Forty-two billion dollars. That is the loan. Broadcom extends it to Anthropic. Anthropic routes roughly a third of it back into Broadcom's TPU rental contracts. Broadcom books the hardware sale and the interest on the same ledger, in the same quarter.
Sit with that. The vendor sells the shovel, underwrites the purchase, and collects rent on the mine. That is not a supply chain; it is a bank wearing a fab's logo.
Now the audit. The filing behind this story has never surfaced on SEC EDGAR. Broadcom declined comment. Anthropic declined comment. The dataset carries an October 2026 timestamp — a future date relative to anything I can independently verify. So state the boundary plainly: I cannot confirm the digits, and neither can you. What I can confirm is the architecture, because the architecture already exists — Nvidia's reported capital commitments, AMD's investment-for-deployment structures, the Google-Broadcom TPU lineage. The pattern is live. The numbers are hearsay.
That gap is the entire trade. Narrative follows logic, never precedes it.
The Vocabulary of Vendor Financing
Vendor financing is not new. In 1999, Lucent and Nortel extended credit to CLECs — competitive local exchange carriers — so those carriers could buy Lucent and Nortel switches. The revenue landed on the equipment makers' income statements immediately. The cash arrived later, or never. When the CLECs defaulted, the write-downs ran into the tens of billions and the telecom bubble collapsed under its own receivables.
The Anthropic structure is the same machine with better branding. Broadcom is simultaneously three things: hardware supplier, equipment lessor, and lender. Anthropic's draft prospectus, per the report, discloses the conflict itself — a rare case of an issuer flagging its own structural flaw before regulators do. When the borrower tells you the lender sits on both sides of the table, believe the borrower.
The mechanics, as reported: $125.2 billion in five-year TPU compute leases; $161.2 billion committed to Broadcom alone, roughly 31% of Anthropic's total $518 billion in compute obligations; $42 billion in convertible notes covering about one-third of the Broadcom lease book. Against that, Anthropic's Q2 2026 revenue is cited at $11.6 billion — call it $46.4 billion annualized.
Run the arithmetic and the story breaks open. And note the choice that precedes the arithmetic: Anthropic rents rather than builds. That single decision converts capital expenditure into operating leases — faster to scale, but it hardens fixed costs into rigid annual cash obligations. A company that rents its compute cannot pause the bill when demand softens. It can only default.
What makes this dangerous is not the size of any single deal but the reflexivity of the whole stack. Broadcom's return is tied directly to Anthropic's eventual IPO — the chipmaker has stopped selling chips and started taking equity-like exposure to its customer's exit. That is not a supplier relationship. That is a joint venture where one side wrote the check and the other side signs the note.
The Gap Is the Signal
Here is the number nobody leads with. Annualized revenue: roughly $46 billion. Total compute commitments: $518 billion. Even spread evenly across five years, the annual compute obligation is about $103.6 billion — roughly 2.2 times annualized revenue. If the commitment term is longer, the annual burden eases; if shorter, it detonates. The report never discloses the maturity ladder. That missing variable is the single most important number in the entire structure, and its absence is itself informative.
I spent 2017 auditing whitepapers for tokenomic fallacies — 50-plus of them — and I published the result as "The Zombie Chain" because 80% had no utility under the vesting schedule. The lesson never left me: the vesting curve, not the vision deck, decides who gets diluted. The same discipline applies here. A compute commitment is a vesting schedule denominated in silicon. If the revenue curve cannot cover the obligation curve, the equity is a residual claim on a shortfall.
Now the accounting. When Broadcom lends $42 billion to a customer to buy Broadcom product, the hardware revenue and the loan asset both appear on Broadcom's balance sheet. That is textbook circular revenue — the cash inflow is Broadcom's own credit, reclassified as a sale. Lucent did it. Nortel did it. The pattern is so reliable it should carry a warning label.
Auditing the code, not the charisma. Ask three questions the report avoids. How is the $42 billion classified on Broadcom's books — held-to-maturity, available-for-sale, or an equity-method stake? Has Broadcom booked a credit provision against Anthropic exposure? And does Anthropic's "adjusted operating profit turning positive" exclude the lease amortization — because "adjusted" almost always means the long-term lease cost was stripped out to flatter the line.
This is where crypto analysts have an edge, and where the AI crowd is still learning. In DeFi, we have watched circular collateral for four years. When a protocol lends against its own governance token, the yield is real and the liquidity is fictional. Yield is the lie; liquidity is the truth. Broadcom's AI revenue line carries the same tell: if a meaningful share of that revenue is self-originated credit rather than third-party cash, the earnings quality is uninvestable at any multiple.
