Fed chair ties higher yields to AI data-center borrowing

Federal Reserve Chair Kevin Warsh put AI infrastructure borrowing inside the rate story this week. Asked why long-term borrowing costs keep climbing, he gave three reasons, including the companies building artificial-intelligence infrastructure. "The so-called hyperscalers are out in the market raising funding," he said, according to Fortune's account of his press conference. "And so the competition for capital is real. And I think it partly explains the increase in yields."

The comment lands on a buildout that has just changed how it pays for itself. Through the first six months of 2026, Microsoft, Alphabet, Amazon, Meta and Oracle spent $329.9 billion on property and equipment, equal to 96 cents of every dollar of operating cash flow they generated, according to an analysis of the companies' SEC filings published by SimianX. In 2019 the same ratio was 36 cents. Combined free cash flow across the five fell from $239.3 billion in full-year 2024 to $14.4 billion in the first half of this year.

The gap is filled with borrowed money. Long-term debt at Alphabet, Amazon, Meta and Microsoft stood at $132.0 billion at the end of 2024 and $341.9 billion on 30 June 2026, on the same filings-based analysis; Alphabet's rose roughly ninefold in eighteen months, and the company posted its first quarter of negative free cash flow since it listed in 2004. Microsoft is the outlier, funding its programme out of cash and paying debt down.

The scale of the issuance is recent. The five hyperscalers sold $121 billion of US corporate bonds in 2025, against an average of $28 billion a year between 2020 and 2024, on BofA Securities figures cited by Fortune. Morgan Stanley put AI-related global debt at close to $236 billion by the end of May, four times the year-earlier pace, and forecasts something near $570 billion for the full year. Goldman Sachs expects the debt-funded share of hyperscaler capital spending to rise from about a quarter in 2025 toward 35% in 2027, when roughly $400 billion of issuance would sit against $1.14 trillion of spending. Those are forecasts, not results.

Credit markets have begun to price the change. The gap between credit-default-swap costs on hyperscaler bonds and on bank bonds has widened to around 60 basis points from roughly zero since October 2025, Apollo chief economist Torsten Slok wrote in a note this week. "What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets," Slok wrote. Oracle, which spent about $1.61 of capital for every dollar of operating cash flow in the first half, was cut to BBB- by S&P.

The financing strain has not shown up in public capex guidance. Seven AI builders spent $657 billion over their last four reported quarters, and every one of them raised, reaffirmed or reiterated its guidance in the most recent round, according to Supercycle's buildout tracker. Amazon guides to about $220 billion for 2026, Alphabet to $195 billion to $205 billion. Oracle told investors it delivered 850 megawatts of new AI data-center capacity and more than 300,000 GPUs in a single quarter.

That is the physical economy the financing pays for: concrete, transformers, switchgear, electricians, pipefitters and the people who operate the buildings afterward. It is also, in the Fed's own account, part of the inflation problem. Policymakers counted the expansion of AI investment among the forces keeping prices elevated, with the buildout drawing on scarce power, chips, materials and workers in an economy near full employment. New York Fed President John Williams has identified it as his primary inflation concern, according to reporting on recent meeting minutes.

So the labor the buildout consumes helps push up the rates the buildout must now pay. Whether that becomes a constraint on hiring depends on whether the capital keeps arriving at a tolerable price, and so far it has: nobody in the tracker cut a number.

Meta is hedging the labor side anyway. The company is testing robots from Watney Robotics, Kinova and ABB inside its data centers to plug in cables, reseat parts and power-cycle servers, several current and former workers told WIRED. One worker estimated a cable-swapping bot could replace up to 80 percent of some people's workloads. Meta declined to comment on the testing; spokesperson Francis Brennan said the company is investing in training and hiring, adding that "America is in the middle of its biggest infrastructure boom since World War II, and there's a major shortage of skilled workers to fill the roles; we need more workers, not fewer."

Both things are being funded at once. Meta launched a program this year to train thousands of people annually in electrical, mechanical and plumbing work at no cost, with guaranteed employment in Louisiana, Ohio, Indiana and Texas for those who finish, and it has partnered with building trades unions on apprenticeships. Two workers told WIRED they believe Meta is also building AI tooling that will let it hire lower-skilled, lower-paid staff and centralize some roles in cheaper locations such as Denver.

The next test of the financing comes in late October, when Alphabet, Microsoft, Meta and Amazon report and update guidance for the first time since Amazon named higher memory prices as a reason its 2026 capital budget rose to about $220 billion.