Kimi is demanding a 30% cut from Microsoft, Amazon, and Google; its open-source model is starting to "collect rent overseas."

Recently, foreign media revealed that Dark Side of the Moon is in talks with Microsoft, Amazon, and Google, hoping to bring Kimi K3 onto Azure, AWS, and Google Cloud, and take up to 30% of the revenue from selling K3-related services on these platforms. If finalized, this could be one of the first significant model revenue-sharing agreements between a Chinese AI company and major US cloud providers.

Negotiations are still in the early stages: the two sides have not yet fully agreed on the revenue sharing methodology, the usage data that Dark Side of the Moon will obtain, and how the tokens sold by the cloud vendors will be audited. All three cloud vendors and Dark Side of the Moon declined to comment on the negotiations.

Meanwhile, Dark Side of the Moon has signed similar agreements with some smaller cloud platforms, and China Software International, a Chinese IT service provider, has also publicly announced that it will share the token revenue and other cooperative derivative revenue generated by Dark Side of the Moon around the Kimi model .

Over the past year, the most important task for Chinese open-source models going global has been to get overseas developers to use them first: open up weights, reduce API prices, get them onto Hugging Face and OpenRouter, and get as many inference platforms as possible to launch the models, thereby gaining overseas influence.

But once the model is open source, how can it continue to generate revenue from the vast overseas ecosystem?

This message suggests that Dark Side of the Moon's attempt is to turn the overseas usage of the model into a new revenue stream: cloud providers can purchase GPUs, deploy Kimi, and then sell inference services to customers. However, once the business grows large enough, open-source model companies can also share in the profits.

Image generated by AI.

01 has already started testing the waters on small and medium-sized platforms?

In addition to the news of negotiations among the three major cloud vendors, Dark Side of the Moon's own developer website has now set up a dedicated "Access Kimi models anywhere" entry, and explicitly stated that K3 can be accessed through third-party inference partners.

Based on publicly available information from various platforms, K3 has been adopted by a number of inference service providers, including Together AI, Fireworks, DigitalOcean, Modal, Baseten, and DeepInfra. If we also include other small and medium-sized inference platforms that have already provided the K3 API, there are more than 10 providers available on the market.

Image: Kimi's reasoning partner

Here is a very important but easily overlooked distinction: not all of these companies are API resellers.

The so-called API reseller is essentially just a distributor. A user purchases the Kimi service on platform A, but platform A itself does not run Kimi. Instead, it forwards the request to Dark Side of the Moon or another upstream provider that actually runs the model. It earns the difference between the API purchase price and the retail price. Where the model actually runs and how it optimizes inference are irrelevant to it.

True model hosting is entirely different. The provider directly obtains the open weights of K3, deploys this 2.8 trillion parameter model on its own GPU cluster, and handles model parallelism, inference engine, memory management, scaling, and stability issues itself. Finally, it sells tokens to customers using its own API. In this case, the model vendor, such as Dark Side of the Moon, no longer bears the inference cost of each call; the cloud vendor is the true provider of computing power.

It has now been confirmed that a significant portion of K3 Providers belong to the latter category.

In late July, Together AI announced a strategic partnership with Dark Side of the Moon, explicitly stating that it would use its own US infrastructure to "natively serve" the Kimi model. Beyond K3, future open weighted models released by Dark Side of the Moon will also continue to be integrated into Together's platform. Modal similarly stated that they are collaborating with Moonshot and vLLM to achieve day-zero support for K3, offering both shared APIs and independent deployment options.

Image: Together AI × Moonshot strategic partnership official announcement page

DigitalOcean disclosed that in order to launch K3, it needed to select its own hardware, adjust the inference stack, optimize the model, and then validate it against the Moonshot benchmark. In other words, when users call K3 through DigitalOcean, the underlying GPU and inference infrastructure belong to DigitalOcean, not Moonshot.

Image: DigitalOcean's Kimi K3 inference deployment architecture, which integrates DigitalOcean's proprietary inference agent and distributed inference stack, and is specifically adapted for dynamic tool calls in K3. Image source: DigitalOcean.

It is currently confirmed that Dark Side of the Moon has established about ten or more third-party inference channels, many of which are K3 hosting providers that actually bear their own computing power; at the same time, some small and medium-sized cloud platforms have further accepted revenue-sharing business agreements.

02 Kimi has already vaguely formed an overseas channel system

These collaborations also reveal that Kimi is exploring a lighter approach to overseas expansion than selling its own APIs. Previously, model companies had to rent GPUs, build inference clusters, and then sell tokens to developers; now, Dark Side of the Moon doesn't need to deploy an expensive global inference network itself, but can leverage cloud providers' existing computing power and enterprise channels to enter overseas markets.

What truly made this model work was that Kimi added a business threshold in addition to opening up the weights.

Kimi K3 uses a custom Kimi K3 License developed by Dark Side of the Moon. The first part of the agreement follows a permissive licensing approach similar to MIT, allowing use, modification, deployment, and commercialization, but adds additional commercial conditions for large MaaS enterprises. If a company engaged in Model as a Service business and its affiliates have total revenue exceeding $20 million for 12 consecutive months, they will need to sign a new commercial agreement with Dark Side of the Moon.

The $20 million threshold is merely the starting point for negotiations and does not imply an automatic 30% cut-off for amounts exceeding that threshold. The "maximum 30%" disclosed by foreign media is one of the specific business conditions that Dark Side of the Moon is currently proposing to major clients.

