AI Buildout Runs on Debt Now, and a Telecom Veteran Sees 1999 Again
Meta has sold $55 billion in bonds and Morgan Stanley sees $3 trillion in AI infrastructure spend through 2028. Runway Growth's David Spreng says the 2002 telecom debt bust is the history that matters now.
By Grace Kim
4 min read

What's News
- AI companies raised $102 billion in venture funding in Q3 2026, about 64% of the global total.
- Meta sold $30 billion of bonds in October and $25 billion more in April.
- Morgan Stanley expects about $3 trillion in AI and cloud infrastructure spend through 2028, only half funded by cash flow.
- By 2002 the telecom sector carried roughly $1 trillion in debt, much of it never repaid.
- Global venture funding hit a record $679 billion through the first three quarters of 2026.
AI companies pulled $102 billion in venture funding in the third quarter of 2026 — roughly 64% of the global total — and the money building the rest of the stack is increasingly borrowed. That shift from equity to debt is the moment the telecom bubble becomes the relevant history, argues David Spreng, founder, co-CEO and co-CIO of Runway Growth Capital, who watched the last cycle from a front-row seat as a venture capitalist in the late 1990s and has spent the years since lending to growth-stage technology companies.
How much debt is funding the AI buildout?
The numbers are already jumbo-sized. Meta sold $30 billion of bonds last October in one of the largest corporate debt sales on record, then returned for $25 billion more in April, with investors reportedly demanding more compensation the second time. Amazon, Alphabet and Oracle have each borrowed tens of billions of their own. Meta's giant Louisiana data-center campus was financed with a record $27 billion bond raised by a joint-venture partner rather than by Meta itself.
Morgan Stanley expects the big AI and cloud companies to spend about $3 trillion on infrastructure through 2028, with only about half of it coming from their own cash flow. The rest, Spreng writes, has to come from somewhere — and increasingly that somewhere is the bond market and private credit.
On the equity side, the boom has hit record scale. Crunchbase data shows global venture funding totaled $679 billion through the first three quarters of 2026, the highest amount ever recorded for that period. A record 27 companies raised rounds of $1 billion or more during the third quarter, collectively accounting for about a third of the total.
The difference matters. Equity carries significant risks, Spreng notes, but it does not impose fixed repayment obligations. When investments underperform, equity investors absorb losses through declining valuations, while companies retain flexibility to pursue long-term strategies. Debt doesn't wait.
What does the telecom parallel actually show?
Telecom borrowed staggering sums against a demand forecast that was widely repeated and rarely checked: the famous claim that internet traffic was doubling every 100 days. Traffic was actually doubling about once a year — spectacular growth, but a fraction of what the borrowing assumed.
By 2002, the telecom sector carried roughly $1 trillion in debt, much of which the chairman of the FCC told Congress would never be repaid. WorldCom had entered what was then the largest bankruptcy in American history. Much of the fiber the industry laid sat dark for years, finally put to work by buyers who paid pennies on the dollar in bankruptcy, through a decade of restructurings and private mergers that rebuilt entire cap tables. Some companies survived. Many did not.
Spreng describes the AI buildout's logic as a kind of infinite regress — the "turtles all the way down" story from Stephen Hawking's "A Brief History of Time." Valuations support borrowing, and borrowing builds capacity. The capacity then becomes proof that the valuations were right, so each claim stands on the one beneath it.
Is today's demand story different?
The echo that concerns Spreng most is the demand narrative. Today's version holds that data centers are effectively leased before they are built and that the only real mistake is building too little. That may prove correct, he concedes, but it is repeated with the same uniformity the traffic forecast enjoyed in 1999 — and much of it rests on projections supplied by the companies doing the building.
One thing has changed: Anthropic is now public, and once companies of that scale report financial results every quarter, investors can check the demand story against audited numbers.
Spreng acknowledges the counterargument: the companies doing the largest borrowing may be among the most profitable enterprises ever built, with cash flows the telecom carriers never approached. A few AI companies appear to be reaching revenue scale faster than any businesses before them. And the internet really did change everything, just as the traffic really did arrive.
None of that, in his view, answers the question a lender has to ask: whether the cash arrives on the schedule the debt requires. Equity can wait out a wrong timeline in a way debt cannot.
Who survives when the music stops?
Spreng is not predicting that the AI buildout must end the way telecom did. The most durable lesson of the telecom bust, he writes, concerned capital structure, which decided who was still standing when demand finally caught up with capacity, long after the technology itself had been vindicated. He has spent years arguing that growth companies fear debt more than they should — and still believes it — but the argument has always been conditional. Debt serves borrowers whose obligations are sized to revenue they can see, and lenders who underwrite fundamentals instead of narratives. It treats everyone else harshly.
His closing standard for the borrowing stacked beneath the AI boom: every assumption may hold, but once investors begin asking what the turtles are standing on, they rarely stop until they reach something solid.
Original: crunchbase.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,582 articles