The Exclusions List: Anthropic's Adjusted Profit and the Crypto Protocols That Copy It

Anthropic is profitable. That sentence is true only if you accept a definition, and the definition is the article.

The reported figures: two consecutive quarters of positive adjusted operating profit, and a gross margin above 80%. Both were reported by the Financial Times, sourced to unnamed people familiar with the matter, and disclosed by the company to what the report describes as a small group of shareholders. No revenue line. No cost breakdown. No audit.

An 80% gross margin would place Anthropic's unit economics above most enterprise software companies. SaaS gross margins cluster in the 70–80% band. Anthropic, on the reported number, clears the top of that band while running the most compute-intensive business model in the industry.

That should trigger a reflex. A gross margin is not a measurement. It is a definition. And this one, per the report's own language, is defined by subtraction.

The subtraction is where a serious reader begins. It is also where a specific class of crypto asset has been operating for two years, using the same architecture, with none of the disclosure. That is the subject here.

Context

Anthropic develops the Claude family of models. Its last private valuation was reported near $60 billion. Amazon has invested roughly $4 billion and distributes Claude through Bedrock. Google and Salesforce are also investors. By any conventional reading, the company is one of a small number of frontier model developers with a plausible path to public markets.

The report carries a September 2024 date. Its news value is that it breaks the "all AI companies burn cash" narrative. Its analytical value is lower, because it contains roughly five data points and not one of them is a denominator.

Set that against the current market. This is a bear market. Capital is scarce. "AI" is the last narrative still worth bidding. In crypto, that has produced a token cohort that markets itself as AI infrastructure — decentralized compute markets, inference networks, agent frameworks, data-labeling DePINs. Their disclosed metrics have migrated from total value locked to "revenue," "compute demand," and "protocol earnings."

The migration is the tell. When an industry adopts the vocabulary of a profitable sector without adopting its accounting discipline, it inherits the vocabulary's ambiguity and none of its constraints. It gets the word "revenue" and loses the obligation that a revenue number carries a matched cost.

This is not an argument about whether AI is overhyped. It is an inventory of a disclosure pattern: a favorable numerator, an undefined denominator, and a reader who is asked to price the ratio. s heart.

The Margin Ladder

Start with the number. "80% gross margin" excludes, by the report's own description, two cost items: distribution partner shares — read Amazon — and model training costs.

Gross margin is revenue minus cost of goods sold, divided by revenue. What belongs in COGS is an accounting policy, not a physical constant. But there are limits. If you remove the two largest cost lines from the denominator, you are no longer computing a gross margin. You are computing a marginal contribution rate on the last dollar of inference.

The self-consistent reading is that Anthropic's 80% is a contribution margin wearing a gross margin's name. The distance between those two metrics, across an industry, is roughly the distance between a healthy business and a leveraged one.

Reconstruct the ladder. Each step adds one excluded cost back into the computation.

Step one — reported: 80%. Treat it as the operating rate on marginal inference revenue with all structural costs stripped out.

Step two — add Amazon's distribution cut. Cloud marketplace economics typically run 20–30% of the partner's booked revenue. If Anthropic books revenue net of Amazon, that cost is already inside the number and the 80% is better than it looks. If it books gross and treats the share as a below-the-line item, the margin compresses. Take the midpoint of the range and assume 15–25% of revenue. Implied margin: roughly 60–68%.

Step three — add model training costs. Training a frontier model is a capital event measured in hundreds of millions to billions of dollars per cycle. Amortized over a model's revenue-generating life, this is a real COGS line for a company that must retrain to stay competitive. Apply a conservative amortization. Implied margin: 40–55%.

Step four — add inference compute, to the extent it is not already captured. For a model API business, inference GPU spend is the single largest variable cost. Third-party estimates place it at 30–50% of revenue. At the midpoint, the ladder lands in the 20–40% band.

That band is not a failure. Twenty percent gross margin on a growing revenue base can be a viable business. It is not, however, the business the headline describes. The headline describes software economics. The reality is closer to an infrastructure business with a software narrative.