Consider the DeFi parallel I keep returning to. When Uniswap V4 introduced hooks, it turned the DEX into programmable Lego — and the complexity spike scared off the majority of developers who could not audit their own integrations. Vendor financing is the TradFi version of that same complexity: each layer of the structure — supplier, lessor, lender, equity holder — adds a failure mode that no single participant can see end-to-end. Complexity is not sophistication. It is unmodeled risk.
The physical layer gets the same treatment. A billion-dollar compute lease implies gigawatt-scale power, land, and cooling — the true hard constraints on any scaling story. The report never mentions them. Nor does it mention the software lock-in: even if Anthropic trains on custom TPUs, its inference stack, tooling, and talent still orbit CUDA. De-Nvidia-fication is a slide in a deck, not a state of the world.
And the concentration is not incidental — it is the design. Three chipmakers now compete on the currency of underwriting, not silicon: Nvidia reportedly committing up to $100 billion, AMD anchoring $5 billion to Anthropic's MI450 deployment, Broadcom layering credit onto TPU leases. The competitive axis has moved from "whose chip is faster" to "whose balance sheet can absorb the customer's capex." That is defensive competition dressed as offense. Arbitrage exposes the cracks in consensus, and the consensus here is that compute demand justifies any financing structure. Consensus is exactly what gets repriced.
One question sits underneath all of it: who ultimately holds the $42 billion in convertibles? If the answer is banks and funds, the risk has already spread into the systemic financial system. If the answer is Broadcom itself, the risk is concentrated on a single balance sheet. Either way, the AI narrative is now collateralized by credit — and credit reprices faster than compute demand.
Zoom out and the ethical layer is not about model safety at all. It is about incentive distortion. When a hardware supplier is also the primary lender, the borrower's incentive shifts from product innovation to debt service and asset utilization. For a company that markets itself on responsible AI, that is mission drift with a balance sheet attached: every dollar of debt service is a dollar of pressure to commercialize faster and cut safety redundancy. The risk does not stay private. If the structure unwinds, the losses land on pension funds, insurers, and retail holders of the convertibles and the equity — private risk, socialized.
The Landlord, Not the Lab, Holds the Optionality
The reflexive read is that Anthropic wins — it gets capital it could not raise, at speed, without building data centers. Wrong framing. Look at who holds the asymmetry.
Broadcom's downside is protected. If Anthropic succeeds, the convertibles convert and Broadcom captures equity upside on top of hardware margin. If Anthropic fails, Broadcom holds senior creditor status plus significant leverage over the infrastructure assets. The chipmaker has a floor and an option. Anthropic has neither — it carries the full operating risk, the rigid lease obligations, and a multi-creditor structure that dilutes its strategic autonomy with every new supplier.
The real contrarian claim is subtler: Anthropic is not becoming a customer with a bank. It is becoming a tenant. And a tenant with two landlords — Broadcom on TPU and credit, AMD on GPU and investment — has no pricing power at renewal. Multi-sourcing looks like hedging; here it is multi-sided dependency. Every supplier can convert debt to equity and dilute the tenant's value.
Flip it once more. The entity with the lowest structural risk is the one nobody is celebrating: Google, which owns its TPU, its cloud, and its models. Its circularity is internal — no counterparty credit risk. Anthropic's circularity is external and therefore fragile. Floor prices bleed, but structure remains. The structure says the vertically integrated player is the safe asset, and the financed challenger is the leveraged bet.
The Takeaway
The reported $2 trillion Anthropic valuation implies a price-to-sales near 43x on annualized revenue — a multiple that only clears if compute commitments are covered by a revenue curve that does not yet exist. The IPO is not a finish line. It is a refinancing event: public-market cash used to retire private credit before the maturity wall. If the window closes, the chain does not bend. It snaps.
Pivot, not panic — but pivot early. Track three signals: whether the S-1 appears on EDGAR, the maturity ladder on the $518 billion in commitments, and the cash-recovery quality of Broadcom's AI revenue line. One of them will tell you whether this is a bubble or a repricing. Until then, the honest position is conditional: if the digits are real, the fragility is real; if the digits are fiction, the fragility still describes every vendor-financed compute deal now on the books. Either way, you are not buying a chip story. You are underwriting a bank.