Silicon Valley is building the future on borrowed money and off-balance-sheet promises. The Federal Reserve has started counting them. Wall Street has already started collecting.
By P. Madhavendra
ICTpost USA, Global AI Intelligence
November 17, 2026
Infrastructure is financed years before its returns are visible. That gap, between the moment the money is borrowed and the moment anyone knows whether it was worth borrowing, is where every great technology boom has produced its great financial failures. On Monday in Milan, a governor of the Federal Reserve measured the gap.
Speaking at the Association for Computing Machinery’s conference on AI in finance, Michael Barr told an audience of quants and computer scientists that the seven largest AI firms will carry more than $800 billion of debt on their balance sheets by year-end. Their obligations off those balance sheets, in long-term leases, energy contracts and guarantees on the resale value of chips and buildings, now exceed $3 trillion. (Federal Reserve, Barr speech, Nov. 16, 2026)
One clarification belongs at the top, because it is the figure that will be misquoted most. The $3 trillion is not debt. It is the Fed’s estimate of the maximum nominal value of contractual commitments, reported undiscounted, that these companies would owe if every lease commenced, every take-or-pay contract ran to term and every guarantee were called. It is a ceiling, not an invoice. But a central bank publishing a ceiling that large tells you how little anyone else had been counting.
Barr was recapping the Fed’s semiannual Financial Stability Report, released Friday, November 13, whose boxed analysis “Leverage in the Data Center Buildout” runs to three pages. They may be the most consequential three pages the Fed publishes this year.
What the Fed Counted
The seven firms in the Fed’s tally are Amazon, Alphabet, Broadcom, Meta, Microsoft, Nvidia and Oracle: five hyperscalers that build and run data centres, and two chipmakers whose products fill them. (TechTimes; MLQ.ai) For most of the past decade, these companies paid for expansion out of operating cash flow so vast they never needed to ask anyone for money. That era is over.
Debt issued to finance “advanced computing” infrastructure passed $680 billion between January and October 2026, roughly five times the $130 billion raised in all of 2023. Outstanding public debt tied to advanced computing now exceeds $1.1 trillion. (Federal Reserve FSR, Nov. 2026)
The share is as striking as the sum: advanced-computing borrowing accounts for roughly 31% of all investment-grade non-financial corporate debt issuance this year, up from under 6% in 2023. Nearly one dollar in three that America’s blue-chip companies raised in bond markets in 2026 went to build or equip a data centre.
And the money is arriving through narrower pipes. What began as plain corporate bonds from pristine credits has become a mix of public and private debt, structured credit and project finance. The Fed notes that private-credit transactions make the full magnitude harder to measure.
The Promises Off the Books
For the seven firms, the Fed breaks the $3 trillion into two parts as of the second quarter: leases signed but not yet commenced, worth more than $1 trillion undiscounted, and take-or-pay contracts, residual-value guarantees and similar commitments exceeding $2 trillion. (Fed FSR via MLQ.ai) Both categories, Barr said, are growing faster than the firms’ on-balance-sheet debt.
A significant share involves highly rated giants indirectly supporting the financing of weaker firms, or underwriting public-utility investment. A hyperscaler leases compute from a developer for 15 years, and the developer borrows against the lease. It signs a power contract, and the utility borrows against the contract to build a substation. It guarantees what a start-up’s GPUs will fetch at auction, and a lender advances money against that promise.
Barr has a word for what this produces: opacity. When investors cannot see where risk sits, they sell everything that looks as if it might carry it. Complexity that is harmless in a rising market becomes an accelerant in a falling one.
The Circular Economy of Compute
The Fed now uses the phrase “circular financing” in official documents. Hyperscalers and the big GPU producers hold more than $100 billion in cumulative equity stakes in large-language-model labs and cloud start-ups, which spend much of that capital renting or buying chips and compute from the companies that invested in them. (Barr, Nov. 16)
For the investor, the stake is a rounding error. For the start-up, it is oxygen. For the supplier, it books as revenue that Wall Street then capitalises as proof of insatiable demand. The Fed’s conclusion is understated but unmistakable: for the buildout to be viable, those smaller firms must eventually replace supplier money with customer money. Until they do, part of reported AI demand is the industry financing its own order book.
The Winners
Follow the debt and you find risk. Follow the money and you find the people who have already been paid.
