Moonshot Wants 30% of What the Clouds Earn From Its Model

BEIJING, 27 AUG 2026 — Moonshot AI is reportedly in early talks with Microsoft, Amazon and Google about hosting its Kimi K3 model on their clouds, and is asking for as much as 30 per cent of what those platforms earn from services tied to it.

Any agreement would be the first substantial revenue-sharing arrangement between a Chinese AI developer and a major American cloud. The talks are preliminary and may produce nothing.

The usual arrangement, reversed

Normally a cloud platform hosts a model and takes a cut of what the model earns. The platform provides distribution, billing, compliance and the enterprise relationship, and it is paid for supplying them.

Moonshot is proposing the reverse. It wants a share of what the cloud earns, claiming that the model, not the infrastructure, is the scarce input.

The claim has weight. A hyperscaler can always build more capacity, but it cannot easily conjure a frontier model that customers ask for by name. But asking for 30 per cent of a platform's revenue on services tied to your product is an unusually confident opening position, and the reporting is explicit that splits remain under negotiation.

Token auditing is where this gets difficult

The detail most likely to decide the outcome is a practical one. Token-usage auditing is still under negotiation.

A revenue share is only as good as the measurement behind it. If Moonshot is owed a proportion of what the clouds earn from K3, Moonshot needs to verify how much the model was used, which means some visibility into usage on infrastructure it does not control.

Hyperscalers do not readily grant that. Usage data is commercially sensitive, it maps to identifiable customers, and enterprises buy cloud services partly on the promise that their consumption is nobody else's business.

Now add the political dimension. The party seeking that visibility is a Chinese company, and the customers are American and European enterprises. Even a technically clean arrangement — aggregated counts, audited by a third party, no customer identifiers — would need to survive a procurement conversation in which somebody asks what a Beijing-based firm learns about usage patterns. That conversation is harder than the commercial one.

The percentage matters less than what it applies to

Thirty per cent is the number in the headlines, and in a negotiation like this the base is worth more than the rate.

Services tied to K3 could mean several things. At the narrowest, it is revenue from inference calls to the model and nothing else. At the widest, it includes the storage holding the customer's data, the networking moving it, the vector database supporting retrieval and the managed services wrapped around the deployment — the pull-through that makes hosting a model attractive to a cloud in the first place.

Thirty per cent of the narrow base is a modest sum. Ten per cent of the wide one is very much larger. The headline figure describes only the rate, which is the half of the term that was easiest to report.

This is also why an early-stage negotiation can run for months without either side moving on the percentage. The definition is the deal.

What actually restricts this

The regulatory position is more permissive than many assume.

American export controls on artificial intelligence have concentrated on hardware — which chips may be sold, to whom, and through which countries. Model weights, particularly openly published ones, have not been restricted in the same way, and a hyperscaler hosting an openly available Chinese model is not obviously prohibited from doing so.

The real constraints are likely to be commercial and procedural rather than statutory. Government and defence customers buy cloud services under accreditation regimes that specify what may run where, and a cloud provider weighing this deal is weighing the reaction of those customers as much as any legal text. Political attention is also its own constraint: a policy that does not exist today can exist in six months, and a contract signed now has to survive that.

The question facing Microsoft, Amazon and Google is therefore less about permission and more about what doing it costs them elsewhere.

Why the clouds might want this anyway

Why would any hyperscaler entertain terms like these?

Enterprise customers increasingly choose infrastructure by what runs on it. A cloud that cannot offer a model a customer wants loses not only that workload but the storage, networking and data services attached to it. Model availability has become a reason to pick a platform rather than a feature of one.

Chinese open-weight models have been the strongest performers relative to price for some time, which is why they keep appearing in the same conversations. We looked at the practical constraint when Kimi K3 was released and found that the weights were free and the 1,561GB required to hold them were not. That is exactly the gap a hosted offering fills: the model is available to anyone, and running it at scale is not.

So the clouds are not paying for access to something secret. They are paying for the customer who wants it served, supported and billed on one invoice.

What a listing would need this to prove

Moonshot raised more than US$2bn in May and is preparing for a possible Hong Kong listing. That context shapes how these talks should be read.

A company approaching public markets needs to demonstrate revenue that is durable and diversified rather than dependent on domestic enterprise deals and consumer subscriptions. Recurring international revenue, contracted with named hyperscalers, is close to the ideal evidence.

Which is a reason to be careful with the reporting. Talks that are early, exploratory and may not conclude are worth something to a company in this position even if they never close, because the fact of them is itself a signal to investors. Nothing here suggests the reporting is inaccurate; it suggests the timing is convenient, and that is worth noticing.

What it means from here

For businesses in this region the practical consequence, if any deal lands, is straightforwardly good. A capable model available on the cloud you already use, billed through an existing account with the compliance paperwork already done, removes most of what makes adopting a Chinese model awkward for a regulated buyer.

That matters more here than in the United States. ASEAN enterprises have less political reason to avoid Chinese models and more price sensitivity, and the barrier has been operational rather than ideological — procurement, support, data residency, someone to call. A hyperscaler hosting arrangement resolves all four.

The number to watch is not the 30 per cent, which will almost certainly be negotiated down. The real indicator is whether the auditing problem gets solved, because that mechanism is the foundation any future deal of this shape would be built on, and nobody has built one yet.