The figure should have stopped the room. A $42 billion convertible loan. Not spread across a portfolio. Directed at a single customer. Broadcom, a chipmaker, lending its own balance sheet to Anthropic so that Anthropic can purchase Broadcom's TPUs. The disclosure — according to a prospectus reportedly not yet registered on SEC EDGAR — arrives alongside $518 billion in compute commitments and a whispered $2 trillion valuation.
I have spent twenty-two years reading code and balance sheets for the same reason: both omit what they cannot afford to disclose. My first instinct, trained by the Parity reentrancy audit in 2017 and the UST feedback-loop post-mortem in 2022, is not to analyze the structure. It is to verify the inputs. The inputs here are missing. Trust is a variable; verification is a constant. This document, if it exists at all, has been confirmed by no one. That absence is the first finding.
Context: The New Ledger
For a decade the AI supply chain was legible. A lab raised equity. It paid a vendor for silicon. The vendor booked revenue. Cash moved one direction. The transaction closed. Every party could be priced independently.
What the reported Broadcom-Anthropic structure describes is something else. Broadcom is no longer selling chips. It is underwriting the purchase of its own chips, leasing back the resulting infrastructure, and holding a convertible claim on the customer's future equity. Supplier, landlord, creditor. Three balance sheets, one counterparty. The vendor has become the customer's bank.
This is not a new instrument. It is vendor financing — the same mechanism that gutted Lucent and Nortel in 1999-2001, when they extended credit to competitive local exchange carriers so the carriers could buy their switches. The architecture is identical. The sector label has changed. The names on the collateral have changed. The failure mode has not.
Core: The Arithmetic of the Gap
Strip the narrative. Run the numbers that are reported.
Anthropic's Q2 2026 revenue is cited at $11.6 billion. Annualized: roughly $46.4 billion. Total compute commitments: $518 billion. Even under a generous five-year amortization, the annualized obligation is approximately $103.6 billion.
The revenue covers 45% of the annualized commitment. The gap is not a rounding error. It is a structural deficit of more than two to one. Code does not lie, but it often omits the truth. So do term sheets. The prospectus reportedly does not disclose the maturity structure of the $518 billion. A five-year reading is the most benign. A ten-year reading lowers the annual obligation but extends the period over which Anthropic must sustain compounding growth. That omission is not incidental — it is the single most important variable in the solvency equation, and it is absent.
Now examine the Broadcom side. The $42 billion convertible note covers roughly one-third of the $125.2 billion, five-year TPU lease — a ratio the document itself concedes. On paper, Broadcom is protected. If Anthropic succeeds, the note converts to equity and Broadcom captures the upside. If Anthropic fails, Broadcom holds senior creditor status and "significant leverage over the infrastructure assets."
That asymmetry is the tell. Broadcom has downside protection. Anthropic absorbs the operating risk. The customer has become the collateral.
There is a second-order question the document reportedly does not answer: how is the $42 billion classified on Broadcom's own balance sheet? If Broadcom recognizes hardware revenue and loan interest simultaneously — lending to a customer to buy its own product — then the "AI revenue" line is partly self-referential. The cash recovery quality is unknown. In 2000, Lucent booked revenue against receivables that never converted to cash. The mechanism rhymes.
The Anthropic side is equally opaque. The cited "adjusted operating profit" turning positive almost certainly excludes the amortization of long-term lease obligations — because "adjusted" earnings routinely do. If the compute cost is capitalized or deferred rather than expensed, the profitability signal is manufactured by accounting choice, not by cash. This is the same maneuver I modeled in the Impermax yield-farming structure in 2020: a reward distribution that was mathematically solvent only when impermanent loss was excluded from the denominator. Remove the exclusion, and the yield was negative from day one.
Contrarian: What the Bulls Get Right
Here the bulls are not entirely wrong, and intellectual honesty requires saying so.
First, the compute race is real. NVIDIA's reported $100 billion commitment to OpenAI and AMD's $5 billion investment tied to MI450 deployment are not fictions. The competitive currency has shifted from silicon performance to the ability to underwrite a customer's capital expenditure. Broadcom is not acting irrationally; it is acting defensively, matching rivals who moved first.
Second, if AI demand compounds as projected, the gap closes. A lab growing revenue at 100% year-over-year halves the ratio every twelve months. The structure is not inherently fraudulent; it is a leveraged bet on exponential demand.
Third, the conflict disclosure itself is a signal. A prospectus that flags the supplier-lender dual role is exercising a caution rare in this cycle. Most issuers bury the conflict; this one names it.
But the counterargument is precise. The bull case requires the growth assumption to hold across the full maturity of the commitments. It requires no demand plateau. It requires the IPO window to remain open long enough to refinance private debt with public capital. The structure does not fail because AI fails. It fails if AI merely slows.
The Kill Switch
Define the failure conditions explicitly, as I do in every review.
Trigger one: Anthropic's quarterly revenue growth decelerates below the rate required to service the annualized commitment. At current coverage, that threshold is low. A single soft quarter compresses it further.
Trigger two: the IPO is delayed past the maturity of the convertible note. Without public-market refinancing, the debt converts or accelerates. Either outcome transfers control.
Trigger three: a payment or performance default accelerates the majority of lease obligations while simultaneously restricting access to the $42 billion facility. The liquidity crisis then resolves through the same counterparty that caused it. Broadcom becomes both the creditor and the rescue.
Trigger four, the systemic layer: if the convertible notes are held by banks and funds rather than by Broadcom itself, the credit exposure has already propagated into the shadow banking system. The identity of the noteholders is undisclosed. That silence is the loudest variable in the model.
Takeaway
Hype builds the floor; logic clears the debris. The reported numbers may be exaggerated; the structure they describe is not. Vendor financing has returned to the AI supply chain, and it has returned under a banner of strategic partnership rather than credit risk.
The question for the next eighteen months is not whether Anthropic can build compute. It is whether the noteholders can be identified before the cycle turns. Follow the EDGAR filing. Follow the maturity structure. Follow the cash, not the revenue.
The ledger is eating its own tail. The only open question is how many hands are holding the other end.