Apollo Global Management and Blackstone arranged a $35 billion debt package financing Anthropic's multi-year access to Alphabet's custom Tensor Processing Units through a special purpose vehicle structure that closes this week. The loan represents the largest single financing tied to AI compute infrastructure and marks the first time major credit platforms have underwritten model-training capacity as a discrete asset class separate from corporate balance sheets.
The SPV isolates Anthropic's chip-access rights from its operating entity, allowing the debt to price against utilization rates and training throughput rather than the company's enterprise valuation. Alphabet provides the TPU capacity under a seven-year take-or-pay agreement priced at an estimated $4.8 billion annually. Apollo leads the senior tranche at $22 billion, with Blackstone holding $13 billion in mezzanine debt priced 425 basis points above the senior layer. The structure includes covenants tied to model performance benchmarks and compute efficiency metrics, not traditional cash-flow ratios.
This matters because it demonstrates that non-dilutive infrastructure financing can now compete with equity rounds for frontier AI labs. Anthropic avoids issuing new shares while securing compute capacity equivalent to roughly 18,000 H100 GPU clusters in TPU terms. The debt pricing implies the credit markets value AI training infrastructure at a 12-14% cost of capital, materially below the 25-30% implied cost in recent venture rounds at comparable labs. Family offices and endowments holding Apollo or Blackstone credit funds now have indirect exposure to frontier model development without venture liquidity terms.
The SPV structure also separates compute risk from model risk. If Anthropic's models underperform, the debt still performs as long as the company meets its chip-utilization floors. Alphabet benefits by monetizing its TPU capacity years ahead of typical cloud contracts, effectively securitizing future revenue at a 9-11% discount rate. The banks syndicated portions of the senior tranche to six insurers and three sovereign wealth funds, all of whom priced the paper closer to project finance than technology lending.
Operators and allocators should watch whether Microsoft, Amazon, or Oracle attempt similar structures with OpenAI, Cohere, or other large model shops in the next six to nine months. The Anthropic deal sets a pricing floor for AI infrastructure debt and establishes utilization covenants as the new credit standard. Any follow-on transactions will reveal whether this is a one-off or the beginning of a distinct asset class. Also track whether Alphabet offers similar capacity packages to smaller labs, which would signal the hyperscalers are competing on financing terms, not just on silicon performance.
The debt closes the week of May 19th. The first utilization report under the SPV covenants is due 90 days after funding.