt.

November 22, 20269 min readLire en français

By replacing their staff with AI, companies are cutting into their own customer base.

In brief
  • The white-collar workers AI replaces first are the same people who buy the cars, the vacations and the dinners out.
  • US growth now leans on spending by the rich and on AI, and more and more of that AI is paid for with debt.
  • If that bet fails, everything could give way at once, as in 2008, and the recovery could come back without the jobs.
  • I'm not against AI; what bothers me is the timing, because machines move faster than people.

At the checkout, the bill is higher than last month.

The cart is the same. At work, the raise you were promised keeps slipping. You’re doing everything right, and it still feels like the ground is moving under you.

We blame inflation, rent, energy prices. I think something more serious is brewing. AI is a revolution, and it will bring enormous value to humanity. But it is landing on societies that are already fragile. And when several crises break out at once and make each other worse, they can form a perfect storm. Here is how it could begin.

  1. The first cloudsCompanies cut staff who are also their customers.
  2. An economy on one betGrowth rests on the rich and on AI, paid for with debt.
  3. If the bet failsEverything gives way together, and the recovery comes without the jobs.
  4. Governments with no room leftDeep in debt, tempted by inflation.
  5. It has happened beforeRome, in the third century.
  6. The perfect stormWhat it could look like today.
A barometer rather than a forecast: the six parts of this piece, from the first clouds to the storm.

The first clouds: the trap

In October, Alphabet, Google’s parent company, said that thanks to its AI, its teams now ship 35% more code with 30% fewer errors.1 The statement said nothing about headcount.

There’s no villain in this story. Companies are caught in a trap much older than AI.

Think of a football stadium. One fan stands up to see better, and for a moment it works. Then the row behind stands, and the row behind that. Soon everyone is on their feet, nobody sees better, and everyone is tired.

Companies are in that stadium. If one keeps its people while its rivals replace theirs with AI, it loses ground, and its shareholders end up bringing in management that will cut. So nobody sits down first. Economists call this trap the prisoner’s dilemma: everyone would gain by sitting down together, but whoever sits down alone loses out.

The problem is that workers are also consumers. White-collar workers, the first in line to be replaced by AI, also drive consumer spending: they buy the cars, the vacations, the dinners out. A machine can do their work. It can’t do their shopping.

Each company gains by cutting staff. Together, they are cutting their own customer base.

If white-collar workers buy less, who keeps the economy going?

An economy on one bet

More and more, it’s the rich. In the US, the top 10% already account for between a quarter and nearly half of all consumer spending, depending on which figures you use.2 And they spend in step with the stock market: when their shares rise, they feel rich and spend.

Growth, meanwhile, depends more and more on AI. Harvard economist Jason Furman looked at where US growth came from in the first half of 2025. He found that 92% of it came from purchases of computers and software, largely for AI data centers, the huge warehouses full of servers.3 Not every dollar went to AI,4 but the picture holds: the US economy rests heavily on this one bet.

That bet is increasingly paid for with debt. Morgan Stanley expected AI-related borrowing to more than double this year.5 Some lenders are starting to worry. CoreWeave, a company that rents servers to AI firms, owes about $35 billion and now pays close to 15% interest a year to borrow.6 That’s the kind of rate lenders charge when they’re not sure they’ll be paid back.

Google and Amazon make enough money to pay their debts. The weak links are the more fragile companies that build and run the data centers, like CoreWeave, and those who lend to them.

At the end of October, the first gusts hit. According to the Financial Times, Blackstone told investors to expect Anthropic to get a “junk” credit rating after its stock market debut, the grade for borrowers judged risky. Anthropic’s chip project, North Star, only got its financing because Broadcom, a highly rated chip maker, guaranteed it.7 OpenAI cut the size of its funding round.8 On Friday, October 30, the Nasdaq 100, the index of the biggest US tech stocks, fell 3.5%.7

Three weeks later, the wind picked up. On November 20, the day after Anthropic’s market debut, the S&P 500 lost 3.7% in a single session, capping its worst week in more than 18 months. Nvidia fell 14%, and the Nasdaq Composite wiped out all of its gains for the year.9 Yet the money keeps flowing, on credit: the same day, Apollo and Goldman Sachs sold $44 billion of bonds to buy chips for SpaceX, and investors wanted more.10

If the bet fails

In July, the Bank of England warned that expected profits at AI-linked companies could be revised down sharply, and all at the same time.11

