Anthropic’s IPO filing landed this week carrying a number larger than most entire funding rounds, tucked inside the risk section rather than the pitch: Broadcom has agreed to lend Anthropic up to $42 billion. Not to sell it chips. To lend it the cash to lease the chips Broadcom itself helps build.
Reuters reported the arrangement on October 1 off Anthropic’s prospectus. The structure is the story, not the headline figure. For a year this site has tracked AI companies raising equity to buy compute. This is a different machine. The company that designs and co-manufactures the hardware is now also the financier underwriting its customer’s ability to rent it. And Anthropic’s own lawyers flagged the result, in writing, as a conflict of interest.
What the filing actually commits to
The $42 billion comes as convertible notes, meaning the debt can later be swapped for Anthropic equity. The facility is sized to cover roughly one-third of a five-year, $125.2 billion commitment Anthropic has made for tensor processing unit computing capacity, according to the aiweekly brief on the filing. Google designs those TPUs; Broadcom co-develops them; the first next-generation capacity comes online in 2027, the same year Anthropic is projected to become Broadcom’s single largest compute customer.
So Broadcom occupies three seats at once: hardware supplier, lease counterparty, and lender. If the notes convert, it takes a fourth seat as shareholder. The prospectus names the hazard plainly, citing “potential conflicts of interest” and warning that Broadcom’s own pricing and hardware decisions could determine whether Anthropic can secure enough infrastructure to meet its obligations. A company does not usually describe its largest financing partner as a risk to its supply unless the lawyers insisted.
The commitment sits inside a much larger stack. Anthropic disclosed roughly $518 billion in total infrastructure obligations, about 80 percent of it non-cancellable, with restricted cash already deposited for Broadcom’s benefit back in April 2026. Against that stands a company that booked nearly $4.6 billion in 2025 revenue and an operating loss topping $8 billion, now seeking a valuation above $2 trillion. Two customers each account for 12 percent of revenue, and 47 percent of sales flow through the Amazon and Google marketplaces. The obligations are concrete and front-loaded. The revenue to service them is projected.
Why this reads as vendor financing, not a vote of confidence
A $42 billion loan from your chip partner looks like conviction. Read the mechanics and it looks more like a company moving margin around to keep demand for its own silicon visible on a schedule.
Mizuho estimated Broadcom will earn roughly $21 billion from Anthropic in 2026 and about $42 billion in 2027, per FinanceFeeds. The 2027 figure matches the loan ceiling almost exactly. Put differently, Broadcom is lending Anthropic close to one year of the revenue Broadcom expects Anthropic to pay it. The chipmaker’s broader AI guidance, roughly $115 billion in AI semiconductor revenue for fiscal 2027 and about $230 billion for 2028, rests substantially on this class of customer continuing to spend. Lending the customer the money to spend keeps the projection intact.
Markets price cash-paid chip sales and vendor-financed ones differently, because vendor financing trades some margin for added credit risk and lumpier outcomes. That is why the detail belongs in a filing’s risk section and not its growth story. Broadcom shares barely moved on the disclosure, trading around $351, still well below their June high, which suggests investors have already absorbed that a chunk of the AI order book is being underwritten by the sellers.
This is not an isolated structure. It is the dominant one. In June 2026 an Apollo and Blackstone facility worth about $35 billion financed Google chip leases with Broadcom backstopping the senior tranches. Nvidia spent the summer weighing a financing guarantee for OpenAI that reignited the same circular-deal worry. One tracker now puts total AI circular financing near $800 billion, with Broadcom leading vendor commitments at roughly $350 billion. Google has separately committed up to $40 billion of its own to Anthropic, part of the same loop this site traced when it covered Anthropic’s sprint toward a $2 trillion IPO against OpenAI’s retreat from its own. The Broadcom loan is that loop tightening one more turn, now disclosed under oath ahead of a public listing.
The real exposure is correlation
Circular financing works fine while the music plays. Customers use the capacity, generate cash, service the leases, and everyone’s projection holds. The structure’s weakness is that it couples risks that used to be independent.
The clearest tell is a single clause in the prospectus: a payment or performance default could accelerate Anthropic’s lease obligations and, at the same time, restrict its access to the very $42 billion facility meant to cover those payments, as the Finimize breakdown highlighted. The backstop and the obligation are wired to the same switch. Stress the one and you lose the other precisely when you need it. That is not a hedge. That is a cliff with a handrail bolted to the edge that falls away with you.
Robert Leitao of Rothschild called the setup “quite a concentrated bet” on two companies generating enough revenue to make the numbers work. He is describing systemic exposure, not a company-specific one. Broadcom’s AI guidance, Anthropic’s lease schedule, and the convertible notes all resolve to the same question of whether end demand for Anthropic’s models grows fast enough, soon enough. If it does, the structure prints money for everyone. If demand merely slows, the supplier, the lender, the lessor, and the prospective shareholder are all the same balance sheet taking the hit four ways.
What a buyer should take from a filing
If you buy AI capacity or build on a frontier lab, this is not a spectator’s financial-engineering story. It is an entry for your own risk register.
I spent two decades in enterprise IT operations, much of it on procurement and vendor management at a large telecom, and the hardest lesson there was never about price. It was about concentration. A supplier who is also your financier, your landlord, and your single source is not a vendor relationship; it is a single point of failure wearing four hats. We learned to ask who stands behind a provider’s capacity, not just what the provider charges, because when an upstream dependency fails it rarely fails politely or alone. An AI lab financed by its own chip supplier is that exact shape, now at eleven-figure scale.
Three practical moves follow. First, read your provider’s compute-financing structure as part of your due diligence, not as background noise; a prospectus that flags its largest partner as a conflict of interest is telling you where the correlated risk sits. Second, keep a second model and a second provider genuinely qualified across your real workloads, because the failure mode in a circular structure is lumpy and correlated, and a lumpy upstream default is your outage, not just an investor’s loss. Third, price the counterparty risk the way you price everything else, against your own usage, the way we have argued about reading the AI build-out as committed liability rather than announced ambition.
The $42 billion number will dominate the coverage. The sentence that matters is the one Anthropic’s own lawyers wrote: the partner lending it the money to buy compute could also decide whether it gets enough. The AI buildout is increasingly financed by the people selling the shovels, and that arrangement is only as stable as the demand it assumes. This filing, like the Pentagon’s loan to the company building the data centers and the compute-equals-GDP thesis behind the whole race, is a reminder that the money and the hardware now move in the same small circle. When you depend on anyone inside it, you have inherited the circle’s risk, whether or not it shows up on your invoice.