Broadcom’s $42bn Anthropic facility turns chip supply into project finance
The commercial contest for AI accelerators is no longer being fought only through architecture, process nodes, memory bandwidth and software. Balance sheets have entered the specification.
Anthropic’s planned flotation has disclosed an agreement under which Broadcom could provide up to $42bn of financing for the AI company’s infrastructure programme. The facility would help Anthropic meet part of a five-year, $125.2bn commitment to lease tensor-processing-unit computing capacity.
The planned financing could cover roughly one-third of that commitment. Broadcom may appoint another financing partner and the debt instruments could be convertible into Anthropic shares.
Broadcom’s role stretches across several parts of the same transaction. It helps design Google’s TPUs, supplies technology used in the infrastructure, participates in equipment leasing and may provide the finance used to pay for it. Anthropic is expected to become Broadcom’s largest compute customer in 2027.
That circularity does not make the arrangement improper. Suppliers have long financed machinery, aircraft, telecoms networks and industrial equipment. Vendor finance can allow expensive infrastructure to be deployed before it generates enough cash to pay for itself.
The scale and concentration are unusual, however. The financial return depends partly on a customer buying or leasing capacity built around the financier’s own technology. The customer’s ability to repay then depends on demand for AI services growing quickly enough to cover enormous fixed infrastructure commitments.
Anthropic identified the relationship as a potential conflict of interest in its prospectus. Decisions by Broadcom over hardware availability and pricing could affect Anthropic’s ability to obtain the computing capacity on which its business depends.
The documents also show the possible consequences of a default. Certain payment or performance failures could make a substantial portion of Anthropic’s lease commitments immediately payable while restricting its ability to draw on the financing facility intended to meet those payments.
This is a different type of dependency from a conventional component supply agreement. A shortage, late qualification or price increase would not affect only the bill of materials. It could alter the economics of the customer’s infrastructure and the financing behind it.
There is also a question of how clearly the hardware risk has been priced. AI accelerators and networking systems have short commercial cycles. Their useful economic life depends on power efficiency, model architecture, software compatibility and the arrival of newer devices. Equipment that is valuable when a financing agreement is signed may not retain the expected residual value several years later.
Broadcom is not alone in using its financial strength to support demand for computing hardware. What makes the Anthropic arrangement notable is the combination of its size, the customer concentration and Broadcom’s presence on both the supply and finance sides.
If Anthropic’s demand projections are right, the structure could secure a large customer and fund the infrastructure needed to support it. If they are wrong, the separation between semiconductor demand and credit risk will look much less tidy.