That “all at the same time” is the danger. On paper, a bank’s loans look varied: a chip maker here, a property fund there. But many rest on the same bet. If it goes wrong, they don’t fail one by one: first the shock hits AI’s inner circle, from chip makers to the data centers and the people who finance them, then, domino by domino, the whole economy. I think banks are more exposed than they look, even if, according to the Chicago Fed, big banks still lend relatively little to data centers and their total exposure is hard to measure.12

That was the mistake in 2008. Subprime loans, those risky home loans, had been made all over the United States, and people thought they were unrelated. When house prices fell everywhere at once, they gave way together.13 Banks stopped lending, companies cut jobs, and people who had never touched a subprime loan lost their job or their home.

There would be one big difference, though. After a crash, business eventually picks up and companies hire again. This time, they could hand the work to AI instead. The blow would land on people who are already financially stretched, and the recovery wouldn’t give them their jobs back.

Growth could come back without the jobs.

Governments with no room left

In 2008, governments cushioned the fall. They borrowed and rescued the banks.

Next time, it will be harder. According to the IMF, six of the seven G7 countries, including the US and France, already owe more than their economy produces in a year.14 Much of their revenue comes from taxing work. Fewer paychecks mean less revenue, just when more people need help.

When a government can neither raise taxes nor spend less, the easiest way out is to pay with money that is worth a little less each year. Debt is fixed in a currency, like the euro or the dollar. When prices and wages rise, the taxes a government collects rise with them, but its debt doesn’t move. So each year it weighs a little less. Inflation quietly pays off part of the public debt, and savers foot the bill: their money buys less than before.

For this to work, interest rates have to stay below inflation. That’s where central banks come in. After 2008, and again during Covid, the US Federal Reserve created money to buy securities on a massive scale, much of it government debt: its balance sheet went from about $900 billion in 2008 to nearly $9 trillion in 2022.15 The goal was to support the economy. It also made government borrowing cheaper.

20052010201520202025$2tn$4tn$6tn$8tn2008: $0.9tn2022: nearly $9tn2026: $6.7tn2008: $0.9tn2022: nearly $9tn2026: $6.7tn
Total assets of the US Federal Reserve, in trillions of dollars, 2003–2026. Source: Federal Reserve, via FRED (series WALCL).

That matters a lot, because a government almost never truly pays off its debt: it borrows again to repay what comes due. As long as rates stay low, that holds. When they rise, the debt swells on its own, and the pull of inflation gets stronger.

This has been done before. After 1945, the US and the UK cut their debt by 3 to 4% of GDP a year on average, by keeping interest rates below inflation.16 I think most governments are tempted to do it again, each in its own way, the US most of all.

History shows that when money, debt and power all wobble at once, the storm can go much further.

It has happened before: Rome in the third century

Rome went through one in the third century.

Under Augustus, the Roman silver coin was essentially pure silver. By the 260s, the coins in circulation held less than 5%.17 The emperors had kept the look of the coin and quietly emptied it, largely to pay their soldiers.

At the same time, everything else was shaking. The legions were making and unmaking emperors instead of defending the frontiers, and the Empire split in three, with a Gallic empire in the west and the kingdom of Palmyra in the east.18 Prices, meanwhile, went through the roof. In 301, Emperor Diocletian tried to cap them by law.19 But at those prices, sellers lost money on every sale: goods vanished from the markets, and the law was eventually dropped.20

A debased currency, an army that no longer defended the borders, a divided empire, prices out of control: none of these crises would have been enough on its own. Together, they nearly brought the Empire down.

Today there’s no need to debase the coins: our banknotes contain no metal, and money is created by the central bank. Money has changed form, but the risk of such a chain reaction has not gone away.

The perfect storm

I’m not against AI. It will end up bringing prices down through the abundance it can create, and new kinds of work will appear, as they always have. My problem is the timeline: machines improve in months, while people can take a generation to retrain and find new work.

Nobody can predict what the storm would look like today. We can sketch its outline.

It starts with a shock: a big AI company going bust, Asia suddenly unable to make chips, a pandemic, or a cyberattack on a scale we’ve never seen. AI-linked stocks fall together, and the rich, who carry a large share of consumer spending, start spending less.

Those who lent to build the data centers take losses. Banks lend less, companies cut costs, and AI is there to do the work of the people they won’t rehire.

Governments collect less tax and spend more to help people who lose their jobs. Their debt, already heavy, costs more to refinance. Pressure builds on central banks to keep rates low and create money. Prices end up rising faster than pay.

