What the Moonshot Talks Actually Are

Reuters reported on August 26 that Moonshot AI, the Beijing startup behind the Kimi family of models, is in early negotiations with Microsoft, Amazon, and Google to host Kimi K3 on Azure, AWS, and Google Cloud. Three people familiar with the private talks say Moonshot is seeking up to 30 percent of the revenue generated from K3-related services on those platforms. If any of these deals close, they would be the first major revenue-sharing arrangements between a Chinese AI developer and a major American cloud provider.

Read that sentence again, because it is doing more work than it looks like. The same Washington that has spent two years blocking advanced AI chips from reaching China is now watching the country's most aggressive model lab negotiate shelf space inside the three clouds that power most of the American enterprise. The talks are preliminary, the sources stress, and there is no certainty they produce agreements. But the fact that they are happening at all tells you where the open-weight model market is heading.

Why 30 Percent Is the Whole Story

The number matters less than what it signals. Moonshot wants up to 30 percent of K3-related cloud revenue, a cut that sources say is in line with the terms the startup has outlined for other large customers using its open-weight model. This is not Moonshot selling tokens through its own API and hoping enterprises show up. This is Moonshot using the hyperscalers as its global sales force, and charging rent for the privilege.

Everything about the deal structure is still unresolved, and the open items are the interesting ones. Revenue split mechanics, data access, and auditing token usage are all still being worked out. Token auditing is the one that should make anyone in this industry stop and think. Usage-based AI billing runs on token counts, and in a revenue-share deal the party doing the counting is the party paying the share. The Next Web put it plainly: both sides need to agree on how to count the units of text a model processes, in an arrangement where the meter belongs to the other side of the table. That is a trust problem dressed up as an accounting problem.

Kimi K3, the Model Too Big to Self-Host

You need to understand the model to understand why Moonshot needs the clouds at all. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters per token, drawing from 896 experts of which 16 fire per token. It runs 93 layers, reads a 1,048,576-token context window, and ships in native MXFP4 quantization with MXFP8 activations. The full open-weights release on July 27 weighed in at roughly 1.56 terabytes across more than 90 safetensors shards.

The headline numbers are legit. Artificial Analysis puts K3 at 57 on its Intelligence Index, the best score of any open-weight model and fourth overall behind three closed frontier systems. Arena.ai ranks it first on web interface building. Independent testers describe performance comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on complex, multi-step tasks. Moonshot's API pricing starts at 3 dollars per million input tokens and 15 dollars per million output, with 30 cents on cached input, which puts it in the same headline bracket as Claude Sonnet 5. For context on the open-weight race: DeepSeek's V4-Pro sits at 1.6 trillion parameters. K3 is nearly double that.

Here is the catch that defines this whole story. Open weights mean you can download the model, fine-tune it, and modify it. But almost nobody can actually run it. Serving a 2.8-trillion-parameter checkpoint at full context takes a multi-node GPU cluster that all but a handful of organizations on Earth can assemble. Open weights do not remove the need for a cloud provider, as The Next Web put it. They change who the cloud provider is competing with.

The License Nobody Read: Open Weights, Conditional Terms

Kimi K3 shipped under a bespoke document called the Kimi K3 License, tagged on Hugging Face as license:other. It is not MIT. It is not Apache 2.0. It grants the usual rights to use, copy, modify, and distribute, then attaches two conditions that a permissive license would never carry.

The first is a revenue gate. If you run a model-as-a-service business and your group revenue passes 20 million dollars over any consecutive 12 months, you must sign a separate agreement with Moonshot before using K3 commercially. The second is an attribution rule: any product built on K3 with more than 100 million monthly active users, or more than 20 million dollars in monthly revenue, must display the name Kimi K3 prominently in its interface. Purely internal use is exempt. For a startup or a research lab, the practical effect is zero. For a hyperscaler reselling inference at scale, it is a negotiation. That distinction is exactly why this cloud deal talk matters: the license was written with the big distribution partners in mind.

The Bessent Problem: Blacklist Threats and Distillation Accusations

None of this is happening in a political vacuum. US Treasury Secretary Scott Bessent said last month that he might add Moonshot to a trade blacklist. US officials have accused the company of distilling technology from Anthropic's most sophisticated model, Fable, to build Kimi K3, and of illegally acquiring Nvidia chips. Moonshot has rejected the distillation claims, telling China's National Business Daily that its performance gains came from original changes to the underlying architecture, not from copying a competitor.

So here is the tension sitting in the middle of the story. Three American hyperscalers are reportedly negotiating with a company the Treasury Secretary has publicly floated blacklisting, over allegations the same government is investigating. The clouds get a frontier-class model that tests well and undercuts Western pricing. Moonshot gets distribution it could never build itself, plus a revenue story to take into its planned Hong Kong listing, reportedly at a 30-billion-dollar valuation after a sevenfold jump in six months. Whether any of this concludes may depend less on the terms than on whether Washington decides it is permissible at all.

The Questions Nobody Has Answered

1. Who audits the tokens? The party counting usage is the party paying the share, and no independent mechanism has been described.

2. What does data access mean? A Chinese company hosting a model on Azure or AWS raises immediate questions about what it can see of the prompts and outputs flowing through it. Enterprise customers will want an answer before they route anything sensitive through K3.

3. Can a deal survive a blacklist? If Bessent follows through, every one of these contracts becomes a compliance problem the day after it is signed.

4. What does this do to the price war? Chinese models are already far cheaper than Western offerings. Putting them on US clouds at a 30 percent revenue split could compress inference pricing further, which nobody in Mountain View or Seattle is going to enjoy.

5. Where is Alibaba in all of this? Reuters notes Alibaba is also seeking revenue-sharing agreements with major users of its own open-source model. Moonshot is not the only Chinese lab that figured out this playbook.

What This Means: The Operating Model Shift

The old model was simple: a lab trains a frontier model, runs its own API, and sells tokens directly. The new model, and K3 is the clearest example yet, is different. Open weights are the marketing, the cloud revenue share is the business, and the license is the lawyer. The hyperscalers become the distribution layer, the startup gets reach it cannot buy, and enterprises get frontier-class capability through clouds they already use and already trust, assuming the data-access questions get answered.

For anyone running infrastructure, the practical takeaway is blunt. A 1.56-terabyte checkpoint is not something you casually self-host, and the license terms mean you should read the fine print before you build a resale business on top of it. The glory days of downloading an open model and shipping it as your own product are getting more complicated, on purpose.

What Comes Next

Watch three things in the coming weeks. Whether Bessent's blacklist threat turns into an actual designation. Whether any of the three clouds announces a firm agreement instead of a preliminary discussion. And whether Moonshot's Hong Kong listing prospectus starts counting hypothetical US cloud revenue as part of its growth story. For the rest of us, this is what the second act of the AI decade looks like: the models are becoming commodities, and the real fight is over distribution, trust, and who gets to audit the meter.

— Allan Ali, Sylt.ing