For most of the past two years, the story of Chinese Artificial Intelligence (AI) models going overseas was a story about adoption. Give the model away as open weights, slash the Application Programming Interface (API) price, upload it to every platform you can find, and hope developers somewhere actually use it.
On September 18, that story entered a different chapter. Amazon Web Services (AWS) announced that Kimi K3, the flagship model from Chinese startup Moonshot AI (月之暗面), is now available as a fully managed model on Amazon Bedrock, Amazon's enterprise platform for building generative AI applications. Enterprise developers anywhere can call it directly and pay by the token.
That sounds like another routine "model joins a cloud catalog" story. It is not. According to multiple Chinese financial outlets, this is the first time a Chinese AI company has entered a revenue-sharing arrangement with one of the world's three largest cloud providers — meaning Moonshot does not merely watch Amazon earn money from its model; it takes a slice.
The deal also acts as a live test of a question the entire open-weight AI industry is now wrestling with: once you give your model away, can you still get paid?
What Actually Went Live
Let us start with the facts that are confirmed.
AWS made the announcement on its machine-learning blog on September 18 (U.S. time). Kimi K3 is now generally available on Bedrock through U.S. and global inference profiles, meaning companies can route workloads either to an American region or a global one depending on their compliance needs. The model supports Amazon's standard Invoke and Converse APIs as well as OpenAI-compatible endpoints, so it slots into code that developers already have.
Kimi K3 itself is a substantial piece of technology:
- 2.8 trillion total parameters, which AWS describes as the first open-weight model at that scale (parameters are the adjustable values inside a neural network; the count is a rough measure of model capacity).
- Native vision support, so it can process images alongside text.
- A 1-million-token context window — roughly three-quarters of a million words it can hold in a single conversation, enough to analyze very large codebases or stacks of documents at once.
- Scaling efficiency about 2.5 times better than the previous Kimi K2, according to Moonshot's own figures.
AWS is positioning it squarely for long-horizon coding, multi-document knowledge work, and AI agent workflows. Two details matter for enterprise buyers. First, on Bedrock, K3 sits inside the same security boundary as closed, proprietary models, with the same unified access controls, encryption, and audit logging. Second, K3 is the first open-weight model on Bedrock to support explicit prompt caching — developers can deliberately cache repeated sections of a prompt, such as a big codebase or a long set of tool instructions, which cuts both latency and input cost on tasks that reuse the same context.
Amazon also published prices. On the global standard tier, K3 costs US$3 per million input tokens and US$15 per million output tokens — identical to Moonshot's own official API price. On the U.S. region, the prices are US$3.30 and US$16.50, about 10 percent higher. So an American company running the model on American infrastructure pays a small premium rather than a discount.
These items — the launch, the specs, the prices — are documented either in Amazon's own materials or consistently across reporting. The financial relationship behind the launch is where things get less transparent, and more interesting.
The "Model Landlord" Model
Neither Amazon nor Moonshot has publicly disclosed the commercial terms. Amazon's announcement does not mention revenue sharing at all, and an Amazon representative pointed reporters back to that same announcement. Moonshot did not respond to requests for comment from several publications including the South China Morning Post.
So "revenue sharing" is not a number anyone has officially confirmed. But there is a paper trail that explains why observers are confident a separate commercial deal exists — and it lives inside Kimi K3's own license.
When Moonshot released K3 in July 2026, it revised the model's open-weight license. The license lets ordinary developers download, deploy, and use K3 commercially for free. It also contains a threshold clause: if a company (including its parent or affiliates) operates a MaaS — Model-as-a-Service — business and earns more than US$20 million in aggregate annual revenue from it, that company must reach a separate commercial agreement with Moonshot before using K3 commercially.
Amazon clears that threshold by an enormous margin. AWS reported in July that its AI business alone had surpassed US$25 billion in annual revenue run rate, and Bedrock is a textbook MaaS operation: Amazon hosts the model, customers call it over an API, and they pay per token. Therefore, for the K3 launch to be legitimate under Moonshot's own license, the two companies must have negotiated and signed separate terms. That is the hard logic underpinning the revenue-sharing reports — even though no one outside the two companies knows the actual percentage.
The figure most often cited publicly came earlier. On August 26, Reuters, citing three people familiar with the matter, reported that Moonshot was simultaneously negotiating with Microsoft, Amazon, and Google, seeking to place K3 on Azure, AWS, and Google Cloud and take up to 30 percent of the revenue from K3-related services on those platforms. Crucially, "up to 30 percent" was described at the time as Moonshot's opening position in early-stage talks — not an agreed rate, and not specific to Amazon. It should be read as an ask, not an outcome.
Chinese media have taken to calling the resulting arrangement "model collecting rent" (模型收租): Moonshot hands the model to a cloud provider, the provider deploys it on its own infrastructure and sells access to its own customers, and Moonshot earns a recurring cut of the traffic without building or operating any overseas data centers itself.
Why This Is Different from Earlier Deals
Kimi K3 was technically callable through a major American cloud before AWS. Microsoft's Foundry platform already offered it, but under a model ID (FW-Kimi-K3) that signals how it was delivered: the underlying inference was provided by Fireworks AI, a third-party inference company. Foundry handled enterprise access and governance; Fireworks ran the model. There was an intermediary between the model maker and the cloud.
