Moonshot AI’s commercial run has a concrete number attached: two billion dollars. The Chinese company aims to end 2026 with annualised revenue of US$2 billion, after annual recurring revenue already passed US$1 billion in August. That was up from US$300 million in June, a jump that shows how the open-weight Kimi K3 model has become a commercial asset rather than just a research project.

Moonshot released Kimi K3 in July. It is a mixture-of-experts model with 2.8 trillion total parameters, 104 billion activated during inference, and a context window of just over one million tokens. The weights can be downloaded and distributed outside Moonshot’s services, but the licence introduces precise commercial thresholds. A company operating a Model-as-a-Service business that generates more than US$20 million in aggregate revenue over any consecutive 12-month period must enter a separate agreement with Moonshot. Very large consumer products face attribution requirements above 100 million monthly active users or US$20 million in monthly revenue. Internal enterprise deployments remain outside these obligations.

Distribution goes beyond Moonshot’s own services. The company has held talks with Microsoft, Amazon and Google to host K3 on their cloud platforms, seeking up to 30% of revenue generated from K3-related services. The discussions, still active in September, also covered data access and mechanisms for auditing token usage. According to Reuters, similar revenue-sharing arrangements have already been reached with smaller cloud providers, including Chinasoft International.

On the independent side, OpenRouter listed 16 providers serving K3 as of mid-September, including DigitalOcean and DeepInfra. List prices were around US$2.40 per million input tokens and US$12 per million output tokens, with variations by provider. Activity on those independently hosted instances does not necessarily represent API revenue paid directly to Moonshot.

The pricing position is aggressive compared with Chinese rivals. DeepSeek V4.1 Flash costs US$0.15 per million uncached input tokens and US$0.60 per million output tokens off-peak; Alibaba Cloud lists Qwen3.5 Plus from its Singapore region at US$0.40 per million input tokens and US$2.40 per million output tokens for requests up to 256,000 input tokens. K3 remains more expensive, but none of the companies publicly disclose inference costs, discounts or margins.

The reckoning with costs

The growth reported to investors says nothing about the cost of generating it. Moonshot has not released figures on gross margin, operating profit, or the cost of the workloads behind Kimi. In July, shortly after K3 launched, the company temporarily stopped accepting new subscriptions because requests were approaching the limits of its existing compute clusters. It gave priority to paying users and announced a future coding-focused tier. Compute capacity, not demand, was the bottleneck.

This is where the open-weight model reveals its commercial nature. The K3 licence is not free software: the weights are open until a project crosses certain thresholds, then a negotiation begins. That shifts value from technical distribution to contractual relationship. Cloud providers are not simply resellers of compute capacity; they are channels that must share revenue with the lab that trained the model. For an enterprise evaluating a self-hosted deployment, the US$20 million threshold may seem distant, but for an API platform it is a very concrete constraint. The result is an incentive to keep internal deployments directly controlled or to negotiate early. The distinction between internal use and commercial services is one of the trade-offs that AI-RADAR helps evaluate in its frameworks on /llm-onpremise, because contractual cost adds to infrastructure TCO.

Meanwhile Moonshot has continued raising capital. Bloomberg reports that the company raised about US$2 billion in May at a valuation above US$20 billion, followed by a US$3.5 billion round in July that valued it at US$35 billion. By September it had raised more than US$5.5 billion since founding and had confidentially filed for a Hong Kong IPO that could seek about US$3 billion, with Goldman Sachs, China International Capital Corp and Deutsche Bank. An ongoing round was valuing the company at about US$50 billion.

Investors are betting on the ability to turn ARR into profit. But without public figures on inference costs, gross margins and operating expenses, annualised revenue remains a partial metric. The Moonshot case becomes a test for the whole open-weight segment: opening weights can fuel adoption and lower barriers, but monetisation runs through commercial clauses that capture value only when the ecosystem reaches enough scale. Anyone downloading K3 for self-hosted use on their own hardware pays Moonshot nothing up to the thresholds, but takes on VRAM, compute power and operating costs for a 2.8 trillion parameter model. The real equilibrium point is not the per-token price, but who controls the margin between serving cost and recurring revenue.