There is a second exclusion, and it is more consequential for the profit line. "Adjusted operating profit" conventionally excludes stock-based compensation. SBC is non-cash. It is also compensation. For a private company paying below-market cash salaries in equity, SBC is the real cost of the labor that produces the product. Excluding it from "operating profit" is a wide convention and a contested one. The contest is legitimate: a metric that removes the cost of your workforce is not a measure of whether the workforce is sustainable. It is a measure of whether the workforce is currently being paid.

Put the two exclusions together and the reported figures describe a company close to profitable on its most favorable definition and unmeasured on every other.

Now the third exclusion, which is not a cost at all. It is a reader. The figures were disclosed to a small group of shareholders. That is not a public financial statement. It is a selective transmission, and selective transmission has a function. It is what you do before you need a market to believe a story and after you are not yet required to prove it.

The Distribution Channel

Amazon's role deserves its own pass, because it sits on both sides of the ledger. Amazon is an investor and a distributor. The distribution agreement gives Anthropic a second major model channel in the market, alongside Microsoft's channel for OpenAI. It also imposes a toll on every dollar that flows through Bedrock.

A toll is a COGS line. The question the report never answers is whether the toll is booked above or below the reported margin. In a net-revenue presentation, the toll is invisible and the margin looks strong. In a gross-revenue presentation, the toll is visible and the margin looks weaker. Both presentations comply with accounting rules. Only one of them is informative to a reader trying to estimate free cash flow.

This matters because the distribution channel is also a concentration risk. If a material share of Anthropic's revenue arrives through Amazon, then the company's investor and its largest channel partner are the same entity. That is a relationship, not just a contract. When a channel partner is also an equity holder, its incentives on pricing, placement, and disclosure do not always align with the developer's.

I have spent enough time inside partner economics to be unmoved by a favorable margin until I know the reporting basis. Net versus gross is not a rounding decision. On a channel-heavy business it is often a fifteen-to-twenty-five-point swing in the headline number.

The Disclosure Channel

A private company disclosing positive financials to a subset of shareholders is not fraud. It is choreography. The Securities and Exchange Commission regulates what a company may say to the public. It regulates far less about what a company says to its existing investors. The zone between those two rulesets is where pre-IPO signaling happens.

The timing matters. If a company intends to file within twelve months, it begins to seed the belief that its unit economics work. A credible outlet is a useful instrument for that seeding. An anonymous leak is deniable, attributed to no one, and reaches the institutional allocators who price the next round and the eventual listing.

None of that makes the numbers false. It makes them unverified. The report does not claim the figures were audited. It says they were disclosed. Those are different verbs, and the gap between them is where every subsequent mispricing lives. A disclosed number is a claim. An audited number is a claim with a witness. The market frequently prices the two as if they were the same thing, and it is wrong to do so every time.

The Crypto Twin

Here is the pattern, mapped across. Take the three moves — favorable numerator, excluded costs, selective disclosure — and find them in the AI-adjacent token cohort.

Move one, the favorable numerator. Crypto protocols historically marketed TVL. TVL is gameable, everyone knew it, and the market eventually discounted it to near zero. The current cohort markets "revenue." Revenue sounds harder to fake. It is not, because in a token system revenue can be routed.

Consider a decentralized compute network. It pays token incentives to GPU providers. Those providers pay the network a fee to rent capacity. If the network counts the fee as revenue and does not net the incentive, the "revenue" line can be manufactured by the treasury at will. This is Anthropic excluding training cost while reporting an inference margin, with one difference: in crypto, the excluded cost is often the same asset that funds the numerator. The number is paid for with the thing it measures.

Move two, the excluded costs. The crypto version of stock-based compensation is emissions. Token emissions are compensation to validators, providers, and liquidity providers. A protocol reporting gross revenue while excluding emissions is running Anthropic's adjusted-profit logic with a worse disclosure discipline and a more self-referential cost base.

