Moonshot Wants U.S. Clouds to Host Kimi K3

Moonshot AI is negotiating with Microsoft, Amazon and Google over deals that could put its Kimi K3 model on the biggest U.S. cloud platforms. Reuters reports that the Chinese startup is seeking as much as 30% of revenue generated from K3-related services, although the talks are still early and may not produce agreements. The unusual part is not simply that a Chinese AI company wants access to American cloud infrastructure. It is that Kimi K3's open-weight model may need those clouds precisely because its enormous size makes running it independently difficult for most customers. 

Moonshot is negotiating with all three major U.S. clouds

The discussions involve Microsoft's Azure, Amazon Web Services and Google Cloud, according to three people familiar with the talks cited by Reuters. Moonshot is reportedly seeking up to a 30% share of revenue generated through K3 services on those platforms, broadly matching terms it has offered major customers using the model. Microsoft, Google and AWS declined to comment, while Moonshot did not respond to Reuters' request for comment. Because the negotiations are private and unresolved, the proposed revenue split should be treated as a reported negotiating position rather than a finalized commercial arrangement. 

If completed, the agreements would create a particularly interesting business model. The cloud provider would supply the computing infrastructure and distribution, while Moonshot would receive a share of the resulting usage revenue. That would give developers access to Kimi K3 without having to purchase and operate the hardware needed to run a model of its scale. It would also give Moonshot a route into enterprise customers that may prefer buying AI through an existing cloud platform rather than building a separate infrastructure stack.

Kimi K3 is open-weight, but that does not make it easy to run

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model, meaning it contains a very large collection of learned parameters but activates only part of that network for each piece of work. Moonshot's technical paper describes 104 billion activated parameters and a context window of up to one million tokens, alongside native vision capabilities. The model is designed for long-horizon coding, knowledge work and reasoning rather than simple chatbot exchanges. 

Open weights change who can access and modify a model's underlying parameters, but they do not eliminate the cost of inference. A model can be legally downloadable and still be impractical for an ordinary company to operate at useful speed and scale. Reuters notes that K3's 2.8-trillion-parameter size makes self-hosting expensive for many customers, which creates an obvious role for cloud providers. 

The cloud route is already proving useful for Kimi K3

Moonshot does not have to start from zero. Databricks made Kimi K3 available through its Foundation Model API in August, giving customers access to the model with centralized governance, spending controls and auditing. Databricks says K3 can be used alongside models from Anthropic, OpenAI and Google, allowing organizations to select different models for different workloads rather than committing to a single provider. 

That existing deployment shows why the reported negotiations with the hyperscalers could matter. The opportunity is not merely to make K3 downloadable; it is to turn an open-weight model into a service that fits the same procurement, security and monitoring systems enterprises already use. Databricks says K3 is initially hosted in the United States and reports that its customers can use the model through a standardized API, which removes much of the infrastructure work that would otherwise fall on the customer. 

K3 has enough performance to attract serious cloud interest

Kimi K3 has also generated independent attention because its performance is competitive with leading proprietary models on some demanding evaluations. Artificial Analysis ranked it second on its AA-Briefcase agentic knowledge-work benchmark, behind Claude Fable 5, with an Elo score of 1543. But the same evaluation found a significant trade-off: K3 averaged $10.57 per task and nearly 56 minutes per task, reflecting high token use and many model turns. 

That trade-off helps explain why cloud economics are so important. Strong model quality can attract customers, but a model that consumes large amounts of compute can become expensive when used continuously. A cloud provider can optimize scheduling, hardware utilization and access across many customers in ways an individual company may struggle to reproduce. For Moonshot, the goal is therefore not simply to prove that K3 is capable; it is to make that capability practical enough that companies will pay to use it repeatedly.

Revenue sharing would change who controls the AI business

Traditional cloud AI arrangements often revolve around infrastructure consumption: customers pay the cloud provider for computing resources or an API service. A revenue-sharing agreement puts the model developer more directly into the economics of every request. If Moonshot receives a percentage of the revenue, it has an incentive to drive usage through the cloud platform rather than simply selling software licenses or giving customers downloadable weights.

That could also make model distribution more competitive. Cloud customers increasingly want the ability to switch among models depending on price, quality, privacy requirements and the type of task. Databricks is already presenting K3 in that way, describing it as one model in a broader portfolio rather than a replacement for every proprietary system. 

The biggest obstacle may be politics, not performance

The reported talks arrive while U.S. officials are scrutinizing Moonshot. Reuters reports that Treasury Secretary Scott Bessent has discussed potentially placing the company on a trade blacklist, while U.S. officials have accused Moonshot of intellectual-property theft and illegally acquiring Nvidia chips. Moonshot has rejected claims that K3's performance came from distillation, saying its gains resulted from changes to the underlying architecture. 

That makes a potential Azure, AWS or Google Cloud agreement much more complicated than an ordinary model-hosting deal. A cloud provider would have to consider not only technical demand and commercial terms, but also data access, compliance, auditing and the broader U.S.-China technology relationship. Reuters says those issues are already part of the negotiations, with data access and token-use auditing among the unresolved questions. 

What happens if the deals go through

A successful agreement would give Kimi K3 something open-weight releases often lack: simple access at enterprise scale. Developers could use the model through familiar cloud infrastructure without purchasing the hardware needed to operate a 2.8-trillion-parameter system themselves. Moonshot, meanwhile, would gain distribution through three of the world's most important AI infrastructure companies.

But there is a larger question behind the negotiations. If U.S. clouds become willing to host powerful Chinese-developed open-weight models, the boundary between AI model origin and AI infrastructure could become much less meaningful. The model may be Chinese, the servers may be American, and the customer may be anywhere in the world. For now, none of the three reported deals is final. The next development to watch is whether commercial demand for Kimi K3 is strong enough to overcome the political and operational questions surrounding its deployment.