Reuters reported this week that Moonshot AI, the Beijing lab behind the Kimi models, is in early talks with Microsoft, Amazon and Google about revenue-sharing arrangements that would put its Kimi K3 model on Azure, AWS and Google Cloud. According to the report, Moonshot is asking for as much as 30% of the revenue those services generate.
Nothing is signed. The talks are described as early-stage with no guarantee of a deal, and the sticking points are the ones you would expect: how the split actually works, what data access each side gets, and — the detail worth pausing on — how token usage gets audited. Three of the largest companies on earth and a Chinese startup cannot agree on who counts the meter.
If any of the three signs, it would be the first significant revenue-sharing pact between a Chinese frontier lab and a major US cloud provider.
Our take: Two years of US policy has been built on controlling where the chips go. This is weights walking in the front door. Export controls govern silicon crossing a border; they say close to nothing about a model trained in Beijing being offered as a first-class managed service inside American cloud regions, billed in dollars, to American enterprises. That is not a loophole anyone hid — it is a category the rules never covered. And the hyperscalers have no structural reason to say no. They sell compute and margin. They are indifferent to whose weights are burning it.
Why 30% is the real number
A 30% revenue share is app-store pricing, and Moonshot is on the wrong side of it — it is the one asking to receive, not to pay. Read it as a statement about what the lab thinks is scarce. Not training capability: Moonshot has that, and the third-party numbers back it up. Artificial Analysis has assessed K3 as broadly comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8, and Arena.ai has placed it first on a benchmark for building web interfaces. What Moonshot does not have is distribution to Western enterprise buyers, and it has evidently concluded that is worth giving up a large majority of the revenue to rent.
That is the same conclusion Alibaba's Qwen team reached from a different direction, and Alibaba is among Moonshot's backers. Chinese labs have stopped trying to win on model quality alone. They are competing on landed cost inside somebody else's sales channel.
What the buyer sees
Strip out the geopolitics and a procurement team sees a model at rough parity with the US frontier, available through the same console, the same billing relationship and the same compliance paperwork as everything else they already buy. Provenance stops being a technical question and becomes a policy one — and policy questions get answered by whoever is cheapest when budgets are tight.
Which is exactly why the auditing dispute matters more than it sounds. If nobody can agree how to count tokens, nobody can agree how to count what a Chinese model is being used for, by whom, on what data. That is the clause a regulator would read first.
What to watch
- Whether one signs alone. Three simultaneous negotiations means the first mover sets the terms and the other two match or refuse. A single signature converts this from a report into a precedent.
- Washington's response time. The chip-control apparatus took years to build and is still being enforced through indictments. There is no equivalent framework for inbound model weights.
- Moonshot's IPO. The company is reported to be IPO-bound. A signed US cloud distribution deal is a very different prospectus from a domestic-only one.
- Enterprise appetite. Watch whether any named Western customer is willing to be the reference account. Availability is not adoption.
For anyone actually building on these models, the practical takeaway is unchanged and slightly more urgent: the provider list is getting longer and more political at the same time. Our stack resilience playbook covers how to keep a workflow portable enough that a policy decision made in Washington or Beijing is an inconvenience rather than an outage.