Move three, the selective disclosure. This is where crypto inverts Anthropic and lands somewhere worse. Anthropic disclosed to a small group. Crypto discloses to everyone and reconciles for no one. The public dashboard is the leak, permanently, and it never states what it excluded. It shows a number and calls the number revenue. The exclusion list is not hidden. It simply does not exist.

Audit Note: What a Dashboard Does Not Reconcile

I spent eight months in 2026 auditing the API integration between a leading AI-agent framework and its smart wallet layer. The headline finding was a race condition that allowed agents to bypass multi-signature requirements under specific latency conditions. The finding that matters for this article is the revenue dashboard.

The framework reported transaction volume from agent-executed operations. I reconciled a sample against on-chain settlement across a two-week window. A material share of the counted volume was internal: agent-to-agent calls that moved no external capital, settled against the framework's own escrow contract, and returned to the same wallet cluster that originated them. The dashboard's definition of "volume" was, in effect, any state change our indexer observed.

Nothing in that definition was false. Every transaction existed. Every signature verified. The number was real and the number was meaningless, because the operator had chosen a definition that maximized the numerator and documented no denominator. When I raised it, the response was that the metric was "standard industry practice." It was, in fact, standard industry practice. s heart.

That is the same structural event as Anthropic's 80%. A metric is chosen by the party that benefits from its definition, and the exclusion list is where the report stops. The intent does not need to be deceptive for the effect to be identical. A favorable definition, repeated often enough, becomes the number the market quotes, and the exclusion list never travels with it.

The difference is remediation. Anthropic can be remediated by an S-1: a filing forces revenue, cost lines, and risk factors into one document under penalty of law. The agent framework cannot be remediated, because there is no document and no penalty. There is a dashboard, and the dashboard is marketing.

The Reconciliation Test

I use a three-column test on any financial claim, whether it comes from a frontier lab or a token protocol.

Column one: the disclosed metric, stated as an equation, with every variable named and measurable. Column two: the excluded items, each with a direction and an order of magnitude. Column three: the reconciliation — cash in, cash out, and the legal identity of the entity holding both.

Walk Anthropic through it. Column one fails partially: the equation is named — gross margin — but its variables are withheld. Column two fails partially: two exclusions are disclosed, their magnitudes are not. Column three passes only in the sense that a real corporate entity exists, with real bank accounts and a real auditor who will eventually look. Call that a deferred pass. It is a pass, which is more than most crypto assets can claim.

Walk the AI-token cohort through the same test. Column one fails: there is no equation, only a checklist of adjectives. Column two fails: there is no exclusion list, only a dashboard. Column three fails: the cash often never leaves the protocol's own contracts. A test that a company can half-pass and a token cannot pass at all is a useful test, because it separates two things the market insists on pricing identically.

Which Protocol Is Bleeding

In a bear market, the question readers actually ask is narrower: which assets are safe, and which are structurally fragile. The reconciliation test answers it, because a protocol whose revenue survives the test is a protocol that is not funding itself with its own token.

Apply the directions. A protocol reporting revenue net of emissions is telling you its users pay. A protocol reporting revenue gross of emissions is telling you its treasury pays. The first can survive a decline in its token price. The second cannot, because a lower token price raises the real cost of every unit of revenue it reports. The reported metric deteriorates precisely as the market deteriorates. That is not a business. It is a reflex.

A metric whose cost side is denominated in your own token is a function of your own price, and a function of your price is not a business.

I have watched this exact construction before. During 2022 I retreated into the mechanics of Terra's UST stability mechanism and analyzed the seigniorage flow logic rather than the price chart. The mechanism was internally consistent and externally dependent on a price that could not be defended at scale. I published a geometric argument that the de-peg was a structural certainty under high volatility, not a tail risk. It was downvoted for being too abstract. The abstraction was the point. The failure mode was embedded in the definition of the mechanism, and it was visible before it was testable.

The current AI-token cohort has inherited the pattern without inheriting the warning. Its revenue is a function of its token price. Its cost is a function of its token price. The two cancel in the best case and compound in the worst. That is the Terra geometry, restated in an inference market.

