Test report DSG-8877 · Rev D · tested October 30, 2026

AI Datacenter InfrastructureDevice under test

Broadcom Backs $63 Billion Anthropic Chip Deal With $43 Billion Risk Buffer

Broadcom signed a $63 billion financing deal with BofA, Blackstone, Citi and Morgan Stanley to lease XPUs to Anthropic, backstopping up to $43 billion in residual value risk.

Read
4 min
Words
732
Node
3nm
Operator
Elena Vasquez

Spec summary

  1. Broadcom signed a $63 billion (83.5 trillion won) financing deal announced on the 26th with BofA, Blackstone, Citigroup and Morgan Stanley.
  2. Broadcom's residual value support could reach up to $43 billion (57 trillion won) in potential maximum exposure.
  3. The deal funds about 1.7 GW of the 3.5 GW TPU-based computing capacity Anthropic agreed to in April, with supply starting in 2027.
  4. Blackstone solely provides the subordinated loan funds; BofA, Citigroup and Morgan Stanley arrange the senior loans.
  5. CEO Hock Tan said demand for AI computing is growing at an unusually rapid pace and capital exists to support it.
Broadcom leads 83.5t-won financing, backs Anthropic AI chip risk - Chosunbiz
Fig. ABroadcom leads 83.5t-won financing, backs Anthropic AI chip risk - Chosunbiz — bsdrouin / pixabay

Broadcom has signed a $63 billion (about 83.5 trillion won) financing deal with Bank of America, Blackstone, Citigroup and Morgan Stanley to fund Anthropic's AI computing buildout — and it's willing to absorb up to $43 billion of potential losses to get the money moving.

The company announced the agreement on the 26th local time. The structure is a lease, not a purchase: Anthropic rents the compute instead of buying the hardware outright, and the banks get Broadcom, not Anthropic, standing behind the equipment's long-term value.

How does the deal actually work?

The transaction runs through a newly established special purpose vehicle. The financial firms provide loans and investment funds to the SPV, which uses the cash to buy Broadcom's custom AI accelerator, the XPU, plus networking equipment. The SPV then leases that gear to Anthropic, which pays rent to secure the compute resources it needs to develop and operate its AI models.

Financing splits into two layers. Bank of America, Citigroup and Morgan Stanley arrange the senior loans, while Blackstone alone provides the committed funds for the subordinated tranche. If recovery of funds runs into trouble, senior lenders get repaid before the subordinate ones — standard waterfall mechanics, but at a scale that dwarfs most leveraged deals.

What is Broadcom actually guaranteeing?

Not the debt. Broadcom's commitment is what the briefing materials call "residual value support," and the distinction matters for anyone pricing this deal.

Here's the mechanics: if the residual value of the semiconductors and equipment falls below a certain level — one calculated from the amounts recoverable under future lease contracts — Broadcom makes up the difference. That's separate from guaranteeing repayment of the full amount the financial firms lend.

The numbers are large. According to Broadcom's transaction briefing materials, the potential burden from residual value support could reach up to $43 billion, roughly 57 trillion won. To put that in context, the exposure ceiling equals nearly 70% of the entire financing package. Broadcom projects the burden will gradually decline as lease payments come in, and the company stresses the $43 billion figure is a potential maximum exposure, not a confirmed cash expenditure.

Where does this fit in the Google-Anthropic picture?

This isn't a standalone deal. It executes part of a broader AI computing infrastructure supply contract previously signed by Broadcom, Google and Anthropic.

The backdrop: Broadcom is co-developing the tensor processing unit (TPU), a semiconductor built for AI computation, together with Google. In April, the three companies announced a collaboration under which Anthropic will secure roughly 3.5 gigawatts (GW) of TPU-based AI computing capacity through Broadcom starting in 2027.

The new financing supports construction of about 1.7 GW of that capacity — just under half of the total April commitment. Supply of the semiconductors and networking equipment is scheduled to begin in 2027.

Why are suppliers now sharing hardware risk?

The pattern behind this deal is spreading. As AI corporations pour astronomical sums into securing data centers and semiconductors, a new model is taking shape: financial firms back the capital expenditure, and the equipment supplier absorbs part of the risk that the hardware loses value.

For the financial firms, the logic is straightforward. They supply funds against long-term lease contracts, which gives them a contracted revenue stream rather than exposure to a chip buyer's balance sheet. For semiconductor vendors like Broadcom, sharing the residual-value risk lowers customers' purchase burdens enough to make large-scale supply contracts executable.

Put simply: Anthropic gets roughly 1.7 GW of compute without a $63 billion line item, the banks get a leased asset with a vendor floor under its value, and Broadcom secures a massive deployment of its XPU and networking portfolio. Somebody has to believe the chips hold their worth — and Broadcom has put up to $43 billion behind that belief.

What did Hock Tan say?

Broadcom CEO Hock Tan framed the deal as evidence that the capital markets can keep pace with AI's compute demands. "Demand for AI computing is growing at an unusually rapid pace," Tan said, adding: "this transaction shows that capital exists to support such growth."

For practitioners, the takeaway is that lease-based AI infrastructure financing has now cleared the $60 billion mark with a top-tier vendor guarantee attached. Expect more of these structures as the 2027 supply window approaches.

via Google News: AI chip (Source)

Filed under

  • broadcom
  • anthropic
  • ai-infrastructure-financing
  • custom-ai-accelerators
  • google-tpu
Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.

498 articles

Same lot · LOT-C204

« Previous articleNext article »