Moonshot seeks 30% revenue share as open-weight AI models monetize

Moonshot AI is negotiating with Microsoft, Amazon, and Google to host its 2.8-trillion-parameter Kimi K3 on Azure, AWS, and Google Cloud, seeking up to 30 percent of K3-related service revenue — a first for a Chinese AI firm with US hyperscalers.

Reuters reported Aug. 26, citing three people familiar with the talks, that unresolved issues include how revenue would be split, data access, and auditing token usage. The discussions are at an early stage and there is no certainty they will result in agreements, the sources said.

Kimi K3, released in July, has 2.8 trillion total parameters with only 104 billion active per token — a 1.8 percent activation rate enabled by Moonshot's Stable LatentMoE architecture. The model posts strong third-party results: Artificial Analysis says it delivers performance comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on complex multi-step tasks, and Arena.ai ranks it first in web interface-building benchmarks. Moonshot's official API prices K3 at about $3 per million input tokens and $15 per million output tokens, roughly 40 percent cheaper on input and 50 percent cheaper on output than OpenAI's top-tier models.

The negotiations come as Moonshot prepares for an IPO on the Hong Kong Stock Exchange, with its latest funding round valuing the company at $35 billion and a listing target of up to $50 billion. If the three cloud giants accept revenue sharing, Kimi gains a recurring revenue stream that scales with overseas usage — a critical proof point for investors ahead of the public debut.

The K3 License: Open Weights, Commercial Threshold

Kimi K3 uses Moonshot's custom license. The first half follows a permissive MIT-style approach allowing use, modification, deployment, and commercialization. But it adds a commercial boundary: if a company operating Model-as-a-Service business and its affiliates generate more than $20 million in total revenue over 12 consecutive months, they must renegotiate a commercial agreement with Moonshot before using K3 commercially.

The $20 million figure is a negotiation trigger, not an automatic 30 percent revenue share. The "up to 30 percent" disclosed by Reuters is the specific commercial term Moonshot is currently proposing to large customers, with the final percentage depending on negotiation outcomes.

Moonshot has already tested this model with smaller platforms. Chinasoft International announced July 20 via a Hong Kong Stock Exchange filing that it signed a token revenue-sharing and joint innovation cooperation agreement with Moonshot, splitting token consumption revenue from Kimi models at an agreed ratio. K3 is already available on more than 10 inference providers including Together AI, Fireworks, DigitalOcean, Modal, Baseten, and DeepInfra — several of which are true model hosts that deploy K3 on their own GPU clusters rather than reselling Moonshot's API.

The pricing convergence across these platforms is notable. Together AI and Modal operate in essentially the same price band as Moonshot's official API, while DigitalOcean is slightly lower. For an open-weight model, independent inference companies could theoretically price differently; the convergence suggests Moonshot is involved in technical adaptation and cooperative operations.

Two Diverging Paths for Chinese Open-Weight Models

Moonshot is not alone in redesigning the open-source business model. Alibaba is considering a similar revenue-sharing mechanism for large commercial users of Qwen, with the new Qwen license setting separate commercial authorization thresholds for large MaaS and AI Work Assistant businesses. MiniMax has adopted a similar approach, requiring commercial users reaching certain revenue scales to obtain written reauthorization.

But not all Chinese model companies are moving toward "monetizing at scale." Tencent's latest Hy4 preview and StepFun's Step 3.7 Flash use the standard Apache 2.0 license without additional revenue thresholds. DeepSeek V4 Pro continues to use the MIT License, likewise without requiring third parties to reauthorize or share revenue once their businesses grow. Zhipu's latest flagship GLM-5.3 has switched to a custom license, but its threshold — requiring a safety review only when consecutive 12-month total revenue exceeds $10 billion — functions more as a model governance mechanism than a revenue-sharing arrangement.

The revenue math is compelling on paper. If third-party cloud platforms worldwide sell $1 billion worth of Kimi services annually, Moonshot taking an average 20 percent revenue share would mean $200 million in income, and 30 percent would mean $300 million — without bearing the full cost of global inference infrastructure alone.

But the limitations are equally clear. Open models are replaced at remarkable speed. Today Kimi leads in performance, and cloud providers may accept higher commercial licensing fees; a few months from now, if DeepSeek, Hunyuan, Qwen, or another model achieves comparable performance with lower inference costs and more permissive licenses, Kimi's bargaining power erodes.

The 30 percent figure looks more like the highest quote a model can command during a period of strength than an established long-term "model tax." Whether model companies can build recurring revenue comparable to software licensing or chip IP ultimately depends on how difficult the model is to replace and how costly it is for developers and enterprises to migrate to competing models.

For investors, the stakes extend beyond Moonshot. Microsoft, Amazon, and Google are locked in a zero-sum war for enterprise cloud workloads. If one hyperscaler hosts K3 and enterprise customers demand it, the others face pressure to follow — potentially setting a precedent for revenue-sharing across the AI cloud market. Moonshot's pre-IPO valuation of $35 billion already prices in substantial growth; a successful revenue-sharing framework with the three largest cloud providers would provide a concrete revenue model that extends beyond API sales.

This article is for informational purposes only and does not constitute investment advice.