What Broadcom actually signed up for
Broadcom has agreed to make available up to 42 billion dollars in financing to Anthropic, covering roughly a third of the 125.2 billion dollar, five-year commitment Anthropic already owes Broadcom for tensor processing unit capacity. The instrument is convertible debt: Anthropic can convert what it draws into equity rather than repaying cash, and Broadcom gets to designate a financing partner on the deal. Anthropic deposited cash into a restricted account for Broadcom's benefit back in April 2026, when the broader partnership was expanded, and Broadcom and Google separately agreed to support roughly 3.5 gigawatts of next-generation TPU capacity starting in 2027.
Broadcom has already run this playbook once before with the same customer. In June 2026 it launched an AI infrastructure platform with Apollo and Blackstone that included a 35 billion dollar financing tranche tied to more than a gigawatt of Anthropic compute. Stack the two together and Broadcom has now wrapped tens of billions of dollars of lending around a single customer's ability to pay for the chips that customer is leasing from Broadcom in the first place, a degree of financial entanglement that would raise eyebrows in almost any other supplier relationship.
Why a chip vendor becomes a lender
The logic is straightforward from Broadcom's side. AI semiconductor revenue is projected to hit roughly 115 billion dollars in fiscal 2027 and 230 billion in fiscal 2028, and Broadcom is positioning Anthropic to become its largest custom-chip customer by 2027. Financing the customer's ability to pay locks in that revenue trajectory and keeps a key buyer from shopping lease capacity elsewhere when capital gets tight. It also puts Broadcom in the same position Nvidia and others have taken with their own AI customers: extending credit to the companies that then spend that credit back on your hardware.
For Anthropic, the appeal is equally clear. Compute is the single largest cost line on its path to an initial public offering at a reported valuation near two trillion dollars, and locking in financing from the same company supplying the chips removes a step from an already complicated capital stack. But removing a step is not the same as removing the risk. Anthropic's own IPO filing says it does not expect to sell the notes before completing the offering, which tells you the company itself is treating this as a disclosure item investors will scrutinize, not a footnote.
The conflict Anthropic disclosed itself
Anthropic's IPO prospectus names the risk directly: Broadcom's dual role as hardware supplier and financing partner creates potential conflicts of interest that could affect Anthropic's access to computing infrastructure and the pricing it gets. That is an unusually candid admission for a company trying to sell investors on a two trillion dollar valuation story, and it exists because the structure genuinely does blur two relationships that are normally kept separate. A supplier with leverage over your financing has leverage over your roadmap.
The filing also flags what happens if Anthropic stumbles: default scenarios could accelerate lease obligations while simultaneously limiting Anthropic's ability to draw further on the financing facility. In plain terms, the mechanism that is supposed to give Anthropic breathing room could tighten precisely when the company needs room most. That is a structural feature worth noting, not a hypothetical edge case, because it is Anthropic's own lawyers who wrote it into the filing.
What Wall Street is already pricing in
JPMorgan's Harlan Sur has said Broadcom's long-term AI guidance may still prove conservative if deployments and supply ramp faster than currently assumed, which reads as bullish on the surface and reflects genuine confidence in near-term demand. Bank of America has gone further in the other direction, though, and modeled residual-value exposure on the financing itself, concluding that losses could stay manageable under moderate default assumptions while flagging that exposure rises meaningfully if Broadcom replicates this financing model across more of its large customers over the next several years. Analysts do not typically build default models for arrangements they consider routine or low-risk.
Context matters here too: six major AI customers' combined chip demand now exceeds secured supply across the industry, with roughly 350 billion dollars of AI semiconductor revenue expected across fiscal 2027 and 2028 combined. That scarcity is exactly what makes vendor financing attractive to both sides right now, since it lets suppliers lock in committed buyers and lets buyers lock in capacity ahead of competitors. It is also exactly what makes the structure fragile if demand growth or the underlying model economics disappoint even moderately, because the financing assumes today's scarcity-driven pricing power persists for years.
The pattern enterprise buyers should recognize
Vendor-financed demand has become a recurring feature of the current AI infrastructure buildout, visible across multiple supplier relationships, where suppliers extend credit so customers can keep buying at a pace that supports the supplier's own growth story. The structure holds together as long as end demand keeps growing fast enough to justify it. When growth slows even modestly, the companies with the least transparent capital stacks are the ones that get caught flat-footed, along with anyone who signed a multi-year commitment assuming the vendor's balance sheet was as solid as its revenue chart suggested.
If your organization is negotiating multi-year AI compute or model-access agreements, this deal is a prompt to ask a blunt question of every vendor: how much of your growth is financed by you lending to your own customers, and what happens to my contract if one of those customers defaults. You are not just buying capacity. You are taking on counterparty risk one layer removed, and that layer is getting more crowded every quarter.
