China’s Kimi K3 Could Soon Run on Microsoft, Google and Amazon: What Is Moonshot AI?

A Chinese AI company facing scrutiny from Washington is trying to put its flagship model on the servers of three American cloud giants. The commercial logic is straightforward. The political logic is anything but.

“The discussions are preliminary and no agreements have been signed,” Reuters reported, citing three people familiar with the talks. The discussions are preliminary, nothing has been signed, and the three cloud companies have not commented.

If completed, the arrangement could become the first major revenue-sharing deal between a Chinese AI company and a major U.S. cloud provider.

Moonshot is reportedly seeking as much as 30% of the revenue generated by K3-related services. The sides still need to resolve how that revenue would be split, what data the cloud providers could access and how token usage would be audited. For an AI service billed by usage, accurate token metering is central to the economics.

K3 Needs More Than Open Weights

“Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window,” Moonshot AI says in its official K3 announcement.

Moonshot recommends running it across supernode configurations with at least 64 accelerators. The model’s weights are open and downloadable. The infrastructure required to run those weights at useful scale is the harder part.

An individual developer can download K3. Very few can supply the accelerators, networking and memory needed to serve a model of this size efficiently. Microsoft, Amazon and Google already own that infrastructure.

Moonshot Has Serious Money Behind It

Moonshot AI is a Beijing-based company founded in 2023 by Yang Zhilin, who studied at Tsinghua University and earned a PhD from Carnegie Mellon before working on AI research connected to Google and Meta.

Kimi is Moonshot’s product line. K3, released in July 2026, is its largest model so far. Reuters reports that Moonshot has raised more than $2 billion and is preparing for a possible Hong Kong listing, citing people familiar with the matter.

Moonshot raised roughly $2 billion at a $20 billion valuation in May, backed by investors including Alibaba, Tencent and HongShan. The company has since attracted additional funding and is reportedly preparing for a possible Hong Kong listing.

That makes the proposed cloud arrangement less a rescue effort than an attempt by a heavily funded AI company to turn an open model into a widely available commercial service.

Washington Is The Complication

Treasury Secretary Scott Bessent has raised the possibility of adding Moonshot to a U.S. trade blacklist. U.S. officials have also accused the company of distilling Anthropic’s advanced model, referred to in the reporting as Fable, to help develop K3 and of improperly acquiring restricted Nvidia chips.

“The technical improvements of Kimi K3 come from our original architectural innovations and training methods,” Moonshot said in response to the U.S. allegations, according to Reuters.

That creates the central contradiction in the story: a Chinese AI company facing possible U.S. trade restrictions is simultaneously negotiating to put its model on the infrastructure of three of America’s biggest technology companies.

The issue is larger than corporate risk. Washington has already restricted exports of advanced AI chips to China over national-security concerns. Any arrangement that puts a Chinese model on American cloud infrastructure therefore sits directly inside an existing technology-control fight.

K3 Has Become Competitive

Third-party evaluations have put the model near the top on specific tasks. Arena.ai ranked K3 first in a benchmark measuring web-interface-building capability, while Artificial Analysis assessed its performance as comparable to OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8 on complex, multi-step tasks.

Those results should not be stretched into a claim that K3 is the overall best model. Benchmark results vary by task, and Moonshot itself acknowledges that K3 trails the strongest proprietary systems overall.

But the results establish something more important for this deal: K3 is competitive enough that putting it behind a major cloud API could have commercial value.

Developers Don’t Get New Access Yet

For developers, nothing has changed. K3 remains available through Moonshot’s own Kimi app and API. It has not been confirmed on Azure, AWS or Google Cloud, and the negotiations could still collapse over commercial terms, technical requirements or political pressure.

If the deals go ahead, customers would gain another way to use K3 without having to purchase and operate the hardware required to run it themselves. That is where the open-weight model business gets interesting.

Open Models Still Need Big Clouds

“We are currently working closely with inference partners and open-source maintainers to align technical details and ensure a reliable rollout across the ecosystem,” Moonshot says in its K3 announcement.

It can distribute K3’s weights openly while allowing Microsoft, Amazon and Google to monetize customers who would rather access the model through a managed cloud service. Moonshot would then take a share of that inference revenue instead of building the entire infrastructure layer itself.

The harder question is whether those cloud providers are willing to share that revenue while Washington is scrutinizing the company behind the model. If Microsoft, Amazon and Google ultimately host K3, the deal would show that open-weight AI is creating a business relationship that national-security policy has not yet fully figured out.