An open-weight model lab has started behaving like a commercial software company: Moonshot AI has told investors it expects $2 billion in annualized revenue by December, roughly twice the run rate it reported for August and about ten times where it sat in the spring.
Key takeaways
- Moonshot's annual recurring revenue climbed from roughly $200 million in April to $300 million in June and past $1 billion in August, with $2 billion set as the year-end target.
- The inflection traces to Kimi K3, a 2.8-trillion-parameter sparse model that activates 16 of its 896 experts per token and lists cached input at a tenth of its $3 base rate.
- Scale is visible outside the company's own numbers: K3 models still push as much as 300 billion tokens a day through OpenRouter alone.
The target surfaced through Bloomberg's reporting on Friday and was picked up across the trade press within hours. What makes it notable is not the size — it would still be a rounding error next to the American labs — but the business model underneath it. Publishing your weights is supposed to be the thing you do instead of building a revenue line.
How an open-weight lab built a revenue curve
Three data points define the shape: a modest spring, a flat early summer, and then a tripling of annual recurring revenue in the eight weeks after Kimi K3 shipped in July, driven simultaneously by API consumption and consumer subscriptions rather than by one enterprise whale. Anyone can announce a model. Very few labs turn a release into a step-change in booked recurring revenue within a single quarter, and that is the claim Moonshot is now making to prospective public-market investors.
The architecture explains part of the pull. K3 is a mixture-of-experts system whose 2.8 trillion nominal parameters never all fire at once — 16 of 896 experts activate per token — so buyers get frontier-scale behavior without paying frontier-scale inference on every request, and a one-million-token context window lets them drop a whole repository or document archive into a single run.
Pricing does the rest of the work. Against $3 per million input tokens and $15 per million output, Moonshot lists cached input at $0.30 — and for the long-running coding agents that reread the same codebase across hundreds of turns, that cached tier is not a discount detail but the number that determines whether a deployment is affordable at all.
Do the benchmarks justify the bet?
On the company's own published tables, K3 edges GPT-5.6 Sol and Claude Fable 5 on ProgramBench by fractions of a point and posts 42.0 on the longer-horizon SWE-Marathon against 40.0 for Claude Opus 4.8 and 14.0 for GPT-5.5, while still trailing the leading closed models on broader reasoning tests. Vendor-supplied scores deserve the usual discount, and no buyer should treat one leaderboard as settled. The direction, though, matches independent measurement all year: Chinese open-weight releases have been closing the agentic gap, as they did when Qwen3.8 Max climbed above every American lab but two on the Artificial Analysis agentic index.
The two problems money does not solve
Margins are the structural one. Freely downloadable weights mean Moonshot monetizes a thinner slice of the value its model creates than a closed vendor does, so even a met target leaves it an order of magnitude behind reported figures near $40 billion for OpenAI and $65 billion for Anthropic. Growth at that gradient still matters, because it establishes a price floor the closed labs have to answer.
Provenance is the messier one. Anthropic alleged this week that Kimi traffic — close to 300,000 requests — was routed straight to Claude Opus while more than 23 million responses were harvested from its models for training, a distillation campaign Moonshot has not publicly rebutted. Separate reporting has questioned the compute story too, describing access to Nvidia Blackwell systems through domestic intermediaries despite export controls. Neither claim slows a revenue chart, but both sit awkwardly in an IPO prospectus.
What to watch next
A confidential Hong Kong filing in early September, reportedly seeking around $3 billion at a valuation above $50 billion after a $20 billion mark in May, is where all of this gets priced. Those multiples only survive contact with reality if the August pace holds without the company discounting its way there. The meaningful signal over the next quarter is therefore mundane: not another benchmark post, but whether the fourth-quarter run rate arrives at list prices.
FAQ
Is Kimi K3 open source?
Its weights are published openly, which is why Moonshot is classed as an open-weight lab rather than a closed-model vendor. That choice lowers its capture rate relative to OpenAI and Anthropic, since customers can self-host instead of routing everything through a paid API.
How much revenue does Moonshot AI actually make?
Investors were told annual recurring revenue crossed $1 billion in August, having been around $300 million in June and $200 million in April. The $2 billion figure is a year-end goal, not money already booked.
What is Anthropic accusing Moonshot of?
Anthropic says Kimi requests were funneled into Claude Opus at scale and that tens of millions of responses were collected to train Moonshot's own models — a practice known as distillation. The company has not answered the specific allegations in public.
