The asset-light illusion
The word comes from "hyperscale", meaning computing built to grow by adding many identical machines rather than upgrading a few large ones. AWS, launched in 2006, is the usual origin story: fleets of standardised servers, management automated in software, capacity added in big blocks as demand arrived. Microsoft and Google followed the same playbook, and the label stuck to whoever was operating cloud infrastructure far beyond the scale of ordinary enterprise IT.
The "asset-light" reputation was always a trick of perspective. Cloud was asset-light for the customer, who stopped buying servers. Somebody still had to buy the servers, and that somebody was the provider, spreading enormous upfront capital costs across a very large book of usage-based demand. For a decade the demand grew fast enough that the arithmetic flattered them and nobody looked hard at the denominator.
Then came the sentence that broke the spreadsheet. On Alphabet's July 2024 earnings call, Sundar Pichai said "the risk of underinvesting is dramatically greater than the risk of overinvesting for us here" (CNBC), and it has functioned ever since as a permission slip for the entire industry. It is a defensible thing for a CEO to say when a genuine platform shift might be underway. It is also, when five companies say it simultaneously and none of them can afford to be the one that blinks, a description of an arms race rather than a . See for what that race now costs.
The gigawatt buildout
The four largest spent roughly $328bn of capex in 2024 and roughly $410bn in 2025. For 2026 Microsoft has guided to about $190bn, Alphabet to $175–185bn (raised in July 2026, on some reports, to $195–205bn), Amazon to about $200bn, and Meta to $130–145bn. Add the guidance up and you are somewhere near $700bn in a single year from four companies. Amazon's number is genuinely uncertain: some post-earnings reports in July 2026 put its target at $220bn rather than $200bn.
Capex on that scale stops being a financial abstraction and becomes concrete, steel and electricity. Microsoft signed a twenty-year power purchase agreement with Constellation to support restarting the former Three Mile Island Unit 1 reactor, now the Crane Clean Energy Center, at a reported 835 MW, with commercial operation targeted for the second half of 2027. Amazon's arrangement at Talen's Susquehanna plant supports a co-located campus sized at up to about 960 MW, though a 2025 SEC-filed agreement is described as covering 1,920 MW for AWS operations in Pennsylvania. OpenAI's Project Camellia in Effingham County, Georgia, is a 3.2 GW campus across roughly 1,400 acres, with Georgia Power delivering the power in phases from 2028 through 2032 and local reporting putting initial investment at at least $20bn.
And the queue keeps forming. NRG said in August 2026 it was aligned on principal commercial terms with an unnamed hyperscaler for a 1.2 GW combined-cycle gas plant in Texas. In July 2026 Hut 8 signed a $9.8bn, 352 MW lease with an unnamed hyperscaler at its 1 GW Beacon Point campus. One 2026 tracker counted roughly thirteen announced projects committing over 9.8 GW of nuclear capacity to AI data-centre infrastructure. This is not an industry buying servers. It is an industry buying power stations, or paying somebody else to restart them.
The telecom parallel
The last time American companies dug this deep a hole for a demand curve that had not arrived yet, it was fibre. US telecom companies issued more than $500bn in new bonds between 1996 and 2001, and spent roughly $500bn burying cable, financed with debt, speculative equity and generous vendor financing from the equipment makers who wanted the orders.
The ending is documented. Global Crossing spent about $15bn building fibre networks and filed for bankruptcy on 28 January 2002 with $22.4bn of assets against $12.4bn of debt. WorldCom filed on 21 July 2002 listing about $107bn of assets, the largest US bankruptcy to that point. Lucent, Nortel, Cisco and JDS Uniphase watched demand evaporate; JDS Uniphase posted a fiscal 2001 loss near $50bn.
The honest difference, and the strongest argument the bulls have, is that the fibre carriers were building against revenue that did not exist. The hyperscalers will earn something in the region of $1tn of operating profit this year, as we noted on 18 August 2026. That is a real counterweight and not a small one. It is also exactly what makes the leverage question interesting rather than obvious: the cash flow is real, and the commitments are bigger.
