Your GPU Contract Was Written by Someone Else's Lender

What Happened

The banks that wrote the loan would rather not hold it. The Financial Times reported on August 5 that a syndicate led by Morgan Stanley plans to break $15 billion of debt on a 2,000-acre data center campus in Hubbard, Texas into a series of bond sales. Nexus Data Centers is developing the campus, Anthropic has leased it, and Google is behind the project. The bonds are still expected to price at speculative grade, because Google’s support only takes effect once the building is finished. The loan also carries a delayed-draw feature tied to construction milestones (Investing.com).

One deal proves nothing. Several signals pointing the same way do. Morgan Stanley models $570 billion of AI-related debt issuance in 2026, and $236 billion had priced by May 31, four times the year-earlier pace. Demand moved the other way. Order books on hyperscaler bonds ran at nearly five times the amount offered in February and slipped below two times by July (Forbes). Whatever the public market will not absorb goes to private credit, an estimated $800 billion opportunity in data center financing through 2028.

Where that debt sits is documented. The BIS laid the structure out in its March quarterly review: a special purpose vehicle buys the data center assets, the hyperscaler keeps a minority equity stake and takes the capacity back through long-term operating leases and offtake agreements. Upfront capex becomes multi-year operating expense, and most of the borrowing stays off the hyperscaler’s balance sheet. Private credit funds and institutional investors are the ones actually holding it (BIS). Moody’s put unrecognized lease commitments across Amazon, Microsoft, Google, Meta and Oracle at $662 billion, or 113% of their adjusted debt. Under current accounting a lease lands on the books when the tenant takes control of the asset, so a build-to-order data center that takes years to finish stays in the footnotes until then (Bisnow).

The scale of it was covered in July. What changed is not the number. It is who holds the paper and what they now ask for.

What This Means for Founders

Private credit does not lend against a building. It lends against contracted cash flow. So the credit agreement carries a condition: line up long-dated contracts with large, creditworthy customers. Data center operators need to satisfy it, and so do the neoclouds renting GPUs on top of them.

There is essentially one way to satisfy it. Put a minimum commitment in front of the customer and turn it into paper. That is why a GPU quote arrives with a two-to-five-year floor attached and a clause saying you pay for the window whether you burn it or not. It reads like a sales tactic. It is a lending condition. The BIS description of these vehicles lists capacity offtake agreements right next to the long-term leases for the same reason.

When lenders tighten, the pressure travels downhill. In recent weeks at least four borrowers, CoreWeave among them, had to sweeten terms to get deals done. What a provider under that pressure adjusts is the customer contract: longer floors, prepayment and deposits, steeper early-termination fees. A seed-stage company with no credit rating of its own meets that change first.

Look at the collateral too. These facilities pledge more than the GPUs. The customer contracts themselves are collateral. If your agreement is part of your vendor’s security package, an early exit or a term change is not something an account executive can hand you. It moves into territory where the lender has to sign off. Read the contract, not the price list, or that part stays invisible.

What You Can Do Now

Three lines in the agreement carry most of the risk. The minimum commitment term and its total value. Whether unused capacity rolls forward or evaporates each month. What early termination costs, and whether a lender has to approve it.

Ask the vendor one question as well: how is this capacity financed? If it sits on their own balance sheet, there is room to negotiate. If it is a private credit facility, the person across the table may not have the authority to shorten your term. Most will tell you if you ask.

When the offer is framed as price versus flexibility, take flexibility in this stretch. Trading 10% off the hourly rate for a three-year floor is hard math in a market where the model you depend on changes several times a year.

Order-book coverage gets published with every issue. How many times over the next hyperscaler bond sells is a few months of advance notice on the commitment term you will be offered next year.