Borrowing Against Chips

Every neocloud carries a funding gap. A neocloud buys chips before its customer pays for them. The customer pays monthly over years. Something has to fund the difference, and at the scale of these fleets, equity alone is too expensive. The answer the industry developed is debt secured on the chips themselves and on the contracts they serve.

CoreWeave built the template. In August 2023 it raised a 2.3 billion dollar debt facility led by Magnetar and Blackstone, secured on its Nvidia H100 hardware. In May 2024 it closed a far larger facility of 7.5 billion dollars, again led by Blackstone. Both were structured as delayed-draw term loans, or DDTLs. A delayed-draw facility is a commitment to lend up to a ceiling, drawn in stages as the borrower actually buys and installs equipment. The borrower pays for money only when it needs it; the lender releases money only when there is collateral to show for it.

The legal architecture matters as much as the size. These loans typically sit in special-purpose vehicles, or SPVs: separate subsidiaries that own a defined pool of servers and hold the customer contracts those servers serve. Lenders take a first claim on the equipment and on the vehicle's equity. Cash from the customer flows into the vehicle and pays the lender before it reaches the parent company. The design isolates each fleet, so a problem in one pool does not automatically infect another, and it ties each loan to a specific revenue stream.

Look at what the lender is really underwriting. The chips are collateral, but they lose value quickly and their resale market is thin. The stronger security is the contract. A take-or-pay commitment from a highly rated buyer is a stream of cash a lender can model. That is why pricing followed the customers. By March 2026, CoreWeave had closed a facility of up to 8.5 billion dollars, its fourth delayed-draw loan, secured on graphics processing units, or GPUs, and a contract with Meta, which Moody's rated A3 and Morningstar DBRS rated A low. The company described it as the first investment-grade financing backed by this kind of hardware. The class of collateral had not changed. The counterparty had.

Delayed-draw loans are one tool among several. The industry uses at least three other channels, and a reader should look for each in a filing.

The first is customer financing. The Nebius agreement with Microsoft, disclosed in September 2025, included large upfront payments from Microsoft. Money paid in advance funds construction without adding debt. It also tells you who holds the bargaining power: a customer that prepays is effectively lending to its supplier, and it will want terms in return.

The second is vendor and supplier support. When Nvidia buys shares in an operator, anchors an offering, or commits to buy unused capacity, it lowers the cost of capital for the operator by reducing the risk lenders see. It is financing by another name.

The third is conventional corporate finance: unsecured bonds, convertible notes and new equity. CoreWeave issued high-yield bonds in 2025 after its listing. Nebius and others have raised money through share and convertible offerings. These sit above the equity and below the secured lenders, and their holders are the first to feel it when a secured structure absorbs the best assets.

Put the channels together and a pattern appears. The best collateral, contracted revenue from strong buyers, goes to secured lenders at the lowest cost. What remains, including uncontracted capacity and the parent company's own obligations, is funded more expensively. That layering works while the contracts perform and the operator keeps winning new ones. It becomes strained if contracts end and must be replaced at lower prices, because the loans were sized to the old ones.

Faraday's read is that GPU-backed debt is really contract-backed debt with a hardware kicker. Ask who the customer is before you ask what the chip is worth.