That impressive revenue number Anthropic’s fans keep waving around? It probably doesn’t mean what you think it means. According to Marcus on AI, the three letters everyone’s celebrating, ARR, hide a sleight of hand that changes the whole story.
Here’s the trick. ARR can stand for two very different things, and the gap between them is where the hype lives.
🔤 Same acronym, two different animals
Marcus on AI lays out the split clearly:
- Annual Recurring Revenue. This is the good kind. Subscriptions, contracts, money that shows up again next year because customers keep paying. It recurs. That’s the whole point.
- Annualized Run Rate. This one only sounds solid. It can mean you took your best month, multiplied it by 12, and called that your year. No promise it repeats. You might never actually book a full year at twelve times your peak month.
When boosters cheer a company’s ARR, they want you picturing the first one. As Marcus on AI puts it, they “rarely tell you” which they mean, “but they really mean the latter.” That’s the sly part. The phrasing borrows the credibility of recurring revenue while describing a projection built on a single strong month.
💸 Why this matters for Anthropic right now
The distinction isn’t academic. Marcus on AI argues the real risk for Anthropic is that its second-quarter revenue may not repeat at the same scale next year. The reason: what he calls the collapse of “tokenmaxxing.” Big customers, ATT among them, are shifting toward cheaper open-source models from other vendors to cut costs.
Think about what that does to a run-rate number. If your peak came from enterprise spending that’s now hunting for cheaper alternatives, annualizing that peak isn’t a forecast. It’s a hope. Recurring revenue assumes next year looks like this year. A run rate assumes nothing of the kind.
This is significant because the entire AI funding narrative leans on revenue trajectories. Valuations, raises, and partnership announcements all cite these figures. If the market is pricing companies off run-rate math dressed up as recurring revenue, it’s pricing off a number that can evaporate the moment a big client switches models.
📉 The Netscape echo
Marcus on AI reaches back to 1995 for a warning. Netscape had revenue, and lots of it, doubling every quarter. The company went public and hit billions in value without ever turning a profit. Then Microsoft shipped a competing browser for free, and Netscape faded, all but gone within a decade.
The parallel he’s drawing: fast revenue growth and no profit can look unstoppable right until a competitor undercuts you on price. For AI labs, the “free competitor” is open source. That’s not a hypothetical. It’s the exact cost pressure already pulling enterprise customers away.
✅ What to actually do with this
For practitioners, investors, and operators reading the next round of AI revenue headlines, a few practical moves:
- Ask which ARR. When a number gets quoted, find out if it’s recurring revenue or an annualized run rate. If nobody can tell you, treat it as the weaker one.
- Look for retention, not peaks. Recurring revenue means renewals. A run rate built on one big month tells you nothing about whether customers stay.
- Watch the switching costs. The open-source shift Marcus on AI flags is the variable that turns a strong run rate into a shrinking one. Track which customers can leave cheaply.
- Discount the projection, not the product. Anthropic’s models can be excellent and its run-rate framing can still be misleading. Both things hold.
Marcus on AI closes with two words worth keeping handy the next time a founder or a fund boasts about ARR: caveat emptor. Buyer beware. The number might be real. The meaning is the part they’re not saying out loud. You can read the full breakdown at the original source.