Another detail surrounding K3 is worth noting. Currently, the prices of K3 sold on multiple independent inference platforms are unusually close: the official API of Dark Side of the Moon is approximately $3 per million tokens input and $15 per million tokens output, while platforms such as Together and Modal are also in roughly the same price range, and some providers such as DigitalOcean are slightly lower.

Different inference companies have different GPU procurement costs, scales, and inference frameworks. Theoretically, they can price open-weight models independently. Therefore, this price concentration is not a natural consequence of open-source models.

While there is currently no evidence that Moonshot has set a uniform price for its partners, it is clear that K3 is not downloaded, deployed, and sold independently by each cloud vendor, but rather entered these platforms with Moonshot's involvement in technical adaptation and collaborative operation.

A distribution network is emerging: Moonshot is responsible for training models, maintaining the model brand, and adapting the technology to providers; the inference platform is responsible for purchasing computing power, deploying models, and finding customers; and finally, some partners share commercial revenue with model manufacturers according to the agreement.

If AWS, Azure, and Google Cloud eventually join, Kimi will gain more than just three new API entry points; it will gain a distribution network that extends from small and medium-sized professional inference platforms to the global large enterprise cloud market.

03 Are the three major cloud providers "easy to negotiate"?

Compared to smaller inference platforms that have already quickly launched K3, Microsoft, Amazon, and Google are still in the negotiation stage. Foreign media reports indicate that the main disagreements focus on three issues: revenue sharing, data access, and auditing of token usage.

These things become much more complex on large cloud platforms. Inference platforms like Together primarily sell model services directly based on tokens, making it relatively easy to calculate how much money K3 generated.

However, AWS and Azure also offer annual contracts, enterprise discounts, reserved capacity, and bundled sales of various AI products for large enterprises. If K3 is just one model in a multi-million dollar cloud contract, the exact amount of revenue that should be credited to Kimi directly determines how much of the revenue Moonshot can receive.

Token auditing also involves data boundaries. Moonshot needs to confirm how many K3 services the cloud platform actually sold before it can complete the settlement; cloud vendors, in turn, need to protect enterprise customer data.

These problems are much more complex to solve than those faced by small and medium-sized cloud vendors.

04 Qwen is also exploring the same model

The Dark Side of the Moon is not the only Chinese model company starting to redesign its open-source business model. Reportedly, Alibaba is also considering a revenue-sharing mechanism for large Qwen commercial users, and the new Qwen license has begun setting separate commercial licensing thresholds for large MaaS and AI Work Assistant businesses.

Alibaba's definition of AI Work Assistant includes both AI coding and standalone office productivity products. This means that in the future, model companies may charge not only cloud platforms like AWS or Together, but also a large number of coding tools, office agents, and other AI applications built on open models. MiniMax has adopted a similar approach, requiring commercial users who reach a certain revenue scale to obtain written authorization again.

However, not all Chinese model companies are moving towards "charging based on scale." Tencent's latest Hy4 preview and Step 3.7 Flash both use the standard Apache 2.0 license and do not set additional revenue thresholds for large-scale MaaS; DeepSeek V4 pro continues to use the MIT License, similarly without requiring third-party businesses to re-license or share revenue after their businesses grow large.

This means that even if an overseas cloud vendor deploys these models on its own and generates a large-scale business by selling tokens, it does not need to pay the model company any additional fees based solely on the model license.

Z.AI's latest flagship GLM-5.3 has switched from the MIT license previously used for some models to a custom GLM-5.3 license. The new hurdle doesn't seem to be for revenue generation: only companies operating MaaS businesses with total revenue exceeding $10 billion for 12 consecutive months are required to pass Z.AI's security review before commercial use. This extremely high threshold primarily targets large-scale global cloud platforms, reflecting more of a model governance and security control issue than revenue sharing.

This led to two different paths emerging in China's open model. One continued to use free and permissive licensing as its biggest competitive weapon, aiming to gain deployment volume and market share with the lowest barriers to entry; the other began to try "free diffusion, charging at scale," first allowing the model to enter as many cloud platforms and AI applications as possible, and then charging participants who had already achieved commercial value at scale.

05 Will the "model royalties" eventually be established?

From a revenue perspective, this model is very attractive. Assuming a global third-party cloud platform sells $1 billion worth of Kimi services annually, if Moonshot receives an average 20% revenue share, that translates to $200 million in revenue; if it reaches 30%, it would be $300 million. More importantly, this revenue doesn't require Dark Side of the Moon to bear the entire investment in global inference infrastructure alone.

However, its limitations are equally apparent: open models can be replaced very quickly. Today, Kimi's performance is leading, and cloud vendors may be willing to accept higher commercial licensing fees; a few months later, if DeepSeek, Hunyuan, Qwen, or another open model achieves similar performance while having lower inference costs and more lenient licenses, Kimi's bargaining power will decline.

Therefore, 30% is more like the maximum price a model can offer during its peak, rather than a long-term "model tax" that has been established. Whether model companies can truly establish a sustainable revenue stream similar to software licensing or chip IP ultimately depends on how difficult the model is to replace and how high the cost is for developers and companies to migrate to competing models.

If this revenue-sharing model is ultimately established on the three major cloud providers, the business logic of Chinese open-source models going global will undergo a significant change. In the past, model companies mainly relied on selling tokens through their own APIs; in the future, they may directly entrust their models to global cloud vendors for deployment and then continuously share the revenue generated from these channels.

This article is sourced from Tencent Technology.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.