The customer sets the price: why GPU loan margins fell from 8.75% to 2.25%

When CoreWeave signed its first delayed-draw term loan with Magnetar and Blackstone in August 2023, the US$2.3bn facility priced at SOFR + 8.75%. GPU-backed lending was new, the borrower was a fast-growing private company and the lenders were private credit funds pricing an untested asset.

By March 2026, CoreWeave's DDTL 4.0, an US$8.5bn facility arranged by MUFG and Morgan Stanley with Blackstone as an anchor, priced at SOFR + 2.25%, with a fixed-rate option at Treasuries + 2.00%. Over the same period disclosed GPU-backed lending grew from US$2.5bn in 2023 to US$42.2bn in 2026 to 22 October, across 31 public loans.

It is tempting to read this as the market simply getting comfortable with GPUs. The documents point to something more specific.

The pricing grid that gave it away

CoreWeave's DDTL 2.0 in May 2024 was one US$7.5bn facility with three prices. Draws funding contracts with specified investment-grade customers paid SOFR + 6.00%. Other investment-grade contracts paid 6.50%. Non-investment-grade contracts paid 13.00%. Same borrower, same GPUs, same lenders, same documents: a seven-point spread set entirely by who was paying for the compute.

That grid is the clearest public evidence of how lenders think about GPU credit. The GPUs are the collateral, but the customer contract is the cash flow that repays the loan. Lenders price the contract first.

Three prices from one sponsor in 2026

The 2026 deals make the same point. CoreWeave borrowed three times in five months:

• DDTL 4.0 (March): one investment-grade customer, a limited parent guarantee, SOFR + 2.25%.
• DDTL 5.0 (May): two non-investment-grade customers, a full parent guarantee, SOFR + 4.50%.
• DDTL 5.5 (August): a diverse book of contracts averaging about three years, a full parent guarantee, SOFR + 5.50%.

The facility with the least parent support priced the tightest. A full guarantee from the sponsor did not close the gap to a strong customer. For borrowers that may be the most useful lesson in the market: the quickest way to a lower margin is usually to change whose contract is financed, not to add more of the sponsor's own balance sheet.

Other sponsors landed in the same place. IREN's Hardware 3 facility, backed by a Microsoft contract, priced at SOFR + 2.50% in May 2026. Nebius's first secured GPU loan, backed by an investment-grade customer contract and without a full parent guarantee, priced at 2.50% in July. Nscale's Ward County and North Carolina loans, signed in August, priced at 2.375% over SOFR or the swap rate.

What else moves the price

Customer credit is the biggest driver, but not the only one. Three other features show up in the tighter deals:

• Project finance structure. The 2.25% to 2.50% loans sit in ring-fenced borrowers, with draws tied to delivery and acceptance, controlled accounts, reserves and cash traps.
• Full amortisation within the contract. The loan is repaid from contracted revenue before the contract ends, so lenders take little residual value risk.
• Ratings. IREN's Hardware 3 and Lambda's Compute I facility carry investment-grade ratings, which opens the deal to a wider pool of banks and insurers.

Where those features are absent, margins stay higher. Nscale's European facility in February 2026 priced at SOFR + 5.00%, stepping to 5.25%, and its Iceland loan with Macquarie in July carried SOFR + 5.50% plus return-based payments. Crusoe's Iceland facility launched to lenders in September at talk of 4.75% to 5.00%.

Questions for the next term sheet

• Whose credit really stands behind the cash flow, and can the contract be assigned to the lenders?
• Would a pooled facility with tiered pricing cost less than one blended margin?
• What is the parent guarantee buying, if a stronger customer would cut the margin by more?
• Does the loan fully amortise before the contract ends, and if not, what covers the tail?