Mark Zandi, chief economist at Moody's $MCO Analytics, warned in December 2025 that bond issuance by the top 10 AI companies was projected to hit a record $120 billion that year, creating a leverage problem that distinguished the AI boom from the dot-com era.
The comparison has since become a fixture of the AI debate: Is this the same kind of debt story that blew up the telecom sector 25 years ago?
The short answer is no. But the long answer is more interesting, and more unsettling.
What the dot-com debt actually looked like
The dot-com era's credit story was, at its core, a telecom story. In the five years after the Telecommunications Act of 1996, telecom companies invested more than $500 billion, mostly financed with debt, into laying fiber optic cable, adding new switches, and building wireless networks. In the U.S. alone, telecom companies issued more than $500 billion in new bonds between 1996 and 2001.
Before the bubble burst, telecom companies raised $1.6 trillion on Wall Street and floated $600 billion in bonds to wire the country in digital infrastructure. The borrowers were names like WorldCom, Global Crossing, Qwest, and Level 3 Communications. Peak annual telecom capital expenditures reached about $213 billion in 2000 (adjusted for inflation), with total spending exceeding $500 billion between 1996 and 2000. At its peak, telecom capex reached 1.0 to 1.2% of U.S. GDP.
The money went into physical infrastructure: fiber optic cables, switching equipment, wireless towers. The growth in capacity vastly outstripped the growth in demand. The business projections that justified all this borrowing rested on a claim, attributed in part to WorldCom, that internet traffic was doubling every 100 days. In reality, traffic doubled about once per year, creating drastic overcapacity and stranded assets.
When the bubble burst, the carnage in credit markets was severe. The telecom industry owed a trillion dollars, "much of which will never be repaid," FCC Chairman Michael Powell testified to the Senate Commerce Committee in July 2002. Global Crossing filed for bankruptcy in January 2002 with total assets of $22.4 billion and debts of $12.4 billion. WorldCom, in its own filing, listed $107 billion in assets and $41 billion in debt. WorldCom's bondholders ended up being paid 35.7 cents on the dollar, in bonds and stock in the new MCI company.
Across the broader telecom sector, bond investors recovered just over 20% of their investments. Moody's annual default rate for speculative-grade corporate issuers peaked in 2001 at 9.98%. Telecom was the single largest source of those defaults, a position the industry held for five consecutive years through 2004.
What's different about AI debt
Zandi's observation that AI-related borrowing now exceeds dot-com-era levels is correct on the numbers. Even after adjusting for inflation, big tech companies are issuing more bonds than during the late 1990s, according to reporting by Fortune.
By the end of 2025, hyperscalers had issued about $121 billion in new debt to fund AI and data center expansion, according to Mellon Investments. UBS credit strategists projected that aggregated hyperscaler capex could top $770 billion in 2026, pushing public market debt issuance to between $230 billion and $240 billion for the year, according to CNBC.
But the structural comparison breaks down in several places. The dot-com-era telecoms were, in credit terms, weak borrowers. Companies like Global Crossing and WorldCom carried debt loads that dwarfed their cash flows. Many never generated a profit.
The AI hyperscalers are the opposite. Operating cash flow for the big five hyperscalers (Meta $META, Alphabet $GOOGL, Microsoft $MSFT, Amazon $AMZN, and Oracle $ORCL) is expected to hit $577 billion in 2025, up from $378 billion in 2023, while debt should climb from $356 billion to $433 billion, according to Bank of America $BAC research reported by Fortune. According to CreditSights, hyperscalers' ratio of liabilities-to-assets fell to 48% in the third quarter of 2025, close to 2015 levels, while the comparable leverage ratio for S&P 500 companies remained steady at just below 80%.
As a group, these are among the lowest-leveraged large borrowers in the investment-grade market. Most hyperscalers entered this period of higher debt issuance with leverage below 1x, or net cash positions, according to Breckinridge Capital Advisors. Even after billions of issuance, some of these companies will be geared 0.4 to 0.7 times, compared to an average leverage of just under three times for the U.S. investment-grade market, according to M&G Investments. This is the fundamental difference that makes a simple headline comparison misleading. The late-1990s telecom borrowers had no cushion. Today's hyperscalers are being graded on their cash flows, not on the speculative value of the data centers they are building.
Where the comparison still holds
The reassuring picture, though, has two meaningful cracks. The first is Oracle. Unlike the other AI hyperscalers, Oracle will have negative free cash flow until 2029, meaning its capex will exceed cash from operations, according to Bank of America. Oracle would need 7.4 years of operating income to pay off all debt at current levels, while investment-grade thresholds typically sit at 2.5 to 3.5 times. In November 2025, Barclays downgraded Oracle's debt to underweight, warning it could fall to BBB-, the lowest investment-grade rating before junk status.
Oracle looks less like its peers and more like the telecoms of old: a company borrowing to build at a rate its cash flows cannot support. The second is the broader ecosystem. The "neoclouds" and infrastructure-as-a-service companies financing AI's buildout through GPU-collateralized debt and private credit operate with risk profiles far removed from the hyperscalers' fortress balance sheets. Dave Friedman, an independent analyst, put this distinction plainly: the bear case is not that AI does not work, but that "the financing structures being used to build the infrastructure embed assumptions about residual values, utilization rates, and counterparty stability that are not well-supported by the data we have."
The pattern worth watching
Zandi himself drew a key distinction that often gets lost. Internet companies during the dot-com era didn't have a lot of debt; they were funded by stocks and venture capital. The telecom companies that did borrow were a different category: capital-intensive operators with high fixed costs, limited revenue visibility, and debt loads that left no room for error.
Today's AI borrowers sit between those poles. The hyperscalers have the cash flows, but the spending trajectory is steepening. Last year, the four biggest U.S. internet companies generated a combined $200 billion in free cash flow, down from $237 billion in 2024, according to CNBC. The more dramatic drop appears to be ahead, as companies invest heavily up front, which means margin pressures, less cash generation in the near term, and the potential need to further tap equity and debt markets.
That trajectory, not the current snapshot, is what makes the dot-com comparison more than a parlor game. The 1990s telecom bust did not begin with weak balance sheets. It began with capital-intensive buildouts justified by demand forecasts that proved wrong. The eventual question for AI credit is the same one that destroyed telecom credit: What happens if the demand projections are off?
The answer, for now, is that the margin of safety is wide. But it is narrowing.