AWS is structured differently. According to coverage of the listing, AWS lists Moonshot AI directly as the model provider, and Amazon itself acts as the hosting and inference provider. This is the direct cloud-to-lab partnership that Reuters described, rather than a resale arrangement.
Amazon was already no stranger to Chinese open-weight models. Back in February, it added six at once — DeepSeek V3.2, MiniMax M2.1, GLM 4.7 and GLM 4.7 Flash, Kimi K2.5, and Qwen3 Coder Next — and Bedrock has since hosted dozens of open-weight models from DeepSeek, MiniMax, Qwen, and others. What distinguishes the K3 deal is reportedly the commercial layer, not the technical one.
Nor is AWS the only platform to have worked through the licensing question. K3 is already available on Alibaba Cloud's Bailian (百炼) platform in two forms: one supplied directly by Moonshot, and one deployed and served by Alibaba itself. The Alibaba-self-hosted version sits in the same structural position as the AWS deal — a large MaaS platform commercializing K3 — and would likewise require a separate agreement under the license.
Beyond the hyperscalers, K3 now runs on more than ten publicly listed inference providers — including Together AI, Fireworks, DigitalOcean, Modal, Baseten, and DeepInfra — several of which received the open weights and deploy the model on their own Graphics Processing Unit (GPU) clusters rather than merely reselling an API.
China's AI Labs Are Negotiating as a Group
The broader significance of the AWS deal is that Moonshot is not alone. The Chinese open-weight industry appears to be moving toward commercial licensing as a coordinated strategy.
On September 16, Zhipu AI (智谱) disclosed on an analyst and investor call that it had signed revenue-sharing agreements with several major cloud providers, both domestic and overseas. Its GLM series will be offered as hosted APIs on those platforms, with revenue split at an agreed ratio, and the income will begin to be recognized in October. Zhipu raised its year-end Annual Recurring Revenue (ARR) guidance from US$2.4 billion to US$3 billion on the strength of this and other business lines.
According to Chinese reporting, Alibaba is also considering revenue-sharing terms for large commercial users of its Qwen models, along with additional commercial licensing conditions for large MaaS and AI-office businesses. MiniMax similarly requires re-authorization for commercial users above a certain revenue scale.
It is important to note that not every Chinese lab is taking this path. Tencent's Hy4 preview and StepFun's Step 3.7 Flash use the standard, permissive Apache 2.0 license, and DeepSeek V4 Pro continues to use the MIT License — both essentially free for large MaaS use with no revenue threshold. So "open weights plus a payment gate" is a strategic choice, not an industry consensus.
The Logic — and the Weak Point
Why would a cloud giant accept a payment gate on a model it could, in theory, download and run for free? The answer the industry keeps arriving at is simple: a license clause creates leverage only if the model itself has demand.
The clause in K3's license does not physically stop anyone from deploying the model. What it does is preserve Moonshot's right to negotiate with any company large enough to be worth suing or losing as a partner. A cloud provider will accept that gate for one practical reason — if not offering K3 would push its users and their token spending toward a rival platform that does. Amazon is in an intensifying competition with Microsoft and Google for enterprise AI workloads; a strong, popular model in the catalog is worth a revenue share if customers genuinely ask for it.
For Moonshot and its peers, the appeal is structural. They turn the world's cloud companies into a global sales and distribution network, gain access to each platform's compliance infrastructure and enterprise relationships, and avoid the enormous capital cost of building inference data centers and sales teams in every market. It is, in effect, a lighter, faster route to overseas revenue than selling API access directly or negotiating custom enterprise projects.
But the same reporting points repeatedly to the weak point. The model only works for as long as the model is irreplaceable. Open-weight models now iterate extremely quickly. If a competitor offers similar performance at lower cost under a genuinely free license — a DeepSeek on MIT terms, a Qwen on Apache — cloud providers and developers can switch, and the "rent" stream evaporates. A threshold license can reserve the right to negotiate; it cannot, by itself, manufacture the bargaining power to win.
There are also open practical questions. No one outside the deals knows how token usage is audited, what counts as attributable K3 revenue inside a giant cloud's bundled business, or how durable the terms are. The US$20 million clause and the 30 percent ask are, for now, evidence that a negotiation happened — not evidence of how much money will move.
Why It Matters Beyond One Company
Strip the story back and the shift is clear. A year ago, Chinese AI labs were handing models to the world to win users. Now they are figuring out how to turn that installed base into a recurring income stream, and they are using American and global cloud platforms — not just their own infrastructure — as the collection point.
For overseas developers, the practical effect is mild: one more strong, long-context model available inside a platform many already use, at a price that matches the maker's own API. For the industry, the effect is larger. Kimi K3 on Bedrock is a proof point that open weights and commercial revenue are not opposites — a model can be freely downloadable for most users while still carrying a payment gate for the handful of giant platforms that profit most from hosting it.
Whether that gate holds depends entirely on whether Kimi K3, and the models that follow it, stay good enough that the world's largest cloud companies decide paying is cheaper than walking away. The first answer is in. Amazon said yes.