In a united world, the big powers could act together. In a world split into blocs, I fear each will turn inward: trade barriers, a hunt for culprits, political anger.

For white-collar workers, that would mean a less certain job, savings that shrink if they earn less than inflation, and variable-rate loans whose repayments become unaffordable. A storm doesn’t hit everyone the same way. The most fragile are hit first.

Which brings back the question in the title, now more urgent. Today, it’s the rich who carry consumer spending. But their spending follows the stock market, so it rides on the same bet. If it fails, who will buy?

You can’t stop a storm, but you can see it coming. I watch three signs: borrowing rates on AI projects, unemployment among young people and the middle class, and what central banks decide. The day all three turn red at once, this scenario will look much more likely to me.

But there’s a more hopeful side to this story, and that is where the next piece picks up. If machines take on much of the work, everyone can create and build.

Sources

  1. Alphabet, CEO remarks on Q3 2026 results, 28 Oct 2026 (company source) — https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/ (figures checked by the editor, 29 Oct 2026; no headcount given)
  2. Justin Fox, Bloomberg Opinion, via Advisor Perspectives, 7 Apr 2026 (Moody’s 45.8% for 2025; BLS Consumer Expenditure Survey 22.9% for 2024; spending “rising and falling with asset prices”) — https://hooks.advisorperspectives.com/articles/2026/04/07/k-shaped-economys-statistic-problems
  3. Fortune, Nick Lichtenberg, “Without data centers, GDP growth was 0.1% in the first half of 2025, Harvard economist says”, 7 Oct 2025 (information-processing equipment and software “accounted for fully 92% of GDP growth over that period”; Furman: “it’s probably not the case the U.S. economy would have recorded almost no expansion at all absent this buildout”) — https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist/ · Jason Furman on X, 27 Sep 2025 — https://x.com/jasonfurman/status/1971995367202775284
  4. Federal Reserve, FEDS Notes, Soto, Thieu, Allen, “The AI build-out and the economy”, 17 Jul 2026 — https://www.federalreserve.gov/econres/notes/feds-notes/the-ai-buildout-and-the-economy-publicly-available-data-to-assess-ais-impact-20260717.html
  5. Reuters, “Global AI debt issuance to top $500 billion in 2026, Morgan Stanley says”, 10 Jun 2026 (“AI-related global debt issuance to more than double to nearly $570 billion in 2026”; about $236bn issued by 31 May, “fourfold more than the same period last year”; coverage of bonds vs loans not defined) — https://finance.yahoo.com/sectors/technology/articles/global-ai-debt-issuance-top-085023868.html
  6. Axios, Matt Phillips, “Why CoreWeave’s woes matter to the market”, 16 Oct 2026 (about $35bn debt + about $16bn lease liabilities at end of June; 2032 bonds yielding “near 15%”) — https://www.axios.com/2026/10/16/coreweave-ai-borrowing-yields
  7. Financial Times, Michelle Chan, George Hammond, “Anthropic may not secure investment-grade rating after IPO, Blackstone tells investors”, 30 Oct 2026 (Blackstone guidance: “high” junk rating expected, per people familiar; Project North Star $63bn, more than $42bn guaranteed by Broadcom; Nasdaq 100 -3.5% on Friday) — https://www.ft.com/content/59383c7f-8d82-493f-bf01-d2e862a1018f
  8. Financial Times, OpenAI cuts funding round target, Oct 2026 ($30bn at $1.4tn → about $20bn at $1.25tn, per people with knowledge of the matter) — https://www.ft.com/content/7a3ad950-1b3b-4d9e-8c15-65a61ae6180f
  9. AP, “S&P 500 sinks 3.7% and Nasdaq composite goes negative for the year as tech sell-off worsens”, 20 Nov 2026 (closing prices; “worst weekly loss in more than 18 months”; Nvidia “closing 14% lower”; Anthropic -8.5% to $88.91, IPO priced at $83 on 19 Nov) — https://apnews.com/article/stock-markets-anthropic-ai-chips-boom-c4c3f4878b12989347f8a930ed813dad
