The target
Moonshot AI is targeting $2bn in annualised revenue by the end of the year, roughly double its August run rate, TechCrunch reported on 11 September, citing Bloomberg. The company told investors its annual recurring revenue topped $1bn in August, up from $300m in June.
That is a tripling inside two months. The driver named is Kimi K3, released in July, which reached the top of several performance leaderboards at a price well below the American frontier models. OpenRouter data shows K3 models generating as much as 300 billion tokens a day on that platform.
The comparison that matters
Put next to the labs it benchmarks against, $2bn is small. Recent reporting puts OpenAI at around a $40bn run rate and Anthropic at roughly $65bn annualised. Moonshot is aiming at a twentieth of the former.
The margin picture differs too. Moonshot’s model weights are freely available, which is the mechanism that drove adoption and also the reason it captures less of the value that adoption creates: anyone can run K3 themselves and pay Moonshot nothing. A closed-weights lab bills for every token.
The caveats in the same reporting
TechCrunch notes that K3 usage has declined slightly in recent months, which sits awkwardly beside a target that requires doubling. It also notes that Anthropic has accused the company of a long-running model distillation campaign involving roughly 300,000 requests routed to Claude Opus — an allegation Moonshot has not, in the reporting, addressed in detail.
The listing behind the number
The revenue target is not disconnected from the calendar. Moonshot is raising at a valuation of about $50bn ahead of a Hong Kong listing that could come as soon as this year, 36Kr reported. A run rate doubling in the final quarter is exactly the number a prospectus wants.
What to watch
Whether the December figure is reported as annual recurring revenue or as annualised from a single strong month — the two are frequently conflated and differ by a lot. And whether the token volumes hold: with open weights, a usage decline on a hosted platform can mean the model is losing ground or that more people are running it themselves, and those two readings point in opposite directions.
