AI no longer mobilizes merely engineers and silicon chips; it has become a frontier for financial innovation.
The Financial Times reveals that Google has established an unprecedented financing arrangement, representing nearly $200 billion in contracts, to back the large-scale roll-out of its AI-specialized chips, the Tensor Processing Units (TPU), historically reserved for its own use and developed with Broadcom.
The operation goes far beyond a simple commercial relationship between the world’s third-largest hyperscaler and the inventor of Claude. In addition to Google, it involves Broadcom, Morgan Stanley, and the Apollo and Blackstone funds. It also includes several data-center operators in a financial architecture designed to fund the infrastructure required to train the next generation of AI models.
According to the British daily, about $150 billion of the arrangement concerns the TPU directly.
A model inspired by aerospace financing
The challenge is as much financial as it is industrial. Anthropic shows compute needs comparable to those of the largest hyperscalers, without possessing their financial muscle or credit rating.
To prevent Google or Broadcom from recording tens of billions of dollars of equipment on their balance sheets, the partners chose a financing model drawn from commercial aircraft financing.
The principle works like this: Google sells the TPUs to Broadcom, who then resells them to a special purpose vehicle (SPV). The SPV is financed by private debt provided chiefly by Apollo and Blackstone. The SPV then leases the equipment to Anthropic.
This mechanism converts a massive investment into a lease contract, while spreading the risks among several financiers.
According to the FT, the robustness of the arrangement rests precisely on this distribution of commitments.
Google guarantees the data-center leases intended to host the TPUs. Broadcom provides a residual-value guarantee on the equipment. This means that if Anthropic were to stop paying rents and the chips lose value upon resale, Broadcom would cover a portion of losses for the most exposed investors.
Morgan Stanley structured the financing vehicles, while Apollo and Blackstone supply the bulk of the capital via private debt funds.
Crypto miners pivot to AI infrastructure
The British daily’s investigation describes an initial operation involving $35 billion of equipment, representing roughly one million TPUs, or nearly a gigawatt of compute power.
A second deployment wave would bring an additional 3.5 gigawatts.
Broadcom also notes, in its financial filings, $128 billion of purchase commitments, of which $55.2 billion must be delivered in 2027 and $72.9 billion in 2028. These sums correspond to the TPUs destined for Anthropic.
Financing the chips, however, solves only part of the equation. There remains the need for data centers powered by electricity to operate them.
To speed up deployments, Google is drawing on several firms historically specialized in cryptocurrency mining, including TeraWulf, Cipher Digital, and Hut 8, which already possess substantial electrical capacity now redirected toward AI infrastructures.
According to the FT, Google guarantees the rents for data centers built for Anthropic, enabling them to secure bond or bank financing on more favorable terms.
The daily identifies five projects totaling 1.4 gigawatts of capacity that have already raised $15 billion in debt, while Google would have provided guarantees on ten projects representing 2.4 gigawatts.
Each participant takes a share of the risk
Beyond the sheer volume of compute, this arrangement could also confer a competitive edge to Google Cloud.
The FT cites a Jefferies analysis suggesting that projects backed by Google’s guarantees borrow at a median rate of 7.1%, versus 9.3% for data-center operators building their infrastructures around Nvidia GPUs.
That difference in the cost of capital could, the bank argues, become a structural advantage for Google’s ecosystem.
The investigation underscores the financialization of AI infrastructures. To absorb the explosive demand for computing power, hyperscalers are increasingly relying on structures that blend banks, private debt funds, industrial players, and data-center operators.
