Can GPU-backed debt scale from a niche private-credit product into a real institutional asset class? Or is it a collateral time bomb with a ticking depreciation clock?
CoreWeave closed an $8.5 billion delayed-draw term loan on March 31, 2026. Moody's rated it A3. DBRS rated it A(low). Both are investment grade. That matters not because of the dollar amount, but because this was the first time a loan secured by graphics processing units received an investment-grade rating from two major agencies. The collateral was silicon, not a power plant or an aircraft fleet.
The same quarter, Moody's awarded its first Aaa rating to a data center securitization, an $830 million Compass deal in February. S&P published formal criteria for rating data center ABS in the same period, a prerequisite for the asset class to scale beyond bespoke club deals.
Neither event is isolated. GPU-backed and data-center-backed debt is shifting from the experimental edge of private credit into the regulated centre of institutional fixed-income markets.
As we wrote last week, compute derivatives crossed the same threshold. Institutional asset class status for AI infrastructure financing is becoming a pattern, not an anomaly.
The numbers so far
Data center asset-backed securities (ABS) issuance grew from $2.6 billion in 2020 to more than $14 billion annually by early 2026, according to Charles River analysts. JPMorgan projects the market will reach $30–40 billion per year by 2027. Over $20 billion in GPU-backed debt exists on neocloud balance sheets as of early 2026. That figure excludes the $27 billion single-datacenter private-credit deal Meta structured with Blue Owl in October 2025.
Law firms are now publishing formal guidance on GPU infrastructure financing. Clifford Chance issued a briefing in July 2026 covering special purpose vehicle (SPV) structures, revenue-based collateral valuation, and cross-border considerations. The Bank for International Settlements devoted a box in its March 2026 Quarterly Review to the "shadow borrowing" embedded in hyperscaler off-balance-sheet structures. Central banks are watching this market for systemic risk.
How GPU ABS works
The structure follows equipment ABS conventions with one critical difference: the collateral depreciates faster than the debt amortises. A GPU loses 30–40% of its value in year one. An H100 that cost $30,000 in 2023 was worth roughly $8,000 by early 2026. NVIDIA releases a new architecture every 12 months, meaning every chip used as collateral today sits on a depreciation treadmill that resets annually.
Lenders compensate with structural protections. Advance rates are 50–70% of hardware value. Principal is front-loaded to retire most of the debt before the collateral value collapses. Loan covenants require real-time telemetry on GPU utilisation. If a cluster goes dark, the lender knows within hours. Residual value swaps, offered by firms like Ornn AI, let borrowers lock a floor price on used hardware at origination.
Investment-grade ratings came from contract quality, not better collateral. CoreWeave's A3-rated $8.5 billion facility is secured not only by GPUs but by a six-year Master Services Agreement with Meta. The rating agencies priced recurring hyperscaler revenue, not secondary chip-market liquidation value, as the primary credit support. That shift from asset-based to cash-flow-based underwriting opens the door to a scalable asset class.
What changes when GPU debt goes investment grade
Before the A3/A(low) rating, GPU-backed lending was a private-credit product. The buyers were Blackstone, Magnetar, and Apollo — firms that underwrite illiquidity premia on bespoke structures. Investment-grade ratings change the investor base. Insurance companies, pension funds, and sovereign wealth funds have mandates that cap their exposure at specific rating thresholds. An A3-rated tranche fits inside those mandates. Legal & General's asset management division reports that insurers are among the most active buyers of data center ABS, drawn by the spread over equivalently rated corporate bonds.
The spread compression tells the story. GPU ABS coupons tightened from roughly 1,300 basis points over Treasuries in the first 2023 CoreWeave deal to around 110 basis points on the February 2026 Compass Aaa-rated tranche. That is a tenfold compression in under three years. The speed of the compression reflects both market familiarity and the sheer volume of institutional capital looking for AI-infrastructure exposure.
The open question
The asset class has never been tested through a downturn. The $20 billion in GPU-backed debt has been originated in a rising market where NVIDIA's annual revenue grew from $27 billion (FY2023) to $130 billion (FY2026). Every assumption about collateral value, contract renewal, and secondary-market liquidity is untested against a correction in AI chip demand.
The single most important variable is the depreciation cycle and its correlation. If multiple neoclouds default simultaneously, triggered by a downturn, a chip-generation shift, or a hyperscaler contract renegotiation, the secondary market for used GPUs would face concentrated supply auction. The advance-rate buffers would narrow fast. Fitch has already published a public consultation on whether GPU depreciation should be formally incorporated into its ratings of securitizations backed by AI infrastructure. That consultation signals that the rating agencies see this as an unresolved structural question.
The next 12 months will be instructive. CoreWeave's Q1 2026 revenue of $2.08 billion (more than double year-over-year) and a near-$100 billion revenue backlog demonstrate that the underlying demand is real. But capital structure resilience is a different question from end-market demand. The bond market has answered "yes" to GPU ABS as a product structure. It has not yet answered what happens when the treadmill speeds up.
