$42 billion. That's the convertible note Broadcom is reportedly handing to Anthropic — and it does not appear on SEC EDGAR. Neither company will confirm it. And yet the structure it implies reframes the entire AI capital stack.
Here's the arithmetic the reporting leaves on the table. Anthropic has committed $518 billion to compute. Broadcom's slice is $161.2 billion — roughly 31% of the total. The $42 billion note covers about one-third of a $125.2 billion, five-year TPU lease.
Code doesn't care whether the loan is real. The exposure math is identical either way.
I've spent the better part of a decade auditing token emission schedules against actual revenue. In 2020 I built a spreadsheet that tracked ten DeFi protocols' inflation against real fees. Eighty percent were pure liabilities. The pattern here is the same — only the denomination changed from tokens to teraflops.
Strip the labels and this is vendor financing: a supplier lending its customer the money to buy the supplier's own product. Lucent and Nortel ran this playbook against Competitive Local Exchange Carriers in 1999. The receivables looked like revenue until they didn't. The telecom bubble didn't pop because demand vanished — it popped because the vendors had financed the demand themselves.
Broadcom now occupies three roles at once: hardware vendor, equipment lessor, and lender. Anthropic's own draft prospectus reportedly flags the conflict of interest — a rare instance of an issuer disclosing its structural flaw before a regulator finds it first.
The numbers, as reported:
- Anthropic Q2 2026 revenue: $11.6 billion → roughly $46.4 billion annualized.
- Total compute commitment: $518 billion.
- Single-supplier commitment to Broadcom: $161.2 billion.
- Convertible note: $42 billion, covering about 33.5% of the $125.2 billion TPU lease.
If the $518 billion amortizes across five years, the annual obligation lands near $103.6 billion. That is 2.2x annualized revenue.
AMD is in for $5 billion, tied to MI450 GPU deployment. NVIDIA has committed up to $100 billion to OpenAI. Every major accelerator vendor is now underwriting its customers' capex. The competitive currency shifted from silicon to balance-sheet capacity.
Here's the piece nobody is pricing. When a vendor finances its own customer's purchase, it books revenue and a receivable in the same transaction. The hardware sale hits the income statement; the loan sits on the balance sheet. Cash doesn't arrive from a third party — it arrives from Broadcom's own credit facility, round-tripped back through the same ledger.
Code doesn't reconcile that. Accounting standards do, and only if someone forces the disclosure.
Three structural faults sit under this.
Fault one: the revenue-commitment gap is not a rounding error. $46.4 billion annualized against $103.6 billion of annualized obligation, assuming a five-year term. Anthropic has not disclosed the term structure. That single missing variable decides whether this is aggressive leverage or insolvency on a schedule. A fifteen-year amortization would cut the annual obligation to roughly $34.5 billion — survivable. A three-year term would push it past $172 billion — fatal. The silence on tenor is not an oversight. It's the load-bearing omission.
Fault two: compute leases are rigid, and that kills the flexibility narrative. I built the 2020 emission model around one question — is the token minted against real demand, or against a promise? Leases answer the same question. Self-built data centers are capex: you can pause a construction phase when demand cools. A take-or-pay lease is opex with a capex spine — the cash leaves whether or not the accelerators spin. Anthropic chose leasing over building, which tells me it lacks the capital, land, and power interconnect to build. That's not a strategy. That's a constraint dressed as agility.
Fault three — and this is the one the reporting buried — nobody is modeling the electrons. $518 billion of compute implies gigawatt-scale power and hyperscale cooling. I've reviewed enough infrastructure filings to know that power purchase agreements, substation interconnect queues, and water rights are the actual bottleneck — not chip allocation. The reporting never mentions a single PPA. A compute commitment without a corresponding power commitment is a contract with no physical substrate. Code doesn't run on contracts. It runs on electricity.
Now the multi-vendor angle. Anthropic is drawing from Broadcom TPU, AMD MI450, and potentially NVIDIA — an ASIC-plus-GPU hedge that reduces single-supplier risk on paper. But each supplier arrives with financing attached. The hedge isn't diversification; it's multiple creditors holding claims on the same balance sheet. Broadcom's convertible note converts to equity upside if Anthropic IPOs. AMD's $5 billion converts to GPU orders. Anthropic isn't a customer choosing vendors. It's collateral being shared.
The ASIC migration itself deserves scrutiny. Moving frontier training off CUDA onto a custom TPU compile stack carries a software-migration tax that no press release quantifies — operator libraries, kernel coverage, tooling maturity. Google spent a decade maturing its TPU stack internally. Anthropic is renting one. The performance delta in training, not inference, remains unverified.
Everyone is reading this as a story about Anthropic's ambition. Read it instead as a story about Broadcom's downside protection — and the asymmetry is the tell.
If Anthropic succeeds, Broadcom's note converts and it captures equity upside on top of hardware margin. If Anthropic fails, Broadcom holds senior creditor status plus what the reporting calls "significant leverage over infrastructure assets." Heads, Broadcom wins the equity. Tails, Broadcom wins the collateral. The chipmaker has engineered a position where it cannot lose the principal — only the optionality. Anthropic, by contrast, carries 100% of the operating risk.
That asymmetry is why the "NVIDIA is defending with $100 billion" framing misses the point. Jay Goldberg of Seaport Research calls it defensive competition — NVIDIA forced to match balance sheets rather than benchmarks. The moat isn't performance anymore. It's underwriting capacity. And Robert Leitao at Rothschild & Co named the real bet: the market is pricing in that two companies will generate enough revenue to service an entire financing lattice.
The contrarian read: this isn't a bubble inflating. It's a bubble collateralizing. The risk migrated from equity to credit — from "will the stock fall" to "will the note pay." Credit moves slower and hits harder.
Watch three signals. First: does Anthropic's S-1 actually land on EDGAR, and does it disclose the term structure of the $518 billion? Second: does Broadcom's next 10-Q show a credit-loss provision against the Anthropic exposure — or, tellingly, the absence of one? Third: does the convertible note's interest rate price Anthropic as investment grade, or as venture debt?
Code doesn't price credit risk. Lenders do — eventually. The only question is who is still holding the note when "eventually" arrives.
