GPUs on Credit: SPVs, Debt and Leasing — the New Finance of the AI Datacenter

Three deals in two days tell the same story: Broadcom and its banks are structuring a $63 billion vehicle to buy accelerators and lease them to Anthropic, SpaceX is closing roughly $40 billion of GPU-backed debt with Apollo, and Meta is preparing equity raises whose total program could reach $95 billion. AI infrastructure is no longer just bought outright: it is financed like an aircraft fleet.

The Broadcom-Anthropic structure is emblematic: a special-purpose vehicle backed by Bank of America, Blackstone, Citi and Morgan Stanley will buy XPU accelerators and lease them to Anthropic, with deliveries starting in 2027 and Broadcom supporting residual value. The vendor sells, the lessee avoids balance-sheet immobilization, and lenders carry a technology asset whose lifespan they must assess.

This is aircraft leasing logic applied to compute: an expensive, standardized asset with resale value and long-term contracts. SpaceX follows the same path with roughly $40 billion of debt to equip its own systems. Even NVIDIA is helping finance its customers, with $3 billion invested in Meta — whose 2026 capex is expected between $135 and $145 billion.

How Financialization Reshapes Risk

Risk does not disappear — it shifts: toward lenders carrying GPUs at residual value, toward lessees whose solvency depends on still-young AI revenues, and toward a chain that has become circular, with the supplier financing its own customer. Elon Musk himself warns that credit is shutting off for neoclouds and that consolidation is coming.

For reading the cycle, these structures are a leading indicator: when debt organizes around an asset, the market deems it durable; when credit conditions tighten, the entire construction chain slows down. The cost of capital becomes as important a variable as the price of a megawatt.

What This Changes for Investors

First consequence: datacenter demand no longer depends solely on hyperscaler balance sheets, but on private credit and banks' appetite for AI assets. SPV and leasing structures let new operators emerge without massive equity — with counterparty risks that are more complex to analyze.

Second consequence: GPU residual value becomes a central issue. If generations succeed each other quickly, an asset depreciated over five or six years may lose its value before being repaid. Infrastructure investors must now read lease contracts the way they read PPAs: duration, indexation, guarantees, residual risk.

How We Apply This at FINXIA

In the TITAN DC AI strategy, we favor infrastructure real estate — buildings, power, grid connections — whose value does not depend on a generation of silicon. Our counterparty analyses now incorporate the tenant's financing structure: a fully debt-financed customer does not offer the same security as a hyperscaler investing from equity.

We also read these structures as demand signals: a $63 billion vehicle tied to 2027 deliveries means datacenter capacity that must be ready on time. GPU financialization accelerates the timeline — and strengthens the value of prepared sites, our core business.

Key Takeaways

AI infrastructure is entering the era of structured finance: bank vehicles, GPU-backed debt, leasing and giant equity programs. This financialization accelerates deployment, but shifts risk toward credit and the residual value of technology assets.

Worth watching: neocloud credit conditions and the first consolidations, pricing on the next GPU-backed vehicles, and whether debt-financed tenants honor their capacity commitments.