Moonshot AI Seeks Revenue Cut From Azure, AWS, and Google Cloud for Kimi K3

Beijing-based Moonshot AI is negotiating revenue-sharing agreements with Microsoft, Amazon, and Alphabet's Google that would allow the three US cloud giants to host and commercially distribute its Kimi K3 model through their enterprise platforms, Reuters reported on August 26, citing three people familiar with the talks. If completed, the arrangement would mark the first time a Chinese AI lab has secured a major revenue-sharing deal with a US cloud hyperscaler — a milestone that carries consequences extending well beyond one company's distribution strategy.

The company is seeking up to 30% of the revenue generated from K3-related services running on Microsoft's Azure, Amazon Web Services, and Google Cloud, according to Reuters — terms consistent with those Moonshot has proposed to other large commercial users of the open-weight model. The discussions remain at an early stage, with no certainty they will produce agreements. Microsoft, Google, and AWS all declined to comment. Moonshot did not respond to a request for comment.

What the talks reveal, however, is more significant than their current status. Kimi K3 is an open-weight model — its parameters are publicly downloadable and modifiable — which means no chip export control prevents it from running on US infrastructure. By negotiating to have Azure, AWS, and Google serve K3 as a managed cloud API, Moonshot would be routing a Chinese-origin frontier AI model directly to US enterprise customers through the most trusted, compliant, and widely integrated distribution channels in the market. Export controls that successfully restricted advanced Nvidia semiconductors from Chinese data centers cannot touch a model that already lives in the open on Hugging Face.

Read more: Microsoft Azure AI Revenue Runs on Three Tech Giants: OpenAI, TikTok, Meta

An Open-Weight Licensing Structure That Rewrites the Business Model

To understand what Moonshot is proposing, the distinction between open weights and open access matters. Kimi K3's parameters were released on July 27, 2026, under a Modified MIT license that technically allows download and modification by anyone. But running a 2.8-trillion-parameter model on private infrastructure is not a task most organizations can undertake. The model requires enormous GPU memory in its compressed format — a level of compute that sits well beyond the reach of most enterprise data centers and is not yet supported by common local inference tools such as llama.cpp or Ollama. As a result, "open weights" is a legal description, not a practical guarantee of self-deployment for the vast majority of commercial users.

Moonshot has structured its licensing around this reality. Any commercial operator generating more than $20 million in annual revenue from K3-related services is required to enter a commercial agreement with Moonshot — one that can include a revenue share of up to 30%, according to Reuters. Alibaba is reportedly pursuing the same licensing approach for its own open-weight model, suggesting the arrangement is emerging as a standard monetization template among leading Chinese AI labs: release weights publicly to maximize developer adoption, then capture value from the commercial tier that actually generates revenue.

DigitalOcean, which already carries Kimi K3 and other Chinese models as managed cloud services, has described the underlying logic as a well-established open-source freemium model — where basic access is free and heavy commercial deployment is monetized. What makes the Moonshot-hyperscaler talks qualitatively different from existing smaller-cloud agreements is scale: Azure, AWS, and Google Cloud together represent the primary channel through which enterprise AI buyers integrate third-party models, through marketplaces such as Azure AI Foundry and AWS Bedrock.

Chinasoft International, a Chinese IT services firm, announced a revenue-sharing agreement with Moonshot in July 2026, though it did not disclose the specific revenue split. The US cloud talks, if successful, would be the first time Moonshot has applied this structure to non-Chinese enterprise infrastructure.

Moonshot's urgency in establishing US cloud distribution has an additional dimension: the company is preparing for a Hong Kong initial public offering, with sources telling Reuters the startup raised more than $2 billion in May 2026 and is targeting a substantially higher valuation in its pre-IPO fundraising. Revenue-sharing arrangements with Azure, AWS, or Google Cloud would provide auditable commercial evidence of international demand for K3 — exactly the kind of revenue diversification and distribution credibility that institutional investors consider when pricing an AI lab's IPO. The company's founder and controlling shareholder Yang Zhilin trained as a PhD student at Carnegie Mellon before founding Moonshot in 2023. Investors including Alibaba, Tencent, IDG Capital, and entities linked to China's National Social Security Fund hold positions in the company.

Kimi K3's Architecture and Why It Needs Cloud More Than Its Predecessors Did

Kimi K3's scale makes the need for managed cloud distribution structurally different from earlier Chinese open-weight releases. The model, with 2.8 trillion total parameters, is the largest publicly available AI model by parameter count, built on architectural innovations Moonshot developed internally, including a hybrid linear-attention mechanism called Kimi Delta Attention.

Kimi Delta Attention, or KDA, interleaves full and linear attention layers in a 3:1 ratio. Standard transformer attention scales with the square of sequence length, which makes extending context windows toward one million tokens computationally prohibitive for most architectures. KDA sidesteps this by using the cheaper linear-attention computation for the majority of layers and reserving full attention for selective positions — enabling faster decoding at very long contexts and substantially reducing key-value cache memory requirements, according to Moonshot's technical documentation.

A second innovation, Attention Residuals, allows any layer to selectively retrieve representations from earlier layers rather than accumulating outputs uniformly through the stack. Moonshot says this cross-layer retrieval improves training efficiency at modest additional compute cost.

The third is the Stable LatentMoE routing framework, which manages the model's mixture-of-experts design. Of the model's 896 separate expert subnetworks, only 16 activate for any given input token. The total parameter count of 2.8 trillion describes the model's full knowledge capacity in memory; the active compute per token is closer to 104 billion parameters, making inference economics substantially more favorable than the headline number implies. Stable LatentMoE uses quantile-based load balancing rather than fragile hyperparameter tuning, which Moonshot says eliminates the routing instabilities that have caused other large MoE models to perform inconsistently at scale.

This architecture enables K3 to run a one-million-token context window at commercially relevant latency and to maintain always-on extended reasoning — its "thinking" mode cannot be disabled — while pricing at $3 per million input tokens and $15 per million output tokens. The always-on reasoning means the effective output-token cost is structurally higher than base pricing suggests, because every response includes the token budget consumed by the internal reasoning trace before the final answer appears.

Third-party benchmarking from Artificial Analysis, an independent evaluation firm, placed K3 as comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on complex multi-step reasoning tasks, and ranked it fourth on its Intelligence Index. Arena.ai ranked K3 first in benchmarks measuring web interface-building capability, according to Reuters. Moonshot's own reported figures — 81.2 on FrontierSWE, 88.3 on Terminal-Bench 2.0 — are company-reported and have not been independently reproduced in full. Artificial Analysis separately found a 51% hallucination rate in its testing, a figure that Moonshot did not include in its own