Moonshot AI expects to reach $2 billion in annualized revenue by the end of the year, Bloomberg reported Friday — double the run rate the Chinese lab stated for August. Moonshot publishes its model weights for anyone to download and serve, which makes the target a live test of a claim the open-weights camp has been making for years: that you can give the asset away and still sell the service at scale.
Source: techcrunch.com
The growth is attributed to K3, released over the summer. Usage has slipped somewhat over the past few months, but OpenRouter data currently shows K3 models generating as much as 300 billion tokens a day on that system.
Put the target beside the numbers it is chasing and it shrinks. OpenAI is reported at $40 billion in revenue, Anthropic at $65 billion. Moonshot's year-end goal is five percent of the first and about three percent of the second. That gap is wider than it looks, because open weights also mean thinner margins: Moonshot cannot price like a lab whose model exists only behind its own API, and it competes on serving a model that its own customers are free to host themselves. The rising forecasts do show that open models can make money. They also show what that money costs.
The 300 billion daily tokens are worth reading carefully. They measure how far K3 has spread, not how much of it Moonshot is paid for. Freely downloadable weights are, by construction, servable by anyone. A token count on a routing platform is a reach metric that an open-weights lab has every incentive to quote and no particular reason to break down.
Earlier this week Anthropic accused Moonshot of a sustained distillation campaign. By Anthropic's account, Moonshot routed almost 300,000 queries from Kimi directly to Claude Opus, effectively substituted Opus output for its own models' answers, and collected more than 23 million responses from Anthropic models to train its own systems. Moonshot's development methods, on this telling, are contested and possibly illegal.
The two stories arrived within days of each other, and they describe the same business from opposite ends. One says Moonshot has built something people use enough to double its revenue in a quarter. The other says a meaningful part of what they are using was assembled out of a competitor's outputs. Both can be true, and if they are, the number that matters is not $2 billion but the ratio between what the training cost and what it would have cost to do independently. That ratio is the actual product of distillation, and neither Bloomberg's figure nor Anthropic's complaint gives it.
My read is that the revenue target is less a statement about Moonshot than about pricing pressure on everyone else. A lab charging for inference on weights the market can download is a lab agreeing to compete on cost of serving. If Moonshot can get to $2 billion doing that, the floor under closed-weights pricing moves down — not because Moonshot wins, but because it establishes that the same capability is available at a margin nobody with $65 billion in revenue wants to defend.
Notably absent from the reporting is any breakdown of where the $2 billion is supposed to come from: API inference, consumer subscriptions to Kimi, enterprise deployments, domestic contracts. A doubling in four months is a different claim depending on which of those is doing the work, and an open-weights lab quoting an annualized figure is quoting the number least constrained by how long it has actually held.
Anthropic is simultaneously the highest-revenue lab in the comparison and the one alleging its outputs were harvested to build a cheaper rival. If both are accurate, the cost of being the frontier now includes subsidizing the people who undercut you, and there is no legal theory currently tested that prices that in.