Moonshot AI 希望从美国云服务商的 Kimi K3 收入中分得 30%

据路透社报道，Moonshot AI 正在与 Microsoft、Amazon 和 Google 就在其云平台上托管 Kimi K3 进行初步讨论，并寻求获得这些服务所产生收入的最高 30%。

这可能成为中国 AI 公司与美国主要云服务商之间首个此类合作安排，涉及 Moonshot 作为全球最大开放模型发布的模型。Kimi K3 发布时曾引发市场震动。

K3 的参数规模达到 2.8 万亿。尽管它在许可层面是开放的，但对绝大多数组织而言，自行托管并不现实，因为为其提供服务所需的算力，只有极少数公司能够集齐。

因此，开放权重并没有消除对云服务商的需求；它改变的是云服务商所面对的竞争对象。Moonshot 可以获得自身无法建立的分发能力，而超大规模云服务商则能获得一款在与美国前沿模型的竞争中表现优异的模型；Arena.ai 将 K3 在网页界面工作方面评为第一。

目前，这项交易的大多数关键内容尚未敲定。收入分成比例、数据访问条款以及如何审计 Token 用量均仍未确定，而 Token 数量正是按使用量计费的基础。

审计问题比听上去更为敏感。双方需要就如何统计模型处理的文本单元达成一致，而在这一安排中，负责计数的一方同时也是支付分成的一方。

数据访问可能是最有可能决定交易成败的条款。一家中国公司将模型托管在 Azure 或 AWS 上，会立即引发其能够查看多少经由系统流转的提示词和输出内容的问题；企业客户在将任何敏感信息接入之前，也会希望获得明确答案。

政治层面的难度远高于会计核算。美国财政部长 Scott Bessent 曾建议将 Moonshot 列入贸易黑名单；美国官员则指控该公司蒸馏其他公司的模型并非法获取芯片。Moonshot 否认了这些指控。

三家美国云服务商若与一家财政部长希望列入黑名单的公司签署收入分成协议，并不是一项简单的安排。这些谈判能否达成，可能较少取决于具体条款，而更多取决于华盛顿是否认定其在法律和政策上可被允许。

对 Moonshot 而言，时机具有重要商业意义。该公司一直在筹备以 300 亿美元估值在香港上市，而通过全球最大云服务商实现国际分发，将实质性改变其向投资者讲述的收入增长故事。

其估值轨迹本身也是一个故事。在六个月内估值增长七倍后，Moonshot 达到了300 亿美元的目标估值。这种重新定价速度，使分发合作变得紧迫，而不再只是机会性的选择。

其估值已经以需要解释的速度上涨。Moonshot 在大约六个月内从约 200 亿美元升至 300 亿美元目标估值；一家仅向中国客户出售模型访问服务的企业，其增长上限低于一家能够通过 Azure、AWS 和 Google Cloud 销售服务的企业。

欧洲也存在与这项交易相冲突的主权论点：过去两年里，当地买家一直被告知要降低对外国模型的依赖。由美国云服务商提供服务的中国模型，并不显然能够回应这一担忧，尽管它可能是成本最低的选择。

超大规模云服务商也有自身的考量。客户持续要求更便宜且能力出色的模型；K3 同时具备这两点。相比托管该模型，另一种选择是眼看企业通过其他方式接入它——而这恰恰是此前每一次云服务商决定提供竞争对手模型时所遵循的逻辑。

30% 的分成比例较为激进，不过据报道，这与 Moonshot 向大型客户提出的要求相符。它也颠倒了通常的安排：通常由平台抽成，开发者获得剩余收入。

目前尚未达成任何协议，四家公司均拒绝置评，除确认没有更多信息外未作回应。由消息人士披露的早期谈判，是最可能悄然终止的交易报道类型之一，尤其是那些需要先解决美国政策问题的谈判。

Moonshot AI is in early discussions with Microsoft, Amazon, and Google about hosting Kimi K3 on their cloud platforms, seeking up to 30% of the revenue those services generate, according to Reuters.

It would be the first arrangement of its kind between a Chinese AI company and the major American clouds, involving a model that Moonshot released as the world’s largest open model and that rattled markets when it landed.

The model K3 runs to 2.8 trillion parameters, which makes it open in licence but impractical for most organisations to host themselves, because the compute required to serve it is beyond what all but a handful of companies can assemble.

Open weights therefore do not remove the need for a cloud provider; they change who the cloud provider is competing with. Moonshot gets distribution it could not build, and the hyperscalers get a model that has been performing well against the American frontier, with Arena.ai ranking K3 first for web interface work.

What has not been settled is most of the deal. The revenue split itself, data access provisions, and how token usage would be audited are all open, and token counts are what usage-based billing rests on.

That auditing question is more delicate than it sounds. Both sides would need to agree a method for counting the units of text a model processes, in an arrangement where the party doing the counting is the party paying the share.

Data access is the provision most likely to decide this. 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, and enterprise customers will want an answer before they route anything sensitive.

The politics are considerably harder than the accounting. US Treasury Secretary Scott Bessent has suggested adding Moonshot to trade blacklists, and American officials have accused the company of distilling other companies’ models and acquiring chips illegally, allegations Moonshot denies.

Three US cloud providers signing revenue-sharing agreements with a company the Treasury Secretary wants blacklisted is not a straightforward proposition. Whether these talks conclude may depend less on the terms than on whether Washington decides they are permissible at all.

For Moonshot, the timing is commercially significant. The company has been preparing a Hong Kong listing at a $30bn valuation, and international distribution through the largest clouds in the world would materially change the revenue story it takes to investors.

The valuation trajectory is its own story. Moonshot reached a $30bn target after a sevenfold jump in six months, which is the kind of repricing that makes distribution deals urgent rather than opportunistic.

Its valuation has already moved at a pace that requires explanation. Moonshot went from roughly $20bn to a $30bn target in about six months, and a business selling model access to Chinese customers alone has a smaller ceiling than one selling through Azure, AWS, and Google Cloud.

There is also a sovereignty argument that cuts against the deal in Europe, where buyers have spent two years being told to reduce dependence on foreign models. A Chinese model served through an American cloud is not obviously the answer to that concern, though it may be the cheapest one.

The hyperscalers have their own calculation. Customers keep asking for cheaper, capable models; K3 is both, and the alternative to hosting it is watching enterprises route to it some other way, which is exactly the argument that has driven every previous decision to carry a rival’s model.

A 30% share is aggressive, though it reportedly matches what Moonshot has been asking large customers. It also inverts the usual arrangement, in which the platform takes the cut, and the developer receives the remainder.

Nothing has been agreed, and all four companies declined to comment beyond confirming nothing further. Early-stage talks reported by sources are the most likely of all deal stories to end quietly, particularly ones that require a US policy question to be resolved first.