Broadcom May Lend Anthropic Up to $42 Billion for TPUs

Google offered Cloud TPUs by the hour in 2018. Anthropic has now committed to $125.2 billion in TPU leases over five years, with Broadcom reportedly agreeing to lend it up to $42 billion.

Reuters’ account of Anthropic’s IPO prospectus describes Broadcom’s convertible note as an agreement to lend, not a $42 billion payment already made. For Broadcom, the competitive unit is no longer only the component it supplies. Its earlier equipment guarantees expose it to residual-value risk; a drawn note would add direct exposure to whether its customer can carry the capacity it has reserved.

Broadcom was backing the financing before it offered to lend

In June, Bloomberg reported that Broadcom was backstopping senior portions of an Apollo and Blackstone debt package used to buy TPUs for Anthropic to lease. A Broadcom-backed portion of a related deal might yield about 5.75%, Bloomberg reported, versus 8%–9% for a riskier portion. Those are not matched loans, so the gap cannot measure Broadcom’s contribution to the borrowing rate. It does show why a supplier’s promise matters to lenders before the equipment earns revenue.

By September, reporting described Broadcom and Nvidia using residual-value guarantees to support AI-equipment purchases. A residual-value guarantee addresses what equipment may be worth if a buyer cannot use or refinance it as expected. Nvidia had also reportedly offered to rent back unused GPUs from young cloud providers in exchange for a share of revenue. That arrangement addresses idle capacity instead. Neither works like a drawn loan to Anthropic, which would make Broadcom a creditor of the lab itself.

Equipment buyers needed financing; lenders needed protection against assets whose earning power depends on future AI workloads. Suppliers with a stake in getting capacity installed had a reason to provide it. Nvidia’s wider infrastructure-financing portfolio includes equity investments, but its GPU rental backstop shows the narrower point: a vendor can support demand by taking on a risk the buyer would otherwise carry alone.

Google’s TPU went from an hourly product to a five-year commitment

Google first described its custom TPU in 2016. Cloud TPUs entered beta in 2018 at $6.50 per TPU-hour. Anthropic’s five-year lease is not a comparable price quote: it buys a different kind of access at a different scale. Anthropic must find enough paid work over five years to justify reserved capacity; Google’s hourly customers could stop buying TPU hours.

Hardware design is becoming more specific to that work. Google introduced separate eighth-generation TPUs for training and inference in April 2026. In June, OpenAI and Broadcom unveiled Jalapeño, an LLM-optimized inference chip that OpenAI said reached manufacturing tape-out nine months after design began. A design tuned to a serving workload can improve that workload’s economics, while making an unused deployment harder to redeploy at the return its owner expected. OpenAI’s announcement did not establish Jalapeño’s deployment volume, and Jalapeño should not be counted as part of Anthropic’s Google TPU lease.

The distinction matters for Broadcom’s accounts, too. Reuters tied the convertible note to Anthropic’s TPU leases, not to a disclosed purchase of a separable quantity of Broadcom accelerators. Broadcom reported $16.7 billion in AI-semiconductor revenue, up 221% year over year, but that aggregate does not isolate custom-inference chip sales. Broadcom also forecast fourth-quarter revenue below analyst estimates. Financing a customer’s capacity and recognizing revenue from a particular chip remain different events.

Inference efficiency can help the lab and hurt the lease

Investor materials reviewed by the Wall Street Journal showed OpenAI and Anthropic reporting inference costs above half of revenue. Training can be a major expenditure in a bounded project; serving a model incurs costs as customers use it. A lab that reserves capacity for years must therefore bridge a fixed obligation with demand, prices and software efficiency that can change throughout the term.

OpenAI engineers reportedly found a way to more than halve inference costs. That kind of improvement can make each paid request more profitable. It can also reduce the capacity required to serve a given volume of requests. Anthropic benefits from cheaper serving if demand expands enough to use its reserved TPUs; if it does not, better software leaves a larger share of the commitment idle. OpenAI has introduced one- to three-year Guaranteed Capacity contracts for its own customers, one way to secure some revenue duration. The cited record does not establish how much of any lab’s longer infrastructure obligation such contracts cover.

A financed deployment is still contestable

Google was reportedly in talks with Marvell about an inference TPU and a memory-processing unit. The talks were not confirmed, but they limit any claim that a long financing relationship permanently secures the chip-design relationship. Anthropic has also confirmed that it is building an in-house silicon team and pursuing a multi-chip approach. A lender can help a customer secure capacity without controlling every future hardware choice that customer makes.

That leaves a precise condition for Broadcom’s arrangement. If Anthropic fills its reserved capacity with paid inference while its TPU economics remain competitive, supplier capital can support a durable relationship. If those conditions fail, an undrawn note is not a loss—but a drawn one leaves Broadcom exposed to a borrower whose lease still has to be paid. Broadcom can help finance the commitment once; Anthropic’s customers must make it worth keeping, hour after hour.