Why it matters to this crash
On 9 September 2025 Oracle reported remaining performance obligations of $455bn, up 359% year on year (Oracle), with $317bn of the increase reported as coming from an OpenAI cloud agreement. That is contracted future revenue, not debt. It is also a promise to build and power an enormous amount of infrastructure for a single counterparty, disclosed as a backlog figure. The widely-repeated "$300bn Oracle-OpenAI deal" headline appears to be extrapolated from the RPO jump rather than a confirmed contract size. Nobody outside the two companies knows the actual terms.
That pattern — the biggest obligations living outside the balance sheet — is the story. Nikkei went through the filings of Alphabet, Microsoft, Amazon, Meta and Oracle and put their off-balance-sheet obligations at $1.65tn, more than the roughly $1.35tn of debt they actually report, and eight times the level of four years earlier. Meta's is around $420bn, nearly triple its recorded debt; Oracle's reached $273.3bn at the end of May, a more-than-thirtyfold rise in four years. We wrote this up on 18 August 2026. None of it is improper: a data-centre lease is not a liability until the asset is available for use, and GPUs ordered under long-term purchase commitments are not liabilities until delivered. But that means at these companies is not a forecast, it is a delivery queue. Goldman analysts put hyperscaler lease commitments at about $1.5tn, against roughly $200bn five years ago.
The part that reaches everyone else is the funding. Hyperscalers and their AI partners raised about $108bn of bonds in 2025 and about $194bn in 2026 through late July. Meta sold $30bn on 30 October 2025 — the biggest corporate offering of that year, reportedly the largest non-M&A investment-grade deal on record at the time, drawing about $125bn of orders (Bloomberg). Amazon has raised about $53bn in 2026 across a $37bn US offering and more than $16bn of euro paper.
All of that has to be bought by someone. Investment-grade companies sold nearly $1.5tn of bonds in 2026 through August, up 36%, and Nomura estimates the roughly $200bn from the biggest tech firms alone equals about 25% of the Treasury's net note and bond issuance to private investors — five times the 2025 share. "Whoever's issuing, be it a government or a hyperscaler," Nuveen's Tony Rodriguez told Bloomberg, "is now competing with more borrowers. And therefore yields have to be higher." The 30-year Treasury yield touched 5.33% on Tuesday 18 August 2026, the highest since 2007. That is with a new author.
What would make this dangerous
The equity market splitting the complex in two, and the split widening. It has already started. Over the twenty trading days to 19 August 2026, Microsoft was up 24.1% and Amazon up 8.6%, while Core Scientific fell 20.8%, Talen 14.6% and Meta 12.9%; over five days fell 15.6% and Nebius 13.6%. We described this as the market sorting the complex by balance sheet rather than losing faith in AI, on 19 August and 18 August. Meta is the one hyperscaler in the losing column and the one leaning hardest on structures that keep borrowing off its own books, including a Louisiana campus vehicle reported at more than $27bn. Watch whether that discrimination reaches the companies still funded out of operating cash flow.
Investment-grade paper having to pay junk yields to clear. In August 2026 a Blackstone-backed QTS "Project Odyssey" bond — $3.9bn, five-year, rated Baa3/BBB−, funding a Microsoft-linked data centre — was marketed at about 7.63%. A rating agency and a bond desk can look at the same asset and disagree that sharply, and only one of them is putting money down.
Contingent liabilities being called. Nvidia has agreed to guarantee up to $105bn of OpenAI's lease payments on an SB Energy campus in Pike County, Ohio, targeting roughly 8 GW, with the guarantee covering the first 4.25 GW phase, starting in stages between 2028 and 2030 and running twenty years. A guarantee like that may sit in a footnote as a contingent obligation rather than as debt, and it converts an OpenAI credit risk into something close to an Nvidia credit risk — from the company that also sells the chips. We covered it on 18 August 2026. Bloomberg Opinion counted $1.6tn of liabilities across Nvidia and the hyperscalers as of June 2026, nearly double two years earlier.
The rate transmission reaching hard enough to matter politically. It already has, mildly: US pending home sales fell 2.3% in July 2026 to an index of 71.2, the lowest since January, with the average 30-year mortgage rate just below the one-year high of 6.81% it hit in late July. Mortgage rates track ; the long end is being set partly by data-centre capex. If AI demand disappoints while the bonds are still outstanding, the buildout stops and the borrowing costs do not.