  10. Apollo and Goldman Sachs, “Apollo and Goldman Sachs Announce Pricing of $44 Billion SpaceX GPU Financing”, press release, 20 Nov 2026 (“approximately $44 billion of senior and subordinated notes to fund the purchase of GPUs”; “oversubscribed and upsized as a result of strong investor demand”) — https://ir.apollo.com/_assets/_96154d71700d7cd17bf1075ac49d39bf/apollo/news/2026-11-20_Apollo_and_Goldman_Sachs_Announce_Pricing_of_44_662.pdf
  11. Bank of England, Financial Stability Report, July 2026 — https://www.bankofengland.co.uk/financial-stability-report/2026/july-2026
  12. Chicago Fed Insights, Cohen, Killen, Lau, “Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence”, Feb 2026 (“large banks are not materially concentrated (as a percentage of tier 1 capital) in data centers”; “difficult to provide a percentage of committed exposure”) — https://www.chicagofed.org/publications/chicago-fed-insights/2026/ai-tail-risk-for-banks
  13. ESMA (European Securities and Markets Authority), Bouveret, Fennell, Horri, “Model risk in CLOs”, Report on Trends, Risks and Vulnerabilities No. 2, 2020, p. 60 (“the crisis showed how default correlation — a key input in securitisation — had been underestimated in credit rating models”; “When house prices declined throughout the U.S., a large portion of mortgages that had been issued in different states…”; cites the Financial Crisis Inquiry Commission 2011 on the effects “on financial markets and the real economy”) — https://www.esma.europa.eu/sites/default/files/trv_2020_2-model_risk_in_clos.pdf
  14. IMF, Fiscal Monitor “Fiscal Policy under Pressure: High Debt, Rising Risks”, April 2026, Table 1.2, gross debt 2026 (projection): US 125.8, Italy 138.4, UK 103.6, Japan 204.4, France 118.4, Canada 110.7, Germany 64.6 % of GDP (6 of 7 G7 above 100) — https://www.imf.org/-/media/files/publications/fiscal-monitor/2026/april/english/text.pdf
  15. Federal Reserve, “Recent balance sheet trends” (total assets about $0.9tn on 27 Aug 2008) — https://www.federalreserve.gov/monetarypolicy/bst_recenttrends_accessible.htm · Federal Reserve, “Federal Reserve Balance Sheet Developments”, May 2022 (about $8.9tn; pandemic purchases of Treasury and agency securities ended March 2022) — https://www.federalreserve.gov/monetarypolicy/May-2022-Federal-Reserve-Balance-Sheet-Developments.htm · New York Fed, “Large-Scale Asset Purchases” (post-2008 Treasury purchases) — https://www.newyorkfed.org/markets/programs-archive/large-scale-asset-purchases
  16. Carmen M. Reinhart, M. Belen Sbrancia, “The Liquidation of Government Debt”, NBER Working Paper 16893, 2011 (financial repression + inflation; for the US and UK, liquidation of “3 to 4 percent of GDP a year”) — https://www.nber.org/papers/w16893
  17. Butcher & Ponting, “The Denarius in the First Century”, 2011 (Augustus’s later Lyon denarii: “essentially pure silver bullion”) — https://warwick.ac.uk/fac/arts/classics/intranets/staff/butcher/publications/butcher_ponting_complete.pdf · Marussi et al., “A Multi-Analytical Approach on Silver-Copper Coins of the Roman Empire to Elucidate the Economy of the 3rd Century A.D.”, Molecules, 14 Oct 2022 (devaluation “reached its peak with the last issues of the emperors Gallienus (260–268 A.D.)”, antoniniani “containing less than 5% silver”) — https://europepmc.org/article/PMC/PMC9607305
  18. University of Oregon, Mapping History, “Crisis of the 3rd Century: Introduction” (AD 235–285; “Legions and their commanders became more interested in ‘emperor-making’ than in ‘empire-defending’”; “the Gallic Empire in the west and the kingdom of Palmyra in the East”; presented in the text as concurrent with the debasement, not caused by it) — https://mappinghistory.uoregon.edu/english/EU/EU08-00.html
  19. Diocletian, Edict on Maximum Prices, AD 301, preamble, tr. Antony Kropff, 2016 — https://kark.uib.no/antikk/dias/priceedict.pdf
  20. Lactantius, “Of the Manner in Which the Persecutors Died”, ch. 7 (a contemporary and hostile witness: “men were afraid to expose anything to sale… the ordinance… was from mere necessity abrogated”) — https://www.newadvent.org/fathers/0705.htm
Next essay · 03 / 05

Innovation for Everyone?

AI doesn't just replace work. It speeds up everything else.

Read on →