The AI Job Cuts Aren't Paying for the Machines. They're Reassuring the Lenders.

The biggest savings figure any company has attached to an AI layoff this year is PayPal's: at least $1.5 billion, promised in the spring alongside a cut of more than 4,500 jobs ^[1]. Snap, shrinking its staff 16 percent, put its number at $500 million a year ^[2]. Those are the largest payroll-savings figures in the year's announcements. The heaviest cutters, meanwhile, are buying machines priced on another scale entirely: Meta raised its 2026 capital-spending guidance to $145 billion ^[3], Oracle projects $90–95 billion ^[4], and Amazon is spending around $200 billion ^[5]. Two orders of magnitude sit between the savings and the spending. The cuts are not paying for the buildout. They are not paying for a rounding of it. So what are they for? A year ago there was a clean answer, and it had nothing to do with arithmetic. In August 2025, Garud Iyengar, a Columbia engineering professor, explained why AI makes such a convenient thing to blame for a layoff.

AI becomes the perfect non-human scapegoat: it’s inevitable, it doesn’t sue, it doesn’t organize, and it doesn’t tarnish the brand. — Garud Iyengar

The other half of his argument was market mechanics: investors pay a premium for companies that look cutting-edge and AI-driven, which creates a powerful incentive to frame layoffs as part of an AI transformation ^[6]. In 2025, that was the whole transaction. The audience for a layoff was the stock market, the currency was narrative, and no balance sheet ever had to be involved. Then the bill moved onto the balance sheet. Big Tech's capital spending runs about $740 billion this year, 69 percent more than 2025 ^[7], and it has overtaken the cash the businesses generate: at the hyperscalers — the handful of companies that run the world's cloud — capital spending now exceeds operating cash flow, and Alphabet and Amazon both posted negative free cash flow in the June quarter, spending more on the buildout than the businesses brought in ^[8]. The gap went to the bond market. Amazon sold $53.8 billion of bonds in March, the largest corporate sale on record, drawing roughly $126 billion in peak demand and lifting its total borrowings to $122 billion; Oracle plans to raise as much as $50 billion in fresh debt this year ^[5]. Meta issued $55 billion in bonds plus off-balance-sheet ventures ^[8]. What lenders did next is the part of the year the layoff announcements leave out. Torsten Slok, Apollo's chief economist, describes the buildout as borrowed money buying assets that may depreciate before they ever pay back, on balance sheets whose leverage keeps rising. The market has repriced accordingly: credit default swaps on hyperscaler bonds have widened roughly 60 basis points, each one a hundredth of a percentage point, since October 2025 — the largest and fastest such move since 2008 ^[9]. And Slok is blunt about what investors now say justifies the loans.

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

His case in point is Oracle, where cloud margins have climbed from 37 to 41.5 percent, exactly the kind of number that can be made to carry the debt ^[9]. In July, S&P turned the demand into a deadline: it cut Oracle to BBB- — the lowest rung of investment grade, one notch above junk — citing aggressive spending and an uncertain path to positive cash flow, and it warned the other hyperscalers to show returns by 2028 or face the same treatment ^[8]. Across the market, the requirement has hardened into plain language: investors now want concrete evidence of sustainable returns ^[10]. Margins, on a deadline, have one fast source. The data-center contracts are signed and the chips are ordered; revenue compounds over years; payroll is the only expense line big enough to move the number within a quarter or two. And payroll moved. More than 165,000 tech jobs were gone by mid-April, with the year projected to pass 300,000 ^[2]. Meta cut 10 percent of its workforce on May 20 while raising that same spending guidance ^[3]. Oracle's restructuring eliminated more than 25,000 jobs and itself cost $2.1 billion before it saved anything, and S&P downgraded it anyway in July ^[2][11][8]. This week the company cut 546 more people from cloud infrastructure, 7.6 percent of that division, its vice president and two senior directors among them, the losses disproportionately landing on employees over 40 — and it presented the move explicitly as cost-cutting to finance the AI infrastructure push ^[4]. None of this pays for the machines. What it does is move a number the lenders watch every quarter, in the direction they now require. The substitution itself, meanwhile, loses money, and the industry's own engineers keep saying so. Bryan Catanzaro, a vice president at Nvidia, the company that makes the machines, has put his own team's arithmetic in one line ^[7].

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

An MIT study this spring found that AI automation is economically viable in only 23 percent of vision-based roles ^[7]. The freshest evidence is Oracle's own internal rollout: after 80 percent of its staff adopted the company's AI coding tools, individual coding tasks fell from several quarters to a week, but customer delivery did not speed up, a security scanner's false positives forced teams to build new verification processes, and the CIO has conceded the company might have gotten a little bit of sticker shock on the model bills ^[12]. Whatever the payroll line saves, the machines cost more than the people they were bought to replace. No company has said it is cutting payroll to keep its creditors happy — what is true either way is that the cuts are functioning as margin evidence, intended or not. The industry, for its part, keeps conceding that the AI framing was never really about the books. Sam Altman has said nearly every company doing layoffs blames AI, whether or not that is actually what the layoffs are about ^[13]; Nvidia's Jensen Huang has said he hates the practice, and Altman's name for it is "AI washing" ^[14]. Epic's Tim Sweeney simply ran the frame in reverse.

Since it's a thing now, I should note that the layoffs aren't related to AI. — Tim Sweeney

Whether there is another way to sell the AI story to investors is being tested right now by a cohort running the opposite experiment, and the prices of that bet are public. Marc Benioff is hiring 1,000 new graduates and interns at Salesforce while Meta lays off thousands, and says the AI boom is the reason to add them ^[15]. The European Central Bank's March survey found that firms using AI most heavily tend, on average, to add workers rather than shed them, and the bank expects no investment-driven pause in hiring over the coming year ^[16]. The stock market, meanwhile, is still paying the 2025 reward: Recruit Holdings, which owns the job site Indeed, has watched its stock more than double after Indeed demonstrated that AI adoption could lift profits without eliminating the job-matching business — a stock investors had first knocked to a two-year low on the fear that AI would do exactly that ^[17]. The old reward never went away. It just stopped being the only one. The argument has a deadline now, and one company is carrying both sides of it. Oracle holds $664 billion in non-cancelable AI contracts and a share price cut roughly in half over the past year; it owes $125.3 billion against $37.1 billion of cash, and its operations burned $5.5 billion last quarter ^[18]. Guggenheim looked at that ledger this month and kept a $400 price target on the stock ^[19][18]. S&P looked at the same company in July and rated it BBB- ^[8]. One ledger, two readers: the equity market is still pricing the contracts, the credit market is pricing the margins, and the payroll line is the only number both of them watch move every quarter. When the 2028 deadline arrives, one of those readers will be wrong in public, and the rest of the industry — mid-layoff, mid-buildout — inherits the answer.