For twenty years the promise of Big Tech was that it barely needed money. Software copied itself for free and the profits piled up untouched. This month Moody's Ratings put its name to the end of that story: it warned that the shift from asset-light to asset-heavy business models "requires unprecedented levels of investment and capital raising" and threatens credit quality at the six companies it tracks, Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave (CNBC, Jul 25). The contradiction is live. These are still among the most profitable companies on earth, yet the people paid to judge whether they can pay their bills just flagged them.
The numbers behind the warning are large enough to feel like a misprint. Moody's projects combined capital spending across the six will hit $785 billion in 2026 and approach $1 trillion in 2027, a mark-up of $85 billion from its own March forecast (Data Center Dynamics, Aug 2026). Direct debt across the group already stands near $460 billion (Oninvest, Jul 2026). Then comes what Moody's counts but the balance sheets mostly do not: about $1.2 trillion in lease commitments, of which roughly $820 billion covers data centers not yet finished (Briefs, Aug 22). Rent promised on buildings that do not exist yet is debt wearing a costume.
Each actor wants something different, which is why the machine cannot easily slow down. Microsoft, Amazon, Alphabet and Meta want to hold investment-grade ratings while outbuilding each other in compute, because falling behind in capacity now means losing cloud customers for years. Oracle, rated Baa2 with a negative outlook by Moody's, two steps above junk, needs the buildout to work precisely because it has wagered its whole credit story on one customer's cloud contracts (Fortune, Mar 7). CoreWeave sits below all of them at Ba3 and funds GPU fleets through private debt arrangements that few outsiders can inspect (Briefs, Aug 22). Moody's wants, above all, to not be the agency that missed this; its note says investors will increasingly focus on whether any of the spending earns an adequate return (CNBC, Jul 25).
The trigger was the July research note. The pressure is older and simpler: chips wear out in four to six years while the buildings and leases backing them run twenty or more, so the industry has matched short-lived assets to very long promises (Axis Intelligence, Jul 31). Meta's newest deal shows the seam. It is marketing more than $12 billion of bonds through a special-purpose vehicle for a nearly one-gigawatt data center in El Paso, Texas, and buyers want yields above 7%, against a 20-year lease starting only in 2028 (Oton Technology, Aug 2026). Nine months earlier, Meta's record $27 billion data center bond priced far easier (Financial Times, reported Aug 2026). Same borrower reputation, much higher price.
The bond market has been repricing before the rating agencies moved. Of 91 hyperscaler bonds issued this year with comparable pricing, 78 traded at higher yields than their issue levels by July 28, a median rise of about a fifth of a percentage point (Reuters, Jul 28). New deals cost more too: the median discount buyers demand to take a fresh bond climbed to roughly 12 cents per hundred of bond this year from about 2 cents in 2025 (Reuters, Jul 28). Yields on long-dated hyperscaler paper over Treasuries widened from around 108 basis points' worth of gap toward 118, meaning roughly a tenth of a percentage point more compensation for the same names in a matter of weeks. None of this is distress. All of it is a toll booth going up on the road the buildout must travel.
Alphabet saw the same thing and chose shareholders over lenders. Last month it raised $85 billion in equity, upsized from $80 billion and the largest such deal on record, including a $10 billion private placement from Berkshire Hathaway (The Playbooks Daily, Aug 2026). Selling stock dilutes owners but never gets downgraded. When the most cash-rich company in the group prefers equity, it is telling you what its own treasury thinks about the price of credit ahead.
Moody's also named the strange loop inside the boom: hyperscalers invest billions in AI labs such as OpenAI and Anthropic, and those labs spend the money back on cloud services from the same investors, what the agency calls a "circular AI ecosystem" (Briefs, Aug 22). Part of the order backlog everyone celebrates is therefore money going around a loop rather than demand coming in from outside. If one lab stumbles, the revenue line of its backers wobbles too. The system shares one assumption about future demand, and shared assumptions fail together.
The historical model is the telecom fiber buildout of 1998 to 2002. WorldCom, Global Crossing and their lenders financed oceans of glass on projected internet traffic, traffic did arrive but at prices nobody underwrote, and the debt broke the builders even though the infrastructure survived to carry someone else's profits. What differs this time: the borrowers sit on hundreds of billions in annual operating cash flow, so nothing here resembles WorldCom's hollow core, and the customers include every large corporation on earth rather than a speculative startup fringe. The counterargument is real. The cloud buildout of the 2010s also looked excessive on spreadsheets and turned out to be the best-spent capital in corporate history; if AI revenue compounds the way cloud revenue did, these warnings will read like the analysts who mocked Amazon Web Services.
Follow the chain and see who pays. If returns disappoint, the first payment comes from wider spreads on the next wave of data center bonds, making each gigawatt more expensive than the last and testing Oracle first, since it carries the weakest rating of the group into its heaviest spending plan (Fortune, Mar 7). The second payment arrives when lease-heavy structures force the agencies to count off-balance-sheet rent as debt, which would push several names toward downgrade thresholds without a single missed payment. The profit takers meanwhile are the bond desks structuring these vehicles and the private credit funds collecting 7% and better on paper backed by a single tenant's signature. Risk moves; it does not vanish.
What confirms this read: further new issues pricing above 7% and concessions widening past 12 cents per hundred of bond, plus Moody's or S&P formally folding lease obligations into adjusted debt metrics. What breaks it: a hyperscaler issuing long-dated bonds at spreads back near 30 basis points' worth of premium, or a quarter where cloud and AI revenue growth visibly outruns the growth in debt, which would make the whole warning premature. Watch the next Meta or Oracle deal; the coupon is the verdict.
The judgment this piece earns: equity markets spent two years asking whether the AI spending is wise, and the credit market has finally started charging for the answer. Buildings depreciate slower than beliefs.
Rent promised on buildings that do not exist yet is debt wearing a costume.
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.