Broadcom Enters Anthropic’s TPU Credit Stack

Broadcom backed payments on senior debt used to buy TPUs for Anthropic to lease. Earlier reporting put the prospective yield on a backed portion near 5.75%, versus 8% to 9% for a riskier portion. Lenders had to assess the chip supplier’s payment promise alongside the lab’s lease obligation.

Broadcom reported $16.7 billion in AI semiconductor revenue for its third quarter, up 221% from a year earlier. Chip sales explain why Broadcom matters to AI infrastructure. They do not explain why lenders would look to the company for protection on a customer’s compute leases. In the coverage record, Broadcom appeared in 27 articles in the earlier comparison period and 56 in the later one. That measures attention, not funded credit exposure.

Broadcom has tied its balance sheet to the long-term capacity its chips help enable, adding customer credit risk to the economics of a custom-silicon win.

Anthropic’s capacity contract gave lenders a supplier to call

In April, Broadcom agreed to produce future versions of Google’s TPUs as its arrangement with Anthropic expanded to roughly 3.5 gigawatts of computing capacity. The arrangement connected chip production to a large, continuing requirement for capacity rather than a single component order.

Anthropic’s IPO prospectus later disclosed a different kind of commitment. Broadcom agreed to lend Anthropic up to $42 billion through a convertible note that could help finance a $125.2 billion, five-year TPU lease commitment. The note’s ceiling equals roughly one-third of the lease obligation. That ratio describes potential support, not money already advanced.

Apollo and Blackstone also finalized a reported $35 billion package to fund TPUs for Anthropic to lease. Broadcom backstopped payments on its largest senior portions. The note and the backstop are distinct commitments; adding their headline amounts would not establish Broadcom’s total funded exposure. Together, they show how Anthropic’s capacity requirement reached beyond the lab and Google to a silicon supplier and private lenders.

Broadcom’s backstop changes the price of a lender’s claim

Earlier reporting on a roughly $36 billion TPU debt deal put the prospective yield on a Broadcom-backed portion near 5.75%, compared with 8% to 9% for a riskier portion. The reported amounts describe stages of the financing, not two completed packages to sum. The tranches also carry different risks, so the yield gap cannot be assigned wholly to Broadcom’s support.

Apollo and Blackstone could offer different claims against the same broad buildout, with different payment protection. Broadcom’s backstop made the supplier relevant to repayment even though Anthropic, not Broadcom, would use the leased compute.

Other vendors have taken related risks through different instruments. Nvidia and Broadcom have used residual-value guarantees to support customer purchases; Nvidia’s infrastructure financing also extends beyond ordinary chip sales. A payment backstop addresses a borrower’s ability to pay. A residual-value guarantee addresses what hardware may be worth later. Neither risk appears in a chip benchmark.

A custom design win can narrow the exit

OpenAI and Broadcom co-designed the Jalapeño inference chip, using OpenAI models in the design process. OpenAI was reported to have committed $10 billion in orders for the chip’s internal use. That is a closer relationship than choosing an interchangeable accelerator from a catalog: the buyer helps specify the hardware and then needs enough work for the resulting fleet.

Financing discussions exposed the commercial constraint. OpenAI and Broadcom reportedly discussed roughly $18 billion for initial chip production, conditional on Microsoft buying about 40% of the chips. Microsoft’s proposed share matters because it would give the prospective production run another buyer. OpenAI and Broadcom were still discussing the financing when that account was published.

Anthropic’s TPU leases and OpenAI’s Jalapeño chip are separate arrangements. They share an asset-specificity problem: a chip tailored to one set of workloads needs committed use, and a supplier has fewer easy alternatives if that use falls short. Credit support can help secure capacity before the customer has demonstrated years of paid utilization, but it cannot create that utilization.

Nor does custom silicon guarantee Broadcom a permanent place in every design. Anthropic is building an in-house chip team and pursuing a multi-chip approach. Alibaba has reportedly delivered more than 100,000 of its own training-and-inference ASICs. Customers can use internal design to reduce dependence on an external supplier, even while they need partners to manufacture or finance capacity today.

The note’s missing terms matter more than its ceiling

Investor materials reportedly showed inference costs exceeding half of revenue at both OpenAI and Anthropic. That measure does not establish whether either lab can meet a particular lease payment. It identifies the operating expense that must leave enough room for one. Broadcom CEO Hock Tan has separately warned about access to capital, borrowing costs and uncertain returns from AI investment.

Anthropic’s prospectus establishes an up-to-$42 billion note, not a $42 billion draw. Anthropic has not disclosed whether it has drawn on the note or detailed its pricing, maturity, collateral, seniority, conversion mechanics or accounting treatment. Reports that Broadcom is trying to arrange more than $50 billion for OpenAI’s custom chip remain unconfirmed. Neither figure should be treated as cash already at risk.

Broadcom’s $16.7 billion in AI chip revenue records demand for silicon; the note’s missing terms leave the cost of its customer-credit exposure unknown. The prospective 5.75% yield belongs to a lender’s claim, not a chip sale. Broadcom’s name behind that claim is the part a semiconductor revenue line cannot show.