GPU Securitization Will End Badly

CoreWeave reported $640 million in interest expense last quarter. Its adjusted operating income was $128 million. The company's debt has grown to $35 billion, up from $11 billion a year ago.

I wrote in August about the structural risks of GPU securitization, the $500 billion in financing that Goldman, BlackRock, and four other firms are lining up for AI compute. That piece asked what could go wrong. Three months later, we have an answer.

It's already going wrong.

CoreWeave is the proof

CoreWeave is the largest independent GPU cloud and the first to test whether GPU-backed debt can survive contact with an income statement.

The company carries $35 billion in debt. About $13.6 billion is in delayed-draw term loans collateralized by GPUs and customer contracts. Another $10 billion sits in unsecured senior notes paying 8.5% to 9.75%. The rest is convertible notes.

Its 2031 bonds yield above 16%. Credit default swaps on CoreWeave debt trade above 1,240 basis points. CDS spreads above 1,000 are distressed territory. Bear Stearns crossed that line in March 2008.

Seventy-seven percent of CoreWeave's revenue comes from two customers. Microsoft accounts for roughly 62%. If Microsoft renegotiates or reduces its cloud commitments, the debt structure doesn't bend. It breaks.

Q3 earnings are due November 12. The company guided for $860 to $940 million in interest expense against $200 to $260 million in adjusted operating income. The gap is widening, not closing.

Nvidia owns 5.96% of CoreWeave while also supplying its GPUs, partnering with its lenders, and setting the hardware standard the collateral is priced against. The circular financing makes the mortgage market's conflicts of interest look quaint.

Your retirement account might already own this

The risk has already left Wall Street's balance sheet.

Valor Equity Partners structured a $5.4 billion GPU deal for xAI. Apollo funded $3.5 billion of that debt, then placed it with Athene, its insurance subsidiary that manages retirement annuities. A single customer's compute contract, backed by hardware that depreciates annually, now sits in a vehicle that pays retirees.

Michael Burry called it out publicly. He didn't argue the deal was illegal. He argued it moves credit risk from sophisticated investors to people who don't know they're holding it.

About $80 billion in data center bonds have been issued in 2026. Roughly $37 billion of that is rated below BBB.

Another $120 billion in AI infrastructure debt has been moved off-balance-sheet through special purpose vehicles by Oracle, Meta, xAI, and CoreWeave. SPVs exist to make debt disappear from the sponsor's balance sheet. The debt doesn't disappear. It finds a new home in the portfolios of insurance companies and pension funds that think they're buying infrastructure.

The depreciation fiction

GPU-backed debt has a problem that mortgage-backed securities didn't. A house might lose value, but a GPU is guaranteed to.

Microsoft depreciates its GPUs over six years. CoreWeave uses six years. Meta uses five and a half. The industry consensus is five to six years of useful life.

Burry argues the economic useful life is two to three years. H100 rental rates fell roughly 70% between mid-2025 and the end of that year, to about $2.50 per hour. The chips didn't break. Nvidia shipped the B200, and the economics of the older generation shifted overnight.

If the real useful life is three years instead of six, the industry is understating depreciation by $50 to $60 billion per year. That gap sits on balance sheets as book value that doesn't exist in resale markets.

Ares Management CEO Michael Arougheti, whose firm manages $500 billion in assets, said, "I don't know how to underwrite the depreciation curve."

When the people writing the checks say they can't model the collateral, the people buying the bonds should listen.

Aircraft are the closest precedent for securitizing a depreciating physical asset. Airlines have issued ABS backed by planes for decades, with 25-year depreciation schedules, about 15% residual value, and a deep secondary market for used planes.

During COVID, when demand fell suddenly, KBRA downgraded 81 aviation ABS bonds. Only 5 of 48 outstanding rated deals avoided breaching cash-trap triggers.

That was with 25-year assets and a secondary market. GPUs depreciate in three to five years and have no comparable resale infrastructure. A three-year-old GPU is often worthless for frontier training.

We built this before

Between 1997 and 2001, investors poured over $750 billion into telecom infrastructure, most of it fiber optic cable. The thesis was that internet bandwidth demand would grow exponentially. The thesis was correct.

By 2001, Merrill Lynch estimated that about 3% of U.S. fiber capacity was carrying traffic. More than 60 telecom companies went bankrupt between 2001 and 2003, including Global Crossing ($12.4 billion in debt) and WorldCom ($41 billion). The fiber still worked, but the demand didn't arrive on the timeline the debt was underwritten against.

GPU compute has higher utilization rates than fiber did at its peak, and the customers are real companies paying real invoices. But the financing structure has the same vulnerability. It needs demand growth to continue at the rate the debt was priced against. Any slowdown in the growth rate, even while demand keeps increasing, changes the math.

Banks that raced to fund neocloud companies are already hesitating. The early enthusiasm is cooling into the question that killed fiber. What happens at 30% growth instead of 300%?

Who holds the bag

The mortgage market had a name for what happens when risk leaves the originator's hands and lands in someone else's portfolio. They called it originate to distribute. The bank that made the loan didn't keep it. When the loans went bad, the banks were fine. The pension funds weren't.

GPU-backed debt is following the same path. Nvidia doesn't hold the debt. CoreWeave's lenders don't hold it for long. It moves through SPVs and insurance subsidiaries until it lands in the portfolios of people who think they own infrastructure bonds.

The SEC cleared this path by exempting data center bonds from post-2008 securitization rules. The risk retention and enhanced disclosure requirements that were installed after the last time Wall Street packaged depreciating assets into yield products and sold them to retirement accounts were removed for this asset class before the first default.

I don't know when GPU utilization rates will disappoint. I don't know which neocloud will be the first to miss a payment on a GPU-backed facility. But CoreWeave's interest expense is five times its operating income. Its bonds trade at distressed levels. Banks are pulling back.

The money has already moved far enough from the people who understood the risk to the people who don't.