This week I read that OpenAI's 2026 actual revenue sits near $35 billion against a year-end annualized target of $70 billion. Anthropic reports more than $9 billion annualized against roughly $4.6 billion in actual revenue. Two companies, two ratios, both landing near 2x. My first reaction wasn't surprise at the scale. It was recognition โ I have audited this exact pattern before, and it wasn't inside an AI lab. It was on-chain, during a bear market, in the wreckage of protocols that reported "revenue" nobody could reconcile.
The Bloomberg framing was narrow: the two companies use different revenue recognition methods, so their numbers cannot be compared. True. But the deeper story is one every crypto native should know by heart, because we invented most of these tricks first.
Context: gross, net, and the arithmetic of appearance
Revenue recognition has two basic modes. Under gross accounting, you book the entire transaction value even though you keep only a slice. Under net accounting, you book only your slice. Anthropic counts the full sales flowing through Amazon and Google cloud channels. OpenAI counts only the net revenue share it receives from Microsoft. Same physical business, wildly different top line.
Then there is "annualized revenue," a run-rate that takes a short window โ a quarter, sometimes a month โ and multiplies it out to a year. It is not audited. It is not a forecast. It is a projection dressed as a fact.
Open source isn't just code; it's a philosophy of transparency. And here is the uncomfortable part: both companies are building the most consequential technology of the decade while publishing numbers no accountant has signed. Investors are pricing multi-hundred-billion valuations against self-defined metrics. That is not a governance failure unique to AI. It is the default state of our entire industry.
Core: the geometry of a number that passes through you
Think of revenue as a river. Gross accounting measures the flow at the delta โ every drop that reaches the sea. Net accounting measures it at the source โ only the water that belongs to the spring. For a company whose product travels through someone else's channel, the gap between delta and source can be three, four, five times. Anthropic's reliance on AWS Bedrock and Google Vertex means much of its "revenue" is water it never owned. This is not a footnote. It is a structural confession: choosing gross accounting admits that your distribution depends on a landlord who takes a cut you have agreed not to show.
Now layer the annualized metric on top. Annualized revenue is a derivative, not an integral. It measures the slope of the last few months and pretends the slope is the level. The steeper the growth, the more violently it lies. A company climbing from a low base to a $70 billion run-rate is, by construction, the worst candidate for annualization โ its derivative is meaningless precisely where it looks most impressive. When I analyzed Curve Finance's invariant formulas during DeFi Summer in 2020, I learned the same lesson from another angle: a curve's instantaneous slope tells you almost nothing about where it settles.
Crypto has a name for this behavior. We call it TVL. Total value locked is the gross revenue of the chain. It counts the same dollar every time it moves between protocols, every time it is wrapped, every time it is bridged. A stablecoin circulating through ten pools is counted ten times. Nobody audits it. Everybody quotes it. And in a bull market, the number only goes up, which is the only property that matters to the people quoting it.
The double-counting in AI is identical. When Anthropic books the full value of a transaction that runs through AWS, and AWS books its own AI-related revenue from the same transaction, the same dollar appears on two balance sheets. This is not fraud. It is the natural outcome of two parties each choosing the method that flatters them, with no standard forcing reconciliation. I saw the same mirroring in layer-2 rollups and their parent chains, in aggregators and the DEXs they route through โ value counted at both ends, celebrated at both ends, reconciled at neither.
The distortion compounds at the valuation layer. Price-to-sales ratios are calculated by dividing market value by reported revenue. A gross reporter inflates its denominator, which makes its multiple look smaller and its valuation look cheaper than it is. A net reporter keeps a conservative denominator, which makes its multiple look larger and its valuation look richer. Two opposite errors, pulling in two opposite directions, applied to companies everyone insists on ranking against each other. Add the annualized-versus-actual gap on top โ a factor of roughly two in both cases โ and the entire comparison collapses. Investors are not comparing businesses. They are comparing accounting choices and calling the result a market.
And notice who sits at the center. For Anthropic, Amazon is both investor and channel. For OpenAI, Microsoft is both investor and channel. The shareholder is the landlord. That structure exists in crypto wherever a foundation, a market maker, and a listing exchange share incentives: everyone profits when the number looks bigger, and no one profits from the audit.
I spent 2017 auditing the early oracle mechanisms of Augur and Gnosis, and the lesson I carried forward was this: most "technical vulnerabilities" are failures of social design wearing a code costume. The revenue recognition problem is the same species. The accounting method is not a neutral technical choice. It is a governance decision, made by the party with the most to gain from the answer.
Contrarian: crypto has no right to feel superior
Here is the part that stings. Crypto's favorite pastime this cycle is dunking on AI's opaque financials. We hold up on-chain data as the antidote โ everything is transparent, everything is verifiable. But transparency of raw data is not transparency of meaning. A blockchain shows you every transaction and tells you nothing about whether that transaction is yours. On-chain protocol revenue is, by default, gross. DEX volume is counted at both the maker and the taker. Fees are often throughput relabeled. TVL double-counts across every chain a token touches, and the sum of all chains always exceeds the sum of all money.
We confuse data availability with accounting truth. They are not the same thing. The AI labs at least have the excuse of operating off-chain, where reconciliation is expensive. We have the ledger, and we still cannot agree on what a dollar of revenue means. We didn't build an immutable record so that we could rediscover, on our own terms, the same accounting tricks Wall Street perfected decades ago.
The pragmatic test is brutal: does the label matter if the money is real? In a bull market, yes โ because valuation is set on the label, not the cash. Price-to-sales multiples, ARR benchmarks, TVL rankings: these are the instruments that move capital, and every one can be gamed by choosing the flattering definition. Decentralization is not a tech stack; it is an accounting philosophy, and right now we are failing our own exam.
Takeaway: report what you keep, not what passes through
The next cycle will not reward the protocol with the biggest gross number. It will reward the one that can prove what it actually keeps โ a net figure, auditable, reproducible, unglamorous. The AI accounting wars are a preview of a question crypto must answer first: in a world where anyone can publish any number, who verifies the number that matters? Until we build a proof-of-revenue primitive that measures at the source rather than the delta, we are all quoting TVL and calling it truth.