The contradiction sits in one line of Alphabet's quarterly accounts. The company generated record revenue and negative free cash flow at the same time, spending twice as fast as it collected, with capital expenditure doubling to $44.9 billion in the quarter while cash left the building rather than piled up (FinanceBuzz, Aug 18). Moody's Ratings did what rating agencies do when cash statements stop matching income statements: it put six names on notice, Alphabet, Amazon, Microsoft, Meta, Oracle and CoreWeave, warning that the AI buildout is eroding free cash flow and loading balance-sheet risk onto companies that investors had treated as untouchable (Forbes, Jul 23). Apple, notably, sat the debt party out.
The actors want different things, and that is where the friction lives. Alphabet and Amazon want share of enterprise AI before anyone else locks in the customers, so they build first and monetize later. Oracle wants to become the compute landlord of the boom, and its fiscal 2026 capital expenditure hit $55.66 billion against a target of $50 billion, overshoot included, with guidance for as much as $95 billion next year (Cryptonomist, Aug 17). Meta wants enough GPUs to make its models competitive and borrowed $30 billion in bonds in October 2025, its largest sale ever, plus separate financing for a $12 billion Texas data center to get there (Cryptonomist, Aug 17). CoreWeave has no ad business or cloud franchise to fall back on; it is leveraged compute all the way down. The rating agencies want these debts repaid in scenarios that include an AI revenue slowdown, which is precisely the scenario none of the borrowers is budgeting for.
The trigger was Moody's paper. The pressure underneath is older: depreciation is arriving on the schedule while AI revenue arrives on hope. The four big US hyperscalers bought roughly $434 billion of property and equipment in the four quarters through March 2026, against about $149 billion of depreciation over the same span (Silicon Analysts, Jul 10). That gap means the cash cost of the buildout lands years ahead of any revenue those assets might earn, and Moody's expects combined capital spending from the six flagged companies to reach around $785 billion this year (Remio, citing Moody's, August 2026). A company can survive thin margins. It cannot survive indefinitely spending more than it collects while also servicing new creditors.
The debt numbers have moved fast enough to spook even the sell side. The five issuers, Microsoft, Alphabet, Amazon, Meta and Oracle, had sold $159 billion of bonds by mid-June, up 47 percent from all of 2025, and the group's investment-grade issuance reached a record $182 billion by mid-July including related names (Cryptobriefing, Jun 12; Yahoo Finance, Jul 18). Goldman Sachs calculates that hyperscaler leverage ratios doubled from 0.9 times to 1.8 times in roughly six months, and that the volume of new supply the bond market can absorb without flinching has shrunk from about $75 billion to $25 billion (Benzinga, Aug 2026). Investors are already charging more: spreads on bonds from Meta, Alphabet and Amazon are widening as buyers demand extra return for the added risk (Cryptonomist, Aug 17).
The consequences run downhill in a straight line. First, the weakest name gets repriced: S&P Global cut Oracle to BBB-, one notch above junk, in July 2026, and Oracle's five-year credit default swaps traded at multi-year highs (Cryptonomist, Aug 17). Second, the strong names lose the cheap-money subsidy; every new data-center dollar costs more to finance, so projects that penciled at 4 percent borrowing fail at 6. Third, the pressure reaches the government bond market itself. Yields on 30-year Treasuries touched a 20-year high near 5.27 percent this week, with the hyperscalers' long-dated borrowing spree named among the causes alongside federal deficits (Reuters via AOL, Aug 21; Mogazmasr, Aug 18). The builders are competing with the Treasury for the same thirty-year money, and both are losing to the yield.
History offers one clean analogue: the telecom buildout of the late 1990s. WorldCom and Global Crossing borrowed tens of billions to lay fiber on the assumption traffic growth would outrun interest costs, and when revenue disappointed the debt decided who survived; the fiber they laid, ironically, became the internet's bargain basement. What differs this time is the quality of the borrower: these are companies with advertising annuities and cloud franchises, not startups with projected cash flows, and their networks rent out by the hour rather than waiting for dial-up demand to arrive. The counter-example cuts the other way: Comcast and the cable operators also debt-funded a network buildout through the 2000s, carried the leverage through the financial crisis, and emerged owning the pipes everyone else needed. Debt kills slow builders with weak demand; it crowns patient ones when demand shows up.
So the question is whether AI demand shows up fast enough. There are genuine green shoots, Google Cloud growing 82 percent among them (24/7 Wall St., Aug 17), but the cash math is unforgiving: Barclays expects hyperscaler capital spending to consume 85 to 90 percent of combined operating cash flow between 2026 and 2028 (Barclays, cited August 2026). And beneath the visible bonds sit the invisible commitments: Moody's counted roughly $662 billion in signed-but-not-yet-commenced data-center leases across the sector back in February, obligations larger than the companies' entire on-balance-sheet debt at the time (Moody's, Feb 2026).
Who pays if this goes wrong? Not the executives who approved the buildout, whose compensation is equity and whose options expire long before the depreciation does. Bondholders pay, along with anyone holding index funds heavy in technology credit, and so do the towns hosting half-built data centers whose construction jobs evaporate when financing costs kill phase two. Who profits meanwhile? Nvidia and the chipmakers collect cash up front, the bond desks collecting fees on each mega-deal profit regardless of outcome, and utilities signing twenty-year power contracts win whichever way the models perform.
The judgment the evidence supports is uncomfortable: the market spent two years pricing AI risk into Nvidia's earnings and almost none of it into the balance sheets funding the chips. Moody's has now moved the question from the earnings call to the credit file. Companies that burn their own cash choose their own speed; companies that borrow have their speed chosen for them, and the lender's terms get stricter every week the revenue lags.
Companies that burn their own cash choose their own pace; companies that borrow have their pace set by people who charge more every month the revenue lags.
Method. This analysis rests on the sources cited below. ARCANE does not publish a proprietary universe, cohort weighting or exclusion list for this piece — the reading is the desk's, argued from the record, not a screened back-test.