The Real Risk in GPU Financing: Why Private Credit Is Pricing the Waterfall, Not the Silicon

IREN closed $2.8 billion of new GPU-backed financing, including a $2.4 billion facility led by Blue Owl and PIMCO at 9% fixed, funding roughly 90% of associated GPU capital expenditures for non-investment-grade customer deployments. On August 27, Lambda closed a separate $926 million asset-backed SPV term loan at SOFR+300 basis points—oversubscribed, tightened 75bp during syndication, rated Baa2 by Moody's. Institutional lenders are now financing individual GPU fleets against contracted compute cash flows, and the pricing tells a precise story about where risk actually sits.

The spread ladder

GPU-backed debt already displays wide differentiation by customer credit quality. IREN's Microsoft-backed $3.65 billion financing clears at roughly 6% blended. CoreWeave's $8.5 billion Meta-backed facility priced at SOFR+225bp. Lambda's NVIDIA take-or-pay deal, Baa2. These behave like project finance.

Move down the credit stack and the numbers shift fast. CoreWeave's DDTLs backing weaker or unrated customers price at SOFR+450bp and SOFR+550bp. IREN's Mackenzie tranche, serving non-investment-grade deployments, costs 9% fixed with a roughly 2.5-year post-funding tenor. A 225-to-325bp spread gap now separates investment-grade offtake from everything else. That gap is the price the market charges for the specific risk that a customer disappears before principal amortizes—leaving the lender holding depreciating silicon.

Secondary prices confirm the depreciation curve

CCIR's trailing executed-sale database, as of August 25, quantifies the haircuts. An H200 after about 2.3 years retains roughly 85% of its original acquisition value. An H100 NVL after 2.9 years: 77%. An H100 SXM after 3.8 years: 42%. An A100 80GB SXM at 5.8 years: 22%. A V100 at 8.4 years: 3.6%.

The decay is non-linear, concentrated around architecture transitions and power-efficiency shifts. Rental economics deliver an even earlier signal: H100 price per effective FP8 PFLOP fell 35.3% quarter-over-quarter in Q2 2026. A GPU can run for years while its revenue-generating capacity collapses.

Vera Rubin changes the collateral math for Blackwell

IREN is borrowing $2.4 billion to deploy Blackwell Ultra / B300-class infrastructure. NVIDIA's Vera Rubin is already entering production, with NVIDIA claiming roughly 10× agent throughput at scale versus Grace Blackwell. AWS announced plans to deploy another 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs during 2027–28. A B300 financed in late 2026 therefore lives through B300 → Rubin → Rubin Ultra within its debt term. Lenders cannot treat these assets the way they treat turbines or fiber conduit—equipment whose utility holds steady across a financing cycle.

IREN's own balance sheet tells the parable

IREN reported $638.8 million of asset impairments at June 30, largely from decommissioned Bitcoin mining hardware—specialized compute that became uneconomic when the company pivoted its sites to AI workloads. The company is now borrowing billions against the next generation of specialized compute. ASIC miners are less fungible than NVIDIA GPUs, so the comparison has limits. But IREN's own books contain a recent, large-dollar record of how rapidly purpose-built computing hardware reprices.

Friday's equity tape reinforced the tension: IREN fell roughly 13% to about $35.10 after earnings, despite AI cloud revenue jumping 110% sequentially and management calling 2026 capacity largely sold out. Blue Owl itself traded slightly higher. Equity markets are punishing the leveraged borrower, not the lender—for now.

The prepayment cushion that reframes the entire thesis

A bearish reading of GPU debt stops too soon if it ignores IREN's disclosed deal mechanics. Recent customer contracts contain prepayments equal to roughly 45–55% of GPU capital expenditures. Combined with debt financing at approximately 90% of GPU capex, upfront funding exceeds the equipment purchase cost.

This is a different animal from a lender advancing 90 cents against a GPU and hoping it stays at 90. IREN's structure starts with material customer cash inside the capital stack. Lambda's Baa2 facility, meanwhile, fully amortizes through December 2030 against an NVIDIA take-or-pay contract—if the offtaker performs, lenders recover principal through operating cash flows, and terminal GPU resale value is largely beside the point.

The right question for every investment committee is therefore how that prepayment cash behaves under stress. Is it trapped inside the SPV? How fast does amortization burn down principal? What triggers release? What happens on customer termination? Those waterfall mechanics—not GPU auction prices—determine whether a lender takes a loss. The strongest GPU financings are contracted-infrastructure instruments where hardware residual value is nearly irrelevant. The weakest are equipment bets dressed in project-finance language. The credit curve already knows which is which. Portfolio construction should follow.

not investment advice