The number that changed how I read tech news wasn't $600 billion. It was six years.
Six years is the useful life several of the largest cloud companies now assign to a server in their financial statements — the period over which the cost of a rack of accelerators gets spread across future earnings. Some of them stretched that estimate from four or five years to six over recent reporting cycles. That one accounting choice is where AI infrastructure investment economics stops being a story about the future and becomes arithmetic you can check yourself. Stretch the assumed life, and this year's profit rises without a single new customer. Shorten it, and the same machines start eating the income statement.
You've watched a company announce a multi-billion-dollar data center program on a Monday and a five-thousand-person reduction on a Thursday, and concluded that one of the two must be a lie. Neither is. They are the same decision, filed under two different line items.
Why do companies cutting thousands of jobs keep raising their AI budgets?
Because capital spending and payroll come out of the same pocket, and only one of them can be capitalized. When a company buys a building full of GPUs, the cost lands on the balance sheet and bleeds into earnings slowly, across that six-year life. When it pays an engineer, the full cost hits this quarter. A company that has promised investors both an aggressive buildout and stable margins has one fast lever available: cut the expense that shows up immediately in order to protect the earnings that fund the expense showing up slowly. Layoffs are not evidence that the AI bet is failing. They are evidence of how it is being financed.
That explains the pattern. It does not tell you which companies are actually exposed. For that you have to compare positions, and the useful comparison isn't optimists versus skeptics — every one of these firms is an optimist. It's how the bet is funded.
Three ways to hold the same bet
I'll use three criteria, and I'd rather name them up front than smuggle them in: where the money comes from, how long the obligation runs against how long the asset earns, and who absorbs the loss if demand lands below plan.
Position one: pay out of operating cash flow. Microsoft, Alphabet, Amazon, and Meta guided toward something in the neighborhood of $400 billion in combined capital spending for 2026 on their late-2025 earnings calls. That is a staggering figure, and it is mostly funded by businesses that existed before any of this — search advertising, Office seats, retail marketplace fees, Instagram. The obligation is a plan, not a contract. If demand disappoints, the consequence is margin compression, an ugly quarter, and shareholders asking hard questions on the call. These companies can slow down. That optionality is the whole asset.
Position two: borrow against a contract. Oracle is the clearest case, which is why it keeps showing up in coverage. Reporting in September 2025 put its compute agreement with OpenAI near $300 billion over five years, and its remaining performance obligations were reported above $450 billion. The capex to serve that is largely debt-funded. The asset earns only if a small number of counterparties keep paying on schedule for years. Credit analysts responded the way credit analysts do — outlook revisions, wider spreads on the company's debt — and cuts across its cloud organization were reported repeatedly through 2025 and into 2026, with one later account putting the cumulative total near 21,000 roles. Treat that headline figure as reported rather than settled; treat the direction as unambiguous.
Position three: borrow against the hardware. The neocloud tier — CoreWeave, Nebius, and the others renting capacity by the hour — has funded expansion with loans collateralized by the GPUs themselves, at coupons reported in the double digits. The problem is not the interest rate. It's that the collateral is a depreciating asset whose resale value depends on the same demand curve the loan assumes. When the collateral and the cash flow are correlated, that isn't collateral. It's leverage wearing a costume.
| Funding source | Obligation vs. asset life | Who eats a demand shortfall | |
|---|---|---|---|
| Cash-flow buyers | Existing ad, cloud, retail profits | Flexible — spending can be paused | Shareholders, via margin |
| Contract borrowers | Debt against multi-year customer commitments | Fixed debt, concentrated revenue | Bondholders, then equity |
| Hardware borrowers | Debt secured by the GPUs | Fixed debt, collateral falls with demand | Lenders, quickly |
Read the table twice and the verdict arrives on its own. The split isn't between companies that believe in AI and companies that don't. It's duration matching — whether the money is due back before the machines have finished earning. Group one can be wrong about timing and survive it. Groups two and three have to be right about timing, because their obligations mature on a calendar and revenue has to show up on the same one.
The friction that shows up before the demand does
Here is the part that surprised me most, because it doesn't come from Wall Street. It comes from utility regulators.
Across a growing list of states, public utility commissions have approved large-load tariffs that require data centers to commit to paying for a minimum share of the power they reserve, sometimes for a decade, sometimes with collateral posted up front and exit fees if the project dies. Ohio's commission approved a version of this in 2025 over objections from the industry. The regulators' reasoning is not anti-technology and not hard to follow: if a company reserves a gigawatt of capacity and the load never materializes, somebody pays for the transmission built to serve it. The tariffs answer the question of who. The answer being written into these dockets is the company, not existing ratepayers.
That matters more than it sounds. It converts a flexible plan into a fixed cost. A firm in position one can walk away from a site; a firm that has posted collateral against contracted power cannot walk away cheaply.
The depreciation schedule is the second piece of friction, and it's self-assessed. The useful-life estimate is a forecast about how fast a generation of accelerators becomes uncompetitive, made by the company that owns the accelerators. Michael Burry made a public argument in late 2025 that the industry's schedules understate real economic depreciation; Amazon had already shortened the life on part of its server fleet earlier that year and booked a charge for it. Both of those can be true at once. The point is that a number driving billions of reported profit is an assumption, not a measurement.
What to check before you believe a capex number
All of this is in public filings. It takes an evening, not a Bloomberg terminal.
- The useful-life footnote in the 10-K property and equipment note. Note the number and whether it changed from last year.
- Capex divided by depreciation. A ratio far above 1 for years means reported earnings reflect an older, cheaper asset base than the one being built.
- RPO or backlog concentration. Look for language about a single customer accounting for a large share. Concentration is disclosed because it is a risk.
- Where the capex cash came from — the financing section of the cash flow statement, not the headline announcement.
- Operating lease and purchase commitments, which capture obligations that don't appear as debt.
- State utility commission dockets for the counties where the campuses are going. They're free, and they're often months ahead of the press.
What I can't tell you
Nothing here answers the question that actually decides this: whether enterprise demand for inference grows fast enough, and at high enough prices, to service obligations already signed. I don't know. Neither do the people quoting $600 billion at you with confidence.
What I'd watch instead are the tells. Backlog converting into recognized revenue on schedule, or not. Capex-to-depreciation narrowing, or not. The first useful-life estimate to get shortened rather than stretched. The first vendor-financing arrangement to be quietly restructured. And the layoffs — because if headcount keeps falling while capex guidance holds, you are watching a company fund a fixed obligation with the only variable cost it controls.
The bet isn't being placed with the money in these companies' pockets; it's being placed with the money in their next five years, and the layoffs are the first installment coming due.
