SpaceX’s $40B Rubin GPU Financing Enters Syndication

By Rebecca Stone

SpaceX has structured a roughly $40 billion GPU lease financing package, with closing targeted before the end of 2026 and Nvidia Rubin hardware due in the first half of 2027.

Goldman Sachs Group and Apollo Global Management are talking with investors about selling portions of the package, according to Bloomberg News; Reuters and The Globe and Mail carried the same account. The shorthand says the banks are “seeking buyers,” but no completed sale to a named buyer has been confirmed.

What does the financing actually fund?

SpaceX described the arrangement as “the largest of its kind for GPUs on record.” That claim is not independently benchmarked. The multi-tranche structure aims to keep debt off SpaceX’s balance sheet: an Apollo affiliate would buy chips and server hardware, then lease them to SpaceX for a little over five years.

The rent payments are expected to total roughly $40 billion in principal. Separately, the financing and a $7 billion lease prepayment are meant to fund approximately $47.3 billion of GPU equipment. The reports do not reconcile the two totals: one measures principal, the other equipment cost.

Goldman leads the loan portion, Apollo is the key bond investor, and Pacific Investment Management Co., or Pimco, also participates. Named pieces include an $8.3 billion senior-secured loan for syndication to banks and the $7 billion prepayment; reporting does not itemize the remainder.

Deal documents identify the hardware as Nvidia’s latest Rubin chips. SpaceX said the fixed rate had not been determined at pricing, leaving investors to price delivery and credit risk without a final borrowing cost.

How does the lease keep financing separate?

Ownership would sit with the Apollo affiliate, while SpaceX would act as lessee. The reports do not specify how auditors or regulators will classify the leases, so their intended balance-sheet treatment remains unproven.

The schedule is tight: expected closing before the end of 2026; delivery in the first half of 2027; and a little over five years of rent from delivery, while a Bloomberg-reviewed presentation describes five years.

Lenders therefore commit capital months before any Rubin chip arrives. Supply, server or site delays would defer rent, making the eventual fixed rate a signal of perceived risk. Fixed rent does not shrink if compute demand falls; SpaceX must use the newest chips enough to meet the commitment.

No residual-value assumptions appear in the reports. Lenders instead rely on the $7 billion prepayment, security behind the senior loan and the credit of one prominent tenant.

Supply may bind before money does: Samsung has reportedly sold out its 2027 HBM supply as HBM4 sales tripled. Financing cannot manufacture accelerators or memory, but it can reserve a delivery slot.

Who is buying the exposure?

Goldman plans to distribute the $8.3 billion loan to banks. Apollo is the key bond investor, and an Apollo affiliate owns the leased equipment. Pimco’s presence points toward large institutional credit investors; its allocation is undisclosed.

Some investors were reportedly asked for stakes of up to $500 million before formal syndication. That figure comes from unnamed people familiar with the matter, not a Goldman, Apollo or SpaceX announcement. An ask is not a commitment; up to is a ceiling rather than an average. No named investor has been confirmed.

Each side gets something: the bank earns fees and a distribution mandate, the credit investor receives rent claims backed by physical assets, and SpaceX gets hardware without borrowing directly for it. The closest parallel is Google’s backstop for $19 billion of TeraWulf leases supporting Anthropic, where a stronger third party helps finance another company’s hardware and sites.

Cited funding of $25 billion at Nvidia and $45 billion to $50 billion at Oracle shows large AI buyers combining debt and equity. SpaceX extends that shift from venture equity and cloud rental toward specialist lessors, following the playbook used in real estate, aircraft and shipping.

What risks remain?

• Rate: The fixed rate remains unset.
• Demand: No buyers are confirmed; the $500 million ticket is unconfirmed.
• Accounting: Final lease classification is unspecified.
• Delivery: Hardware and rent payments depend on first-half-2027 delivery.
• Technology: Newer GPU generations are expected during the five-plus-year term.
• Concentration: One tenant and one Nvidia family support the full $40 billion.

What does the source expect next?

The source offers five forecasts, explicitly its own judgment rather than claims from a named analyst.

• Close: The deal lands in 2026; rate negotiations pose the main risk.
• Benchmark: The fixed rate becomes a reference for GPU-backed finance.
• Copies: A successful placement prompts more lessor-owner structures.
• Supply: Allocation, not capital, constrains Rubin in early 2027.
• Disclosure: Investors seek residual-value, utilization and tenant-obligation data.

What does it mean for developers and cloud buyers?

Most developers will never touch this fleet, but the reports do not specify whether SpaceX will use it for training, inference or another workload. Any capacity effect belongs to 2027 at the earliest.

Large buyers can reserve scarce capacity years ahead, while smaller buyers relying on spot supply may find fewer cheap options. Cited Bedrock pricing shows the spread: Priority costs 75% more and Flex 50% less. The practical response is to reserve capacity where economics work, model multiple price tiers and watch lender disclosures and lease pricing.

For Nvidia, the arrangement provides funded demand and, in effect, a delivery slot. The reports give neither transaction revenue nor a shipment count, so they cannot quantify Nvidia’s results. A cited $289 billion Morgan Stanley forecast for SpaceX keeps the commitment under scrutiny of AI spending.

For lenders, distribution tests appetite for lease-backed AI infrastructure. A separate cited scenario puts $1.5 trillion of AI capex at stake as investors weigh congressional control and interest rates. A successful syndication would validate large-scale GPU leases; a slow book or high coupon would expose their limits. The next hard marker is the fixed rate: clear demand would reinforce financing as a way to secure 2027 Rubin capacity, while weak demand would raise that capacity’s cost and slow further GPU-backed deals.