The short version: In an April 13, 2026 internal memo, OpenAI's chief revenue officer told employees that Anthropic's roughly $30B run-rate was overstated by about $8B — that measured comparably, it was closer to ~$22B, behind OpenAI's ~$25B at the time (Implicator.ai). Nobody is accusing anybody of inventing customers. The entire ~$8B gap comes down to one accounting choice: gross vs net revenue recognition. It's the same choice sitting quietly inside a lot of startup ARR — quite possibly yours.
Where the gap comes from#
When a customer buys Claude through AWS, Azure, or Google Cloud, money flows through the marketplace: the customer pays the cloud, the cloud takes a commission, and the rest goes to Anthropic. The accounting question is how much of that transaction Anthropic gets to call its own revenue.
- Gross (principal) treatment: record the full amount the customer pays, and book the cloud's commission as an expense below the revenue line.
- Net (agent) treatment: record only the slice Anthropic keeps — the marketplace cut never appears in revenue at all.
Same cash. Same profit. A materially different top line. Anthropic books gross; OpenAI reports its Microsoft-channel revenue net of Redmond's cut — so when you line up "$30B" against OpenAI's number, you're comparing a gross figure to a net one. That's the apples-to-oranges the memo was really flagging. Both methods, importantly, are compliant with US GAAP.
Gross revenue answers "how much economic activity flows through this company?" Net answers "how much does it actually keep?" Neither is a lie. The lie is comparing one company's gross to another's net — and then ranking them.
The rule that decides it: ASC 606, principal vs agent#
Under ASC 606 (and its international twin, IFRS 15), the gross-or-net call turns on a single word: control. Do you control the good or service before it's transferred to the customer? If yes, you're the principal and you recognize gross. If you're merely arranging for someone else to provide it, you're the agent and you recognize net.
The standard lists indicators of a principal:
- you're primarily responsible for fulfilling the promise to the customer;
- you carry the risk if the service fails or the customer doesn't pay;
- you have discretion in setting the price.
Run a frontier lab through that test and gross looks defensible: Anthropic runs the inference, owns the model quality, and sets the price — the cloud is closer to a reseller taking a distribution fee. Run a thin API wrapper through the same test and it starts to look like an agent: if you just pass tokens through at a markup and the customer knows they're getting Claude, you may be arranging someone else's service for a fee.
Why this is your problem, not just theirs#
Here's the part that matters for a company with four employees instead of four thousand. The exact same lever is inside your ARR.
If you resell a foundation-model API — you bill customers for usage and pay Anthropic or OpenAI out of it — booking the full customer payment as revenue can make your headline ARR look 2–3x larger than the margin you actually retain. It's not illegal. But it's also not durable: a diligence-stage investor or an acquirer will restate you to net in about ten minutes, and the "growth" that was really just rising cost of goods sold disappears from the model. You don't want to discover that your $3M ARR is a $1M ARR at the table, after you've anchored a valuation on the bigger number.
The move is boring and it works:
- Report the number you keep. Lead with net revenue — your margin after model costs — and show gross separately as "platform volume" if it's genuinely impressive. Sophisticated readers trust the founder who volunteers the net.
- Know your ASC 606 story before you raise. If you're clearly the principal (you control the workflow, own the outcome, set the price, eat the risk), gross is fair — say why. If you're closer to an agent, don't inflate; you'll only get repriced.
- When you compare vendors or competitors, match the basis. Gross-to-gross or net-to-net. A scoreboard that mixes the two — the thing that just triggered a public spat between the two biggest labs — tells you nothing.
The Anthropic–OpenAI fight is a trillion-dollar version of a question every reseller-shaped startup answers whether it means to or not. The labs will settle theirs in an S-1 footnote reviewed by an army of accountants. You settle yours the first time an investor opens your revenue model — so decide which number is really yours before someone else decides for you.
This is the accounting question underneath Anthropic's IPO filing. For the strategic side — what a publicly-traded model vendor means for the startup built on it — start there.