The Training Cost Question, Capitalized

There is an accounting policy inside the Anthropic report worth isolating. Model training cost is excluded from the reported margin. Whether that exclusion is legitimate depends on whether training is an operating expense or a capital investment amortized over the model's useful life. Both treatments exist in practice. The choice changes the reported margin by tens of points and the reported profit by hundreds of millions.

The crypto analog is the treatment of growth incentives. A protocol that pays emissions to acquire users can expense them, or it can call them growth investment. Expense them and the unit economics look poor today. Capitalize them and the unit economics look strong until the incentive stops — at which point the acquired users leave and the capitalized asset writes down. Neither choice is inherently fraudulent. But a reader who cannot see which choice was made cannot price the result, and a reader who cannot price the result will price the narrative instead.

Anthropic tells you the choice. The token protocol does not. That single sentence is the whole comparative argument. It is also, in a bear market, the whole difference between a position you can size and a position you can only hope about.

The IPO Incentive

The final piece is the incentive that produced the disclosure. A company six to twelve months from a public listing has a specific interest: it needs the last private round to price high, and it needs the public narrative to hold through the roadshow. Positive unit economics are the most valuable pre-IPO asset after a functioning product.

That is a legitimate incentive and a directional bias. It predicts what a company will disclose and what it will not. It predicts adjusted operating profit rather than free cash flow. It predicts a gross margin rather than an income statement. It predicts a headline rather than an appendix.

In crypto, the equivalent window is the period between the points program and the token generation event. It is the moment when the protocol has the strongest incentive to report its most favorable metric and the weakest obligation to explain it. Every points program is a soft IPO, run without a prospectus. Every leaked dashboard is a selective disclosure, aimed at a small group of future holders instead of existing shareholders. The mechanism Anthropic used against a small set of investors is the mechanism a protocol uses against everyone who will ever hold its token, simultaneously, with no filing requirement at any point.

s heart.

Contrarian: What the Bulls Got Right

The bulls are not wrong that the direction matters.

A company that reports a positive adjusted figure at all is on a different trajectory than one reporting negative gross margin. The exclusions may be generous, but the underlying motion — revenue growing faster than the cost lines that were included — is real if the report is honest. And the report is plausible. Anthropic has a distribution channel in Amazon, an enterprise sales motion, and a model that customers pay for. That is more than most of the AI-token cohort can claim, and it deserves to be said plainly rather than buried under the accounting critique.

The bulls are also right about stock-based compensation, in one narrow sense. SBC is non-cash. A company can be cash-flow positive while reporting a GAAP loss driven entirely by equity compensation. Excluding it is not always sleight of hand. Sometimes it is a genuine attempt to separate the cash reality from an accounting artifact, and disciplined analysts do the same calculation on their own.

And the strongest bull point: Anthropic disclosed its exclusions. The architecture is legible. A competent reader can rebuild the ladder and decide where the truth sits. Legibility is a form of accountability, even when the number is favorable, and it is more accountability than the token cohort offers.

Where the bull case fails is the assumption underneath it: that the excluded costs are transient. The argument is that training costs fall, inference costs fall, and distribution compresses as the product commoditizes. If that is true, today's 20–40% contribution margin becomes tomorrow's 60% gross margin, and the bulls are early rather than wrong.

But compute costs are structural, not cyclical. The cost curve has not bent at the rate the narrative requires, and every cycle that fails to bend it re-prices the margin lower. Until it bends, the bulls are underwriting a margin that exists only in the definition. They may be right about the destination and wrong about the timeline, which in a bear market is the same as being wrong.

Takeaway

The question is not whether Anthropic's number is real. It is whether anyone will produce the reconciliation. Every market cycle generates a metric optimized for the reader rather than the operator, and the metric always survives until a denominator appears.

Watch for two events. The S-1, which will force Anthropic's exclusions into three audited columns and settle the ladder one way or the other. And the migration of the AI-token cohort from revenue language into gross-margin language, which will signal that the sector has learned the playbook and skipped the discipline.

The audit was a formality, not a guarantee — and here, there was no audit. s heart.