Layoffs Don't Pay for the Data Centers
On 29 April, Meta raised its 2026 data-center budget to $125–145 billion. The stock fell 9 to 10 percent. On 20 May, Meta cut 8,000 jobs, canceled 6,000 open roles, and moved 7,000 people into AI teams. The stock closed up 0.4 percent.
The layoffs did not pay for the data centers. Meta’s severance charge was $1.18 billion, less than 1 percent of the $140 billion it will spend on data centers this year. The cuts made the quarter look disciplined enough to borrow against.
Henry Chen was one of the 8,000. He is 38, a tech lead in the Bay Area. Meta hired him in 2012 as an AV and conference-room contractor; he taught himself Python. He took paternity leave from August through late 2025. When he came back, the company was using AI much more heavily. On 20 May, he learned he was in the layoff. By the end of October, the public market had started pricing the cost.
The Receipt Is One Percent
For four years, the argument has been about whether AI models can do the work. The numbers say the work was never the question. The idea that layoffs pay for the data centers is wrong by about two orders of magnitude.
Meta reported $1.18 billion of severance in the second quarter. It folded that cost into the employee-compensation line instead of listing it separately. Meta’s 2026 data-center budget is $135–145 billion. The severance is 0.8 to 0.9 percent of that, less than a penny for every dollar of data-center spending. Amazon reported $2.7 billion of severance for 2025 against about $220 billion of 2026 data-center spending. The years do not match, but the ratio is still about 1.2 percent. Microsoft did not disclose a severance figure for either round.
The payroll savings are larger than the one-time severance charge, and still small. Three rounds at Amazon, Meta, and Microsoft removed about 42,800 jobs. At $150,000 to $250,000 per employee, that saves $6 billion to $11 billion a year. The four largest companies have guided about $720 billion of data-center spending for 2026, so the savings are 0.8 to 1.5 percent of that. Severance is one-time, so the first year costs more than it saves. Payroll was also the only large cost these companies could cut in a quarter. That is why they cut it, and why the cut was so small.
A cut that small cannot fund anything. The stock rose because investors read the cut as proactive, not as distress. Layoffs usually signal trouble. Meta attached the cuts to AI, and the market accepted that story. Without the AI story, the same cut would have looked like a company in trouble.
Zuckerberg described two cost centers at Meta: compute infrastructure and people. He said more spending on one means less for the other. Meta’s finance chief, Susan Li, made the same point. A leaner operating model would help “offset the substantial investments we’re making,” she said.
Three earnings reports in late October made the pattern clear. On 28 October, Microsoft reported $39.4 billion of cash capital spending and $18.6 billion of free cash flow. Its stock rose 4.3 percent after hours. On 21 October, Meta narrowed its 2026 data-center budget to $135–145 billion and reported negative $2.12 billion of free cash flow. Its stock fell 7 percent the next day. Alphabet reported $51.4 billion of capital spending, below what analysts expected. It also reported negative $9.1 billion of free cash flow. Its stock fell about 10 percent after hours.
None of this says data-center spending is bad. It says investors mark down spending that has to be financed outside the business. Headcount is the only large cost a company can cut before the next quarter.
The Lender Is the Bond Market
The AI companies buy computing capacity by the year and sell model output by the token. When cash flow does not cover the cost, they borrow.
Sridhar Ramaswamy, Snowflake’s chief executive, described how these contracts work in a 16-post thread on X. He compared the risk to banks and commodity traders. He also compared it to airlines, utilities, and cloud providers.
By September, the five largest cloud companies had issued $222 billion of bonds. The IMF’s October report puts their share of global investment-grade corporate bond issuance at nearly 10 percent, up from less than 1 percent in 2024.
Alphabet is the clearest example. On 22 July, it reported revenue up 24 percent and cloud revenue up 82 percent, with a 34 percent operating margin. It raised its 2026 capital-spending budget to $195–205 billion and reported negative $5.9 billion of free cash flow. The stock fell 7.1 percent the next day. Investors were no longer judging the company only by its closing stock price. They were judging whether it could repay debt. A bondholder waits years for repayment; a shareholder can sell in a day.
The promised revenue is larger than the money behind it. The IMF examined 14 disclosed deals since 2023. It found about $300 billion of equity and $650 billion of debt behind about $1.4 trillion of future AI revenue commitments. Those numbers measure different things, but the debt has a payment date. The revenue has to be earned. Arete estimates that the four largest companies will spend $1.8 trillion through 2028, about $600 billion more than they will generate in cash over the same period. If the revenue does not arrive, the borrowing gets more expensive.
Some of the money is circular. The Bank for International Settlements found that 55.2 percent of investment into AI companies between 2021 and 2025 came from other AI companies. It looked at 1,246 companies. The BIS calls this ordinary supplier financing, not fraud: inputs are scarce and information is uneven. The risk is concentration. Customers, suppliers, and investors are increasingly the same companies.
The Public Price Arrives Early
On 29 October, Alphabet missed revenue expectations and lost about $500 billion of market value. The stock closed 13 percent lower, its worst day since its 2004 IPO. The wider selloff started at the Nasdaq’s 6 October high. By 30 October, the Nasdaq was down 7.8 percent from that high, and the month had taken about $2 trillion off the Nasdaq 100. Intel fell almost 30 percent after its guidance. Nvidia fell 3.1 percent. Amazon fell 1.8 percent after raising its own spending. The Nasdaq had its worst month in more than a year.
Then the private market got a public price. Anthropic filed its S-1 on 26 October. The filing shows $16.3 billion of revenue in the first half, 14 times the same period a year earlier. It also shows an $80 billion GAAP loss, mostly a noncash charge on convertible notes. Anthropic lists $657 billion of infrastructure commitments. Of that, $498 billion is noncancelable. The noncancelable part alone is about 30 times the revenue Anthropic booked in the first half. Spread over the life of the contracts, it is roughly $70 billion a year, whether or not Anthropic uses the capacity.
Anthropic is not a failed business. In the second quarter, $11.6 billion of revenue covered $5.6 billion of computing cost. Revenue is growing faster than cost. But the $498 billion commitment is fixed, and it is larger than the revenue. A week before the filing, investors talked about a $2 trillion valuation. By the end of the week, they were talking about $1.5 trillion. A quarter of the value was gone. The public market priced the cost before depreciation showed up in the accounts.
The layoffs were the part investors could see. The IPO is where they can check the rest of the numbers.
The Bill Arrives on a Schedule
The market has started pricing the cost. The accounting charge is still ahead. A company does not deduct the full cost of a data center when it pays for it. It spreads the cost over the asset’s useful life. That yearly charge is depreciation. A 2026 purchase stays on the income statement for years. In late 2025, Bank of America estimated that revenue would grow 13 percent in 2026 while depreciation and amortization grew 33 percent. For 2027, it estimated 12 percent revenue growth against 30 percent depreciation growth. On that estimate, depreciation grows about two and a half times as fast as revenue. Capital spending is still growing faster than the depreciation it creates. Reported profit therefore looks better than it will once the depreciation catches up.
The useful life is the key number. Alphabet uses six years. Microsoft uses six, up from four in 2023. Meta uses five and a half, up from four in January 2025. Amazon uses six years, but has cut one group of assets to five. Amazon cited “increased pace of technology development” as the reason. The equipment becomes obsolete faster than Amazon’s own schedule assumes. Amazon CEO Andy Jassy described the timing on the July call. Amazon spends data-center capital two years before it can put servers into them, he said. Only then does the equipment start earning money. Michael Burry argues that the industry understates depreciation by about $176 billion from 2026 to 2028. His reason is that companies use five- and six-year lives for equipment that has a two- to three-year product cycle. The exact number is debatable; the direction matches the filings.
An unused GPU-hour is gone, like an empty airline seat. The contract still requires payment. The test is whether demand keeps growing faster than the committed cost. If depreciation growth starts to match revenue growth while capital spending keeps rising, demand is not keeping up.
The Junior Rung Goes First
A junior hire is the only person on a team whose value is mostly in the future. A junior needs two or three years of training before working independently. The salary is paid this quarter; the output arrives in 2028. A senior can produce value in the quarter they are hired. A junior cannot. None of that depends on what AI can do. If you manage a team, you have already made this trade, whether or not you wrote it down. The junior job opening a manager leaves open this quarter is a bet. The bet is that the new hire becomes productive before the equipment is fully depreciated.
Chen is an example of that ladder. The entry-level job he climbed into was open in 2012. Since January 2023, entry-level job postings have fallen 35 percent or more. A Stanford analysis looked at workers aged 22 to 25 in the occupations most exposed to generative AI. It found a 16 percent relative decline in employment. The bottom rung stopped opening.
US software developer job postings are up about 15 percent since February 2025, while all job postings fell 7 percent. Entry-level postings are down about a third from January 2023. Demand did not collapse. It moved toward candidates who can produce value sooner.
You will not find this in a financial statement. No company discloses the cost of training junior engineers against the depreciation schedule. That absence is the point.
The Server and the Salary
You hire a junior engineer. They’re useful in two or three years. You buy a server. It’s still on the books five and a half years later. Meta kept the server and cut the hire.
That’s the part I keep coming back to. The hire would have been independent in 2028. The server will still be depreciating in 2031. Meta cut the one that pays back sooner and kept the one that pays back later. Why? Because the server’s cost gets spread across five and a half years, and the salary hits this quarter.
For Chen, the 1 percent was his job. The layoffs weren’t paying for anything. They were the cost that hit this quarter.
