Two major AI customers can exit on 90 days’ notice before new equipment arrives, although their $26 billion in annualized payments dwarf the financing costs.

SpaceX’s $40 billion agreement to finance Nvidia’s next-generation AI chips through Apollo Global Management exposes a mismatch in artificial intelligence infrastructure: The equipment carries five-year repayment obligations, but major customers can terminate their computing contracts on 90 days’ notice before the new processors even arrive.

The company has substantial demand. Two disclosed agreements with Google and Anthropic provide for combined monthly payments of $2.17 billion, or $26 billion annually at full contracted rates. That is more than twice the approximately $11 billion in illustrative first-year principal and interest on the new financing.

Yet Anthropic’s termination right became available after an initial three-month period beginning in May, while Google’s becomes available after December 31. Both precede the first deliveries of Nvidia Vera Rubin equipment, expected in the first half of 2027.

Neither customer agreement has been identified as supporting the Apollo-financed equipment. Their terms nevertheless reveal a central challenge for AI infrastructure operators: Financing commitments remain fixed even when customer revenue, computing prices and hardware competitiveness change.

SpaceX disclosed October 26 that it had secured a binding commitment for approximately $40 billion in finance leases with an Apollo-managed entity. The financing covers Vera Rubin GPU systems and related equipment, delivered in installments beginning next year.

Each tranche will carry a lease of approximately five to 5.25 years. Quarterly payments will generally include 5% of the financed equipment cost, interest at a fixed rate yet to be determined, and additional expenses. The transaction remains subject to definitive documentation and closing conditions.

Substantial revenue, shorter commitments

SpaceX’s existing contracts demonstrate its ability to attract large customers.

Google agreed in June to pay $920 million monthly for access to computing capacity including approximately 110,000 Nvidia GPUs and associated infrastructure through June 2029. Either party can terminate on 90 days’ notice after December 31, 2026.

Anthropic agreed in May to pay $1.25 billion monthly for capacity including approximately 325,000 GPUs, with payments scheduled through May 2029. Its agreements allow either party to terminate on 90 days’ notice after an initial three-month period.

These rights do not imply that either customer intends to leave. They do mean the stated contract endpoints are not guaranteed revenue streams.

Nor does $26 billion in annualized revenue translate into $26 billion available for Apollo payments. The agreements cover processors, networking, storage and other services, each carrying costs. They also concern existing infrastructure rather than the incoming Rubin equipment.

SpaceX’s latest quarterly filing reported $7.56 billion in third-quarter superintelligence revenue, demonstrating substantial commercial scale. The segment generated $4.95 billion in adjusted EBITDA, but $431 million in operating income after depreciation and other expenses.

The contrast underscores why gross customer payments alone cannot establish investment returns.

Apollo’s exposure shrinks every quarter

The Apollo arrangement amortizes principal through 20 quarterly installments, generally recovering 20% of each tranche’s financed cost annually.

Most Apollo principal is scheduled for recovery within three years

Illustrative repayment schedule, assuming $40 billion financed simultaneously.

Source: SpaceX October 26 disclosure; Fastinform calculations. Interest and fees excluded. Actual payment dates depend on equipment deliveries.

At an illustrative 8% annual interest rate, assuming each $2 billion principal installment is paid at the start of the quarter and interest accrues on the reduced balance, first-year interest totals $2.8 billion. Combined principal and interest would reach $10.8 billion, excluding fees.

That schedule reduces Apollo’s dependence on the residual value of aging GPUs. By the end of year three, 60% of the original principal would have been repaid.

But repayment obligations remain regardless of how quickly the equipment loses commercial competitiveness, unless contractual provisions establish otherwise.

The preliminary disclosure leaves important protections unresolved, including corporate recourse, collateral guarantees and whether returning equipment could discharge outstanding obligations. It also does not explain the potential additional quarter beyond the 20-installment schedule.

Depreciation presents a measurable risk

Newer AI processors can produce more computational work for a given investment in electricity and infrastructure, threatening the pricing power of older equipment.

Yet technological obsolescence and economic depreciation are different. Older GPUs can remain commercially useful for inference, fine-tuning and workloads that do not require frontier performance.

Accounting and credit markets take differing views. CoreWeave depreciates technology equipment over six years, while Nebius adopted five years for servers and networking equipment in 2026. Reuters reported October 1 that banks commonly underwrite GPUs using three- to four-year depreciation assumptions.

SpaceX’s own financial statements show the scale of this expense. Its superintelligence segment recorded $2.473 billion in third-quarter depreciation and amortization, approximately 33% of segment revenue. The figure covers multiple assets, not just GPUs, and does not reveal the useful life assigned to individual processors.

The financial exposure therefore depends less on whether Rubin chips can operate for five years than on whether their revenue and operating contribution remain sufficient throughout the repayment schedule.

Rapid payback could neutralize the risk

SpaceX Chief Financial Officer Bret Johnsen offered the strongest counterargument in the company’s October 26 financing announcement, saying its superintelligence infrastructure investments were achieving payback periods of less than one year.

If that means recovering invested capital through net operating cash generation, the risk of subsequent hardware depreciation would be considerably reduced. But SpaceX has not defined the calculation sufficiently to establish whether it measures revenue, operating contribution or cash investment recovery after expenses.

Nor has the company demonstrated that its historical payback experience applies to the new Rubin systems.

SpaceX can also allocate computing capacity between its own Grok models and external cloud customers, providing flexibility that independent GPU rental providers may lack. Internal deployment, however, does not necessarily generate the same cash returns as commercial contracts.

The company consequently enters the Apollo arrangement with two substantial advantages: an established multibillion-dollar computing business and a financing schedule that steadily reduces outstanding principal.

Its remaining exposure is whether the new equipment can generate sufficient cash as competing chips improve and customer commitments expire.

The decisive test will be whether SpaceX can substantiate its claimed sub-one-year cash payback—and whether the final Apollo agreement protects lenders if that recovery proves slower than expected.