Skip to content
[ 05 ]Journal

AI Just Moved From Equity to Debt. The Collateral Is a Chip That Ages in Three Years.

SpaceX, Broadcom and Oracle are lining up tens of billions in debt to buy AI chips. The build-out is moving from equity to credit markets, backed by collateral that loses most of its value in a few years.

All writing
AI Just Moved From Equity to Debt. The Collateral Is a Chip That Ages in Three Years.

AI Just Moved From Equity to Debt. The Collateral Is a Chip That Ages in Three Years.

For most of the AI boom, the money came from people who were paid to take risk. Venture funds, sovereign wealth funds, and the cash piles of the largest companies on earth. If the bet failed, shareholders absorbed the loss.

In the past week, that changed in a way that is easy to miss and hard to undo.

On October 6 and 7, reports surfaced that SpaceX is seeking about $40 billion to buy Nvidia chips: roughly $30 billion of investment-grade bonds and $10 billion of bank loans, with Apollo expected to lead and a close targeted for 2027. The same week, Broadcom was reported to be arranging somewhere between $30 billion and more than $50 billion of debt to finance chips for OpenAI, depending on whether you read Bloomberg or the Wall Street Journal. Oracle, which reported negative free cash flow of $23.7 billion for fiscal 2026, is reportedly in talks with Apollo and Goldman Sachs on a structure where a separate entity buys the chips and leases them back.

Apollo is reportedly in the room on all three.

The numbers behind the shift

Morgan Stanley has estimated that AI infrastructure will need about $1.5 trillion of outside financing by 2028. That is money the hyperscalers cannot or will not fund from their own cash flow, and venture capital is far too small to cover it. By Crunchbase's count, the entire global venture market raised $159 billion in Q3 2026, and that was considered a strong quarter.

So the gap is going to debt markets. Reuters-cited data puts AI-related borrowing in U.S. leveraged finance at around $88 billion in 2026, up from about $20 billion in early 2025. Bloomberg reporting this week tallied close to $500 billion of AI-related debt across the sector, enough to push the cost of insuring Broadcom's and Oracle's bonds against default to record highs. Broadcom's five-year credit default swaps reportedly widened to 136 basis points on October 8–9; Oracle's traded near 261.

Those are not distress levels. They are, however, the market starting to charge for a risk it used to ignore.

Why chips are strange collateral

Lenders like collateral that holds its value. Buildings, aircraft, and power plants last decades. A loan against them can run for decades too.

GPUs do not work like that. A frontier accelerator is a top-tier asset for maybe two or three years before a new generation offers far more compute per dollar and per watt. It does not stop working, but the rent it can charge falls quickly. Some of the reported deal structures pair long-dated debt, in SpaceX's case reportedly including 30-year maturities, with hardware whose economic value will be a fraction of today's by the end of the decade.

That mismatch is manageable if the borrower's cash flow is strong and diversified, because the lender is really underwriting the company, not the chip. It gets riskier when the chips are the main asset, the main customer is a single AI lab, and the lab's revenue is still being proven. Several of these deals have exactly that shape: a chipmaker or cloud provider borrows, buys hardware, and leases it to one or two model companies whose ability to pay depends on continued growth.

We made a version of this argument in $725 Billion of AI Capex This Year: the spending is real, but the part funded by circular deals and thin customer concentration is where the bubble risk sits. Debt makes that part more fragile, because equity can wait out a bad year and lenders cannot.

It is now happening in a rising-rate world

All of this is being arranged just as the Federal Reserve started raising rates again. The September hike to 3.75%–4.00% was the first since 2023, and most officials expect another by year end. Every new tranche of AI debt will price off a higher base, and floating-rate bank loans will reset upward.

That does not mean the build-out stops. It means the cost of capital for AI infrastructure is no longer set by optimistic equity investors. It is set by credit committees, rating agencies, and insurance spreads, all of which respond much faster to bad news.

What this means for founders

The infrastructure layer is becoming a utility business. Debt-funded data centers and chip fleets need predictable, contracted revenue to service their loans. Expect the big compute providers to push harder for long-term commitments and take-or-pay contracts. If you buy compute, the best pricing will go to customers who can commit, and spot capacity may get more volatile, not less.

Avoid single-supplier dependence. If your product runs on one provider's capacity, and that provider is highly leveraged, its financial stress becomes your outage risk. Build for portability across at least two clouds or model providers now, while it is cheap to do.

There is a real business in financing plumbing. Lenders underwriting GPUs need things that barely exist yet: independent valuation of used accelerators, utilization and performance data, secondary markets for retired hardware, and monitoring that tells a lender when the collateral is losing value. Asset-backed finance created entire industries around aircraft and real estate. Compute is next.

Don't take your cue from the giants' balance sheets. SpaceX can borrow $40 billion because it has a launch business underneath. A startup that borrows heavily against hardware, without contracted revenue, is making the same bet with none of the cushion.

The honest version

Moving AI infrastructure onto debt is not automatically a sign of a bubble. Mature industries are financed this way. Railroads, telecom networks, and power grids were all built on borrowed money.

But those booms also taught a consistent lesson. The infrastructure usually turns out to be useful. The first round of lenders and builders often does not get paid. The fiber laid in the late 1990s carried the internet for twenty years, long after many of the companies that borrowed to lay it had gone bankrupt.

Build on top of the infrastructure. Be careful about being the one who borrowed to build it.

Sources

[ Next ]Keep reading