Apollo's $40B SpaceX Lease Tests GPU Collateral

SpaceX agreed to a reported roughly $40 billion lease with Apollo Global Management to buy Nvidia GPUs. Nvidia has also been trying to find data-center space for companies that already own GPUs. The chips can change hands more easily than the power connection can.

Bloomberg's October 26 report said the lease would finance GPUs for SpaceX's AI initiatives and was set to close before the end of 2026, not described as funded. On the coverage record, its reported size is about 12 times the roughly $3.4 billion loan Apollo was reportedly arranging in February for a Valor-led vehicle to buy chips and lease them to xAI. Neither figure establishes the final collateral or risk-allocation terms of the SpaceX deal.

AI infrastructure is acquiring a second balance sheet: lenders can finance GPU fleets separately from the companies and data centers that use them. But equipment finance stands or falls on a harder question than whether today's operator can pay. If that operator cannot, can the lender place the fleet in another powered facility, with another paying workload, before the hardware loses value?

CoreWeave made the chip a borrowing asset

In 2023, CoreWeave raised $2.3 billion in debt collateralized by Nvidia chips. The cloud provider still had to operate the machines and sell access to them, but lenders could identify equipment rather than rely solely on the value of CoreWeave's business. That distinction gave a fast-growing operator a way to fund hardware purchases without first accumulating the balance sheet of an established cloud company.

Ownership and operation were already coming apart elsewhere. Tether spent roughly $420 million on 10,000 H100 GPUs and a stake in Northern Data in 2023; Northern Data planned to rent the chips to AI startups. By February 2026, the Financial Times reported that tech companies were increasingly using GPU-backed debt and special-purpose vehicles to place hardware obligations outside their own balance sheets. A vehicle can buy the chips, borrow against them and lease them to an operator. The operator supplies the workload; the vehicle owes the lender.

That legal division matters when payments stop. A lender with a claim on identifiable GPUs has a different starting point from a lender with only a claim against a cloud provider. It does not, however, receive a working data center along with the chips.

Apollo moved from financing factories to financing fleet use

Apollo's earlier semiconductor deals attached capital to production capacity. In 2023, Wolfspeed said an Apollo-led group would make a $1.25 billion debt investment, with room for another $750 million, to support its US expansion. In 2024, Apollo was reported to be in exclusive talks to supply more than $11 billion for Intel's Irish plant. Those investments depended on chipmakers building and operating facilities; Apollo was financing the enterprise or the factory that made the chips.

The proposed Valor-led vehicle for xAI presented a different set of parties. Apollo was reportedly nearing a loan to a vehicle that would acquire Nvidia chips, while xAI would lease and use them. That arrangement was reported, not confirmed. xAI had been burning more than $1 billion a month through fall 2025, according to the same coverage. A claim on the equipment would not make xAI's payments immaterial: a borrower could still face a costly search for another tenant if the intended user could not keep paying.

SpaceX takes the equipment-finance question beyond the specialist cloud provider. Bloomberg described a lease for SpaceX's AI initiatives, not a completed transaction or a confirmed copy of the Valor structure. Apollo cannot infer from the proposed amount alone how much risk rests with SpaceX, the equipment, a guarantor or a future customer. Those are the terms that decide what the lender owns when a lease goes wrong.

Nvidia is helping lenders bridge the gaps around the chip

In February, Nvidia disclosed $3.5 billion in guarantees to companies leasing land, power and data-center facilities, four times its third-quarter level. Those guarantees concern the premises needed to run chips; they are not evidence of a guaranteed resale price for GPUs. Nvidia has also offered to rent back unused GPUs from young cloud providers in exchange for a share of revenue. That offer addresses idle capacity, but a revenue-sharing rental is not proof that every financed fleet could be sold at its loan value.

Other parties have begun to put prices and promises around the uncertainty. Nvidia and Broadcom have supported customer equipment purchases through residual-value guarantees, according to September coverage. CME Group and Silicon Data announced futures contracts based on daily benchmarks for on-demand GPU rental rates. A rental benchmark could help lenders compare a fleet's contracted income with the market price of capacity. An announced futures market cannot yet tell them what a repossessed fleet would fetch, how long a move would take, or who would bear the cost.

Lenders have not treated every borrower alike. AI infrastructure companies borrowed more than $100 billion in 2025, while smaller companies paid higher interest rates because investors remained wary of unproven businesses. The chip provides collateral; a lender still has to price the operator.

The lender's exit needs an address and a customer

Nvidia has reportedly been trying to connect GPU holders with data-center operators that have available capacity in the Nordics. That matchmaking identifies what a foreclosure document cannot supply: a suitable rack, power, cooling, networking and an operator prepared to turn hardware into a service. A railcar can change shippers only where tracks and terminals connect them. A GPU cannot earn rent merely because its owner has changed.

Customer contracts supply another part of the system. SpaceX signed a computing agreement with Reflection AI worth up to $6.3 billion for access to Nvidia GB300s at Colossus 2, with payments of $150 million a month through 2029. Banks have likewise marketed more than $56 billion in investment-grade data-center construction loans tied to Oracle's future leases—the building side of the AI infrastructure project-finance stack. In both cases, lenders can point to prospective payments as well as physical assets.

A contract's headline value is not necessarily income a lender can count on for the life of a GPU fleet. Anthropic's IPO prospectus described agreements to pay SpaceX up to $84.5 billion for Nvidia-based compute through 2029, but most of those commitments were cancellable on 90 days' notice. Apollo's underwriting must distinguish contracted access from cash flow that survives a customer's decision to leave. The lien may follow the GPUs out of one building; the megawatts and the next customer's signature remain at an address.