Google Finances TPU Expansion with SPV
The Financial Times reports that Google plans to allocate over $150 billion in TPUs to Anthropic through a special purpose vehicle (SPV), with related long-term contracts totaling around $200 billion, rather than having Anthropic purchase all hardware directly.
The transaction chain is as follows: Google sells TPUs to Broadcom, which places the equipment into a dedicated SPV; funding sources like Apollo and Blackstone provide financing to the SPV through debt, which holds the hardware and offers computing power to Anthropic in the form of long-term leases, while Google guarantees data center-related obligations. This structure separates chip procurement, equipment ownership, customer usage rights, and credit support among different entities.
The first completed transaction was executed by Compute SPV, totaling approximately $35 billion, corresponding to about 1 gigawatt of computing power and around 1 million TPUs. The financing is divided into three layers of debt, with approximately $30 billion of senior debt receiving residual value support from Broadcom; if lease cash flows are insufficient, the SPV can dispose of the TPUs, with any asset sale shortfall borne by Broadcom.
This enhancement of credit risk partially converts the leasing risk of Anthropic, which has no public credit rating, into residual value risk of the equipment and credit risk of Broadcom. The repayment sources for senior debt investors are sequentially Anthropic's rent, proceeds from TPU disposal, and Broadcom's residual value support, thus financing pricing is no longer entirely dependent on Anthropic's independent debt repayment ability.
In April, Google agreed to provide an additional 3.5 gigawatts of TPUs to Broadcom, which will ultimately still be used by Anthropic. Broadcom subsequently disclosed a procurement commitment of $128.1 billion, with $55.2 billion expected to be delivered in fiscal year 2027 and $72.9 billion in fiscal year 2028; insiders say these commitments are almost entirely related to Google's TPU project.
From a market mechanism perspective, this is an AI infrastructure transaction of "equipment manufacturing - asset financing - long-term leasing - residual value guarantee": Anthropic is buying the right to use computing power for the next several years, rather than chip assets; creditors are buying rental cash flows and credit enhancement provided by Broadcom; Google and Broadcom are transforming large-scale hardware demand into orders, data center utilization, and ecosystem lock-in. Beneficiaries include Broadcom, Apollo, Blackstone, data center operators, and the TPU supply chain; those under pressure are Broadcom, which bears residual value support, Google, which provides project guarantees, and Anthropic, which needs to continuously grow Claude's revenue to cover leasing obligations.
Source: Public information
ABAB AI Insight
This structure is closest to aircraft leasing rather than traditional cloud computing procurement. Manufacturers like Boeing, Airbus, and GE have long relied on a model where lessors hold the aircraft, airlines pay rent, assets serve as collateral, and manufacturers or engine suppliers participate in residual value support to drive the expansion of high-priced equipment. The TPU SPV transplant this model to AI: chips do not enter Anthropic's balance sheet, computing customers do not need to make a one-time capital expenditure, and financial institutions obtain debt returns through long-term contracts and disposable hardware. The difference is that the secondary trading, maintenance systems, and residual value history of aircraft have decades of data, while the resale market and performance obsolescence cycle for dedicated AI accelerators have not yet formed an equally mature pricing basis.
The capital path separates and repackages the risks of the three entities. Google ensures the large-scale implementation of its chips through TPU supply and data center guarantees; Broadcom recognizes revenue from selling or delivering customized accelerators and network devices while providing residual value backing for senior debt; private credit capital from Apollo, Blackstone, etc., obtains long-term debt backed by rental cash flows, hardware assets, and large company credit support. Anthropic rewrites one-time equipment expenditure into ongoing leasing costs, shifting capital pressure to the revenue growth curve. The essence of this design is not to eliminate risk but to redistribute risk according to the balance sheets that are most willing and able to bear it.
Historically, the telecom industry also tied long-term capacity demand, project financing, and future utilization rates together during the fiber bubble: network assets can be financed, but if actual traffic and prices do not meet expectations, high fixed costs can quickly backfire on operators and creditors. The expansion of GPU cloud rental company CoreWeave also relies on long-term customer contracts, equipment financing, and cloud infrastructure assets. The TPU SPV is currently in the early stages of AI infrastructure transitioning from "technology companies' self-built capital expenditures" to "private credit assetization"; its key is no longer just the demand for model training but whether an independent secondary computing power and chip disposal market can be established.
Essentially, it belongs to capital concentration: cutting-edge model companies need massive computing power, but their own credit and cash flow may not be sufficient to support hardware purchases worth billions of dollars; thus, tech giants, chip suppliers, and private credit institutions jointly provide equipment, guarantees, and debt. As long as Claude's revenue growth, TPU's replacement value, and data center utilization rates remain valid, computing power can be packaged as a financeable asset close to infrastructure; once model revenue falls below rent, chips quickly become obsolete, or the secondary market lacks buyers, risks will be transmitted back to Google's and Broadcom's balance sheets through residual value guarantees, procurement commitments, and data center assurances.
ABAB News · Cognitive Laws
- Financing does not eliminate risk; it only rewrites risk holders.
- Whether equipment can be securitized depends on whether residual value can be trusted.
- Demand growth is a leasing asset; demand slowdown is credit risk.