Start with Nvidia. The bulk of a data centre’s cost is its chips, which means every dollar of the $680 billion flows disproportionately to one company’s income statement. In August, Nvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build financing platforms aimed at mobilising more than $500 billion of third-party capital for customers buying its hardware. Jensen Huang called his chips an investable asset; Blackstone’s Jon Gray compared financing compute to mortgage lending on homes. (CNBC, Aug. 10, 2026) That is the collateral-guarantee machinery the Fed is now flagging, announced on live television as a product.
The same week the Fed report landed, Nvidia’s board accepted Huang’s resignation, effective immediately, days after launching an independent investigation; finance chief Colette Kress is interim CEO and the company reports earnings on November 17. (AFP via IndustryWeek) The architect of AI-as-asset-class left the building the week regulators started inspecting it.
Then the alternative-asset managers, who have turned data centres into the largest infrastructure trade of the decade. Blackstone, which took QTS private for $10 billion, reported a $150 billion data-centre portfolio with a $160 billion pipeline as of the first quarter. (Savrn capital atlas) Apollo’s MidCap Financial lent $3.5 billion in January to finance compute that was then leased to xAI, in what Apollo called the largest private financing ever executed. (Apollo, Jan. 7, 2026, via Americans for Financial Reform) KKR says it has committed roughly $34 billion of equity to digital infrastructure plus $20 billion to power. (Commercial Property Executive) Blackstone, Apollo and KKR together put $5.34 billion into Williams Companies’ behind-the-meter power projects serving data-centre load.
Utilities are the quieter winners, booking decades of contracted demand underwritten by the world’s best credits. And the banks that arrange all of it collect fees on the way in whether or not the capacity ever fills.
Here is where the Fed’s report turns from a Silicon Valley story into a banking story. The largest US bank holding companies have close to $1 trillion of exposure to large technology firms, data-centre developers, advanced-computing special-purpose entities, private-credit vehicles and utilities, exposure the Fed says it now evaluates in its annual stress test. (Fed FSR via CASRAI; FourWeekMBA) The riskier subordinated and equity tranches of these deals sit with high-yield bond funds, life insurers and business-development companies. The winners, in other words, have distributed the downside with care.
1999, With Better Cash Flow
The Fed’s report compares hyperscaler capital spending and debt since 2020 against the entire telecommunications industry during the broadband buildout that began in 1995. The hyperscalers have not yet exceeded their comparators. The current acceleration suggests they soon will. (Barr, Nov. 16)
The difference is that today’s borrowers have businesses. WorldCom borrowed against a future that had not arrived; Microsoft borrows against a present that already generates hundreds of billions in cash. That is why Barr said he worries less about direct credit losses than about leveraged investors. With hedge-fund leverage at its highest since the Fed began tracking it and risk appetite high, a sharp repricing could trigger rapid deleveraging and what he called “self-reinforcing fire sales,” transmitted through the prime-brokerage desks of the largest banks.
Markets have noticed. In the Fed’s survey of market contacts, 87% named AI as a salient risk to financial stability, up from 50% in the spring, the most-cited risk on the list. (Reuters, via AOL)
Barr reached past telecom to railroads in the 1870s, highways in the 1950s and fibre in the 1990s, citing a Brookings paper by Columbia’s Stijn Van Nieuwerburgh suggesting the AI buildout could rival or exceed all three. (Van Nieuwerburgh, Brookings, Fall 2026) Each felt overbuilt at some point. Each left infrastructure we are glad to have. And each wiped out a generation of the investors who financed it.
The Bill Comes Due in the Bond Market
The AI revolution may well be everything its champions claim. Barr himself believes it will raise productivity and living standards. The Fed’s report does not dispute that the technology works, that demand for compute is real, or that the largest builders have the cash to weather a storm.
What it disputes is the comfortable assumption that building the future is the same as paying for it. The chipmaker has been paid. The asset managers have been paid. The banks have been paid. What remains is $3 trillion of promises resting on the proposition that the customers will arrive before the leases do.
Silicon Valley is building. Wall Street has collected its fees. The Federal Reserve has started counting.
The next test of artificial intelligence will not be run in a laboratory. It will be run in the bond market, and the examiners have taken their seats. editor@ictpost.com
