Broadcom Becomes Anthropic's Compute Banker

Broadcom has agreed to lend Anthropic up to $42 billion through a convertible-note arrangement to finance infrastructure spending, according to Anthropic’s IPO filing. The figure is large enough to invite the wrong shorthand. It is not $42 billion of chips delivered, and it is not a completed cash injection. Anthropic says it does not expect notes to be sold before its IPO.

The arrangement matters because of who sits on each side. Broadcom is already central to the TPU supply chain that Anthropic plans to use through its partnership with Google. Now it may also designate a financing partner, lend against the expansion, and convert the debt instruments into Anthropic shares. The supplier is moving closer to the customer’s capital structure.

That is the Culled layer: AI infrastructure is no longer simply a contest over who builds accelerators. It is becoming a system in which the company that translates a design into silicon can also help finance the buyer’s long-term commitment to use it.

The facility funds a commitment, not a machine order

Reuters reported that the notes could finance roughly one-third of Anthropic’s $125.2 billion, five-year commitment to lease TPU compute capacity. The next-generation capacity is expected to begin arriving in 2027. That timing matters. A lease commitment creates an obligation before it creates usable output or revenue, and a financing facility gives the customer a way to carry that obligation.

But the disclosure also puts limits around the story. Anthropic said certain payment or performance defaults could make a substantial part of its lease obligations immediately due while limiting its ability to use the facility to cover those payments. In other words, the fallback finance is not an unconditional escape hatch. It is part of the same contractual architecture.

The framing is more precise than “Broadcom bankrolls Anthropic.” Broadcom has agreed to a potential financing role linked to a defined infrastructure relationship. Whether notes are issued, when they are issued, and what they convert into remain future decisions with IPO conditions attached.

One counterparty now touches three layers

The physical layer is familiar: Broadcom works with hyperscalers and model labs on custom silicon, networking, and TPU-related systems. As Culled argued in the custom-silicon shift, this position can be more durable than a single chip cycle because it sits behind multiple competing AI architectures.

The financial layer is newer. If the notes are drawn and converted, Broadcom could have supplier revenue, financing exposure, and an equity interest in the same lab. That alignment can make a massive buildout easier to execute. It can also concentrate dependencies that are usually separated between a manufacturer, a bank, and an investor.

Anthropic itself flagged potential conflicts over access to AI infrastructure. That is not boilerplate worth skipping. When capacity becomes scarce, a provider that is also a creditor has more information about the customer’s commitments and more influence over the timing of financing. The relationship may be commercially rational; it is still a tighter knot than an ordinary equipment purchase.

The new unit of AI competition is not only the chip or the model. It is the contract that binds compute, credit, and future equity together.

The test comes before the IPO

The immediate test is not the $42 billion headline. It is whether Anthropic begins to draw on the facility and how its restricted-cash obligations evolve as TPU capacity ramps. The second test is whether the IPO creates a public equity value that makes conversion attractive rather than merely possible.

For Broadcom, the upside is clear: a customer capable of financing a multi-year compute program is a more reliable buyer of the silicon and network infrastructure it helps enable. The risk is that a slowdown in model demand, a delayed capacity ramp, or an IPO below expectations transmits stress across more than one relationship at once.

The proposed facility therefore belongs in the AI-infrastructure story, not merely the venture-finance column. It is another sign that the industry is inventing its own capital stack because the bill for frontier compute has outgrown ordinary procurement.