July 2026. A 261-page S-1 lands on my desk. Eighty pages of it are risk factors. Forty-eight pages describe the actual business. The ratio tells you everything before you read a single number.
The document, reportedly filed confidentially by Anthropic in June and surfaced by Reuters in September, contains a figure that should stop any auditor cold: $518 billion in infrastructure commitments. Against 2025 revenue of $4.59 billion. That is a 1:113 ratio. For every dollar the company earned last year, it has contractually obligated itself to spend 113.
I have reviewed reserve proofs for exchanges that hid $400 million in misappropriated funds. I have traced reentrancy vulnerabilities through assembly opcodes at 3 AM. But this structure — if the reported numbers hold — represents something structurally different. Not fraud. Not incompetence. A deliberate, board-level decision to lock a company into a trajectory so rigid that failure becomes terminal rather than merely costly.
The chain remembers what the ledger forgets. Here is what the ledger shows.
Context: The Shift Nobody Priced In
For three years, the narrative around frontier AI companies was software economics: high margins, low capital intensity, infinite scalability. That narrative is now dead.
Anthropic's disclosed commitments tell a different story. The $518 billion breaks down across four counterparties: Broadcom at $161.2 billion (31.1%), Google at $111.1 billion, Amazon at $110 billion, and Microsoft at $31.4 billion. Approximately 80% — roughly $414.4 billion — is described as binding and non-cancellable.
Non-cancellable. Even if the models stop improving. Even if demand collapses. Even if a competitor releases something better and cheaper tomorrow. The obligation persists.
I spent three weeks in 2022 cross-referencing on-chain transactions against internal SQL databases for a mid-tier exchange. The pattern I look for is the same in every audit: where does the risk sit, and who holds the offsetting hedge? In Anthropic's structure, the risk sits entirely with Anthropic. The hedge sits with Broadcom, Google, and Amazon — all of whom are simultaneously shareholders, suppliers, and in Google's case, a direct competitor.
That is not a supply chain. That is a hostage arrangement with preferred stock.
The asymmetry is the finding. Everything else is commentary.
Core: Systematic Teardown of the Commitment Structure
Let me walk through this the way I would annotate a vulnerable contract — line by line, tracing each obligation to its failure mode.
Finding 1: The Broadcom Ratio — Broadcom receiving the largest single allocation (31.1%, or $161.2 billion) is the most technically revealing data point in the entire disclosure. Broadcom does not sell general-purpose GPUs. Their AI revenue comes from custom ASIC co-design and AI networking silicon — Tomahawk and Jericho switching chips. When a company allocates more to Broadcom than to Google Cloud and AWS combined, the signal is unambiguous: they are building custom silicon at scale, or they are building a training cluster network so large that the interconnect costs exceed the compute costs.
Either interpretation carries high risk. Custom ASIC development requires 18-24 month design cycles, billion-dollar mask sets, and yields that nobody can guarantee. If Anthropic is mid-flight on a TPU-class accelerator, they have already committed the capital but not yet validated the silicon. The bug was there before the deployment. The tape-out risk is real, and it is entirely theirs.
Finding 2: The Circular Financing Loop — The document reportedly references a $10 billion anchor investment from NVIDIA. NVIDIA sells GPUs. Anthropic buys compute. Some of that compute flows through NVIDIA silicon. NVIDIA then invests in Anthropic. This is not a conspiracy — it is a structural pattern with a name: circular financing. Lucent and Cisco did it during the telecom bubble. Vendors financed customers to buy their equipment, revenue appeared on both sides of the ledger, and when demand softened, the whole chain unwound in a single quarter.
I am not saying Anthropic is a fraud. I am saying the financing structure makes it impossible to determine organic demand from induced demand using the disclosed information alone. That is the definition of an information asymmetry that benefits insiders.
Finding 3: The Non-Cancellable Clause — Here is where it gets structurally interesting. In traditional corporate finance, a company that needs $518 billion in infrastructure would issue debt, match duration to asset life, and hedge counterparty risk. Anthropic instead has issued what amounts to a purchase obligation — a liability that produces no corresponding asset on the balance sheet.
Trust is a variable, not a constant. And in this structure, trust runs entirely one direction. If Anthropic fails to pay, the cloud providers and Broadcom have legal recourse. If the cloud providers fail to deliver, Anthropic's remedies are undisclosed. That asymmetry does not appear as a line item. It appears as a structural fragility that rating agencies will eventually notice.
Finding 4: The Governance Overlay — The S-1 reportedly describes a Long-Term Benefit Trust, supervised by figures including Ben Bernanke, holding governance authority alongside a founder-controlled LLC with 50.1% voting power. The stated purpose: to prioritize safety even when that priority conflicts with shareholder financial interests.
I have audited custody solutions for ETF issuers. The question I always ask is simple: who can move the funds, and under what conditions? Here, the answer is that the founders can override financial interests indefinitely — or until 'only two or fewer founders remain.' The exit path from this structure is undefined. For a public market investor, undefined governance exit paths are a discount factor, not a premium.
Contrarian: What the Bulls Actually Got Right
Here is where I have to be honest about the limits of my own analysis. The bearish reading — 1:113 revenue-to-commitment ratio, circular financing, governance entrenchment — is structurally sound. But it misses three things that the bulls are not wrong about.
First, the revenue growth rate is real and steep. If Anthropic grew from effectively zero in 2022 to $4.59 billion in 2025, that is a trajectory that very few enterprise software companies have ever matched. My 1:113 ratio assumes the denominator stays roughly constant. If revenue grows 5x annually, the ratio compresses to something uncomfortable but not fatal within two years.
Second, the Broadcom bet, if it works, is a genuine strategic moat. Every dollar not spent on NVIDIA margin is a dollar that flows into Anthropic's own IP. Google proved with TPU that this path works. If Anthropic is 18 months behind on a successful custom chip, the market may not care by the time it ships.
Third, the safety positioning has real commercial value in regulated markets. Governments and enterprises that need auditable AI deployment may pay a premium for Anthropic precisely because of the governance structure that public market investors find objectionable. The Pentagon's supply chain risk designation under FASCSSA Section 4713 cuts both ways — it signals that Anthropic's safety controls are strict enough to matter.
The bulls are not wrong that this could work. They are wrong that it can be evaluated using traditional metrics. Optimization is just risk wearing a disguise. And this structure is optimized for a world where scaling laws continue to hold and AI demand compounds at 100% annually. If either assumption breaks, the disguise comes off.
Takeaway: The Question That Matters
The real story in this S-1 is not the $2 trillion valuation that has been circulating in unverified reports. It is the $518 billion commitment. That number is the actual entry ticket for frontier AI. Every company that wants to compete at the top of the model stack must now sign contracts that expose it to existential financial risk before it has the revenue to justify them.
This is the moment AI stopped being a software business and became a heavy industry. Heavy industries have cycles. They have bankruptcies. They have asset write-downs. They have creditors.
Anthropic may be the first AI company to go public with a structure that makes it look more like a leveraged infrastructure play than a technology company. If the IPO proceeds, the market will have to decide whether to price it as software or as steel. The difference in multiples is not small.
The question for every reader holding AI-adjacent assets: when the cycle turns — and it will — who holds the non-cancellable paper? Because the chain remembers what the ledger forgets.
Every exit liquidity event is a forensic scene. This one just hasn't happened yet.