SpaceX's $40 Billion GPU Plan Puts Cash Flow to the Test

SpaceX is in preliminary talks with Apollo and banks over financing for a proposed $40 billion purchase of Nvidia GPUs. The prospective financing would rely primarily on investment-grade debt, but key lenders and other terms remain unsettled. Apollo declined to comment. For a company buying chips to rent out computing power, the number to examine alongside $40 billion is the amount customers will pay, and for how long.

The proposed package could comprise about $10 billion in bank loans and $30 billion in investment-grade debt. Completion was expected in 2027, while earlier-stage talks could end without a deal, the Financial Times and Bloomberg said. Neither SpaceX nor the other companies involved had publicly confirmed the plan as of October 7. What backs each portion? When does repayment begin? Do customer contracts run long enough to support it?

How Chips Became a Credit Question

SpaceX has a recent reference point for bond investors. It raised $25 billion in bonds shortly after its mid-June IPO. Demand for that sale was strong, but AI-related bonds subsequently sold off and credit spreads widened. A borrower returning to the market months later may encounter a different price for money even if its plan to deploy the money has improved. The proposed GPU borrowing would test investor appetite for another large obligation whose payoff depends partly on a newer computing business.

Data-center capacity constrained growth in 2025, chip production became a constraint by early 2026, and financing emerged as an obstacle by midyear. The sequence helps explain why banks and bond buyers now sit so close to the question of acquiring GPUs: securing hardware and finding customers do not, by themselves, fund the purchase.

Long-term customer commitments or a backstop guaranteeing a minimum payment for capacity can make some GPU projects financeable, SemiAnalysis said. It said Nvidia has offered take-or-pay backstops to some neoclouds, receiving a share of revenue above the guaranteed level in return. That arrangement gives a lender another party's promise to examine alongside the GPUs. A lender assessing SpaceX could instead look at its customers, the company's wider cash generation, collateral, or some combination. Those routes put different parties on the hook if a GPU cluster earns less than expected.

Revenue Has a Calendar

SpaceX has a case to put before those lenders. Its Memphis computing sites, Colossus 1 and Colossus 2, are generating billions of dollars in monthly recurring revenue, according to company disclosures cited by International Business Times. Chief Financial Officer Bret Johnsen told investors last month that the company believed it was on track to reach $100 billion in annual recurring revenue; he also disclosed a hosting agreement expected to produce approximately $1.11 billion a month beginning December 1.

That prospective payment is a plausible reason to seek more chips. It also illustrates the lender's problem. The December start is an expectation, while the $100 billion figure is management's outlook for an annualized revenue measure, rather than cash already collected to repay a loan. A credit decision would depend on contract length, payment obligations, operating costs and the timing of chip deployment. Those details determine whether customer money arrives before debt payments come due, and how much is left after the equipment has been powered and operated. SpaceX's reported revenue ambitions make the borrowing rationale understandable; they do not set a borrowing rate.

Yorkville analyst Dan Ives argued that SpaceX's AI backlog and Starlink cash generation could support additional debt, International Business Times said. That expands the possible repayment case beyond a single GPU customer. It also raises a structural question for shareholders: would lenders have claims only against identified computing assets and their cash flows, or against more of the company? A lender comfortable with SpaceX's wider business might accept a different package of protections from one relying mainly on chip rentals.

Other GPU Borrowers Show What Terms Can Do

CoreWeave offers a closer financing comparison than a generic corporate bond. It has used delayed-draw term loans to pay for GPUs, with borrowing generally drawn as equipment is purchased and deployed. Interest usually begins accruing on the amount drawn, although an undrawn commitment can carry a fee. An $8.5 billion CoreWeave facility is split roughly between fixed- and floating-rate portions; the floating portion will require hedges on at least 95% of anticipated borrowing. CoreWeave has told investors its loan payments are structured to align with contractual payments from customers.

That structure addresses two different clocks. Delayed draws can keep a borrower from paying full interest before it needs all the chips. Customer payment schedules can help it plan debt service after deployment. Hedges can constrain exposure to rising benchmark rates, though they do not ensure customers keep buying compute or that GPUs retain their value. A bond-heavy SpaceX package could allocate those risks differently from CoreWeave's loans; its eventual draw schedule, rates and hedges would show how.

Two other facilities illustrate why a quoted spread alone cannot price SpaceX's proposal. CoreWeave's $2.6 billion delayed-draw loan carried Term SOFR plus 5.50%, while a roughly $775 million secured Nebius facility carried SOFR plus 2.50%. Nebius said its facility was backed by deployed GPU infrastructure and cash flows from an investment-grade customer. CoreWeave's facility financed committed deployments across a broader customer pool. The borrowers, collateral and customer support differ; neither rate predicts what SpaceX would pay. They show the questions a rate answers once a deal is signed: whose promises support repayment, and what assets can lenders claim if those promises fall short?

GPUs would likely serve as collateral for SpaceX's proposed borrowing, a person familiar with the talks told CNBC. If so, lenders would have an asset to claim, but its sale value could be weaker when demand for the same computing capacity is weak. The eventual collateral terms and maturity would show how much of that risk lenders accept and how much remains with SpaceX. Calling the prospective bonds investment-grade describes the market the parties hope to tap; the final protections bondholders receive will depend on the terms.

For employees with equity, debt sits ahead of common shares when a company's assets are distributed. Debt service draws on cash before any residual value can accrue to shareholders, while borrowing instead of issuing new shares avoids immediate dilution.

Customers have a stake in the financing design, too. A provider that matches equipment purchases, repayments and contracted demand has more room to deliver the capacity it sold. If those schedules diverge, the pressure does not stop at a bondholder's spreadsheet. SpaceX's next consequential numbers may be the contract duration, collateral boundaries and repayment dates attached to the proposed $40 billion, rather than a larger chip order.