The charts show two companies racing toward the same summit. The ledgers show they are not even running the same race.
In a Bloomberg report now circulating through analyst channels, two figures sit side by side that cannot be reconciled. OpenAI's actual 2026 revenue lands near $35 billion against an annualized target of $70 billion. Anthropic's 2025 actual revenue lands near $4.6 billion against a publicly circulated annualized figure above $9 billion. The distance between these companies is not merely scale. It is accounting. One books gross; the other books net; and a market hungry for a scoreboard has accepted both numbers as though they shared a single unit.
I have spent enough years auditing cryptographic systems to recognize the shape of this problem. It is not new. It is simply wearing a new logo, and speaking a new dialect of the same old language.
For readers who do not live inside financial statements, the distinction matters more than it sounds. Gross revenue recognition records every dollar that passes through a business, including the portions owed to partners and platforms. Net recognition records only the fraction the company keeps. Anthropic, which distributes heavily through Amazon Bedrock and Google Vertex, counts the full value of those cloud-channel sales. OpenAI, which leans on direct sales and consumer subscriptions, counts only its net revenue share. The difference is structural, not cosmetic: a single dollar of cloud-mediated revenue can inflate Anthropic's top line several times over while leaving its underlying economics untouched.
Both companies then convert this into "annualized revenue" โ a run-rate that extrapolates a recent period across twelve months. This metric is unaudited, voluntary, and in both cases lands at roughly twice the actual figure. Neither company publishes standardized financial statements. So the industry now holds two headline numbers that are directionally useful, cross-company incomparable, and systematically generous. That is not an accounting footnote. That is the entire valuation conversation.
Here is the mechanism, traced carefully. When annualized revenue overstates actual revenue by a factor of two, any price-to-sales multiple built on it appears halved โ the company looks cheaper than it is. When the gross method further inflates the denominator, Anthropic's multiple is suppressed a second time. When the net method keeps OpenAI's denominator honest, its multiple is comparatively inflated. Two distortions, pulling in opposite directions, stack on top of a shared inflation. The result is a comparison that cannot be made, dressed up as one that has been made.
This is not a financial story with a crypto cameo. It is a crypto story wearing a suit.
Tracing the silent currents beneath the market, I keep returning to 2020. I was analyzing the curve.fi stablecoin pools when I calculated that leveraged algorithmic stablecoins carried a fragility index of 0.85 โ a number that implied imminent collapse. The market ignored it, because the yields were 300% and the yields were the argument. When Terra fell in 2022, the model was vindicated and the lesson was bitter: self-reported metrics are not measurements. They are narratives with decimals.
Total value locked taught us this first. A protocol could count the same dollar across three chains and call it growth. A yield farm could print APY against emissions and call it revenue. Nobody audited the reserve, because the reserve was inconvenient. Liquidity is a mirage; reality is in the reserve. The AI industry has now built its own version of this reflex, and it has done so with the full participation of the most sophisticated capital in the world.
What sharpens the current moment is the identity of the counterparties. Amazon is simultaneously Anthropic's largest investor and its largest distribution channel. Microsoft occupies the same dual role with OpenAI. When the shareholder is also the channel, the gross method does not merely inflate a number โ it inflates the number in a direction both parties find agreeable. The cloud provider earns a larger AI-revenue story; the model developer earns a larger valuation story. The same transaction is counted twice, and each count serves its author.
Based on my audit experience with recursive proof verification, I have learned to distrust any figure whose provenance cannot be inspected. A zero-knowledge proof is valuable precisely because it can be verified without trust. A revenue figure is the opposite: it asks for trust without verification. When a metric is unaudited and self-published, the correct posture is not skepticism alone โ it is the assumption that the metric has been optimized by its author, consciously or otherwise.
The consensus framing is that AI is absorbing crypto โ that the capital, the talent, and the narrative have all migrated. I think that reading is backwards in the dimension that matters most.
What AI has absorbed is crypto's measurement culture, not its technology. Annualized revenue is total value locked with a consulting vocabulary. The gross-versus-net discretion is the same discretion that once let DeFi protocols choose which pools to count and which to hide. And here is the twist: post-FTX crypto is moving in the opposite direction. The survivors are adopting net-basis reporting, proof-of-reserves, audited attestations. Crypto spent a decade learning that unaudited self-reporting destroys trust, and it is finally internalizing that lesson โ at the exact moment the AI industry is discovering how profitable the old habit can be.
When I advised a sovereign wealth fund in Riyadh on a Bitcoin allocation last year, I watched board members interrogate a volatility model line by line. They would never have accepted a run-rate number without a reconciliation. The two industries are now crossing in opposite directions on the same bridge. Patterns emerge when we stop watching the price and start watching the ledger.
The tell will not come from a press release. It will come from an S-1. When either company files to go public, auditors will be forced to reconcile the run-rate with the recognized figure, and the adjustment will be public, permanent, and instructive. The audit reveals what the algorithm omits.
Until then, treat every cross-company AI revenue comparison as a hypothesis, not a fact. Ask for net-basis numbers. Ask for gross margin. Ask who the channel is โ and whether that channel sits on the cap table. The water is rising. The real question is which foundation is load-bearing, and which one was only ever a number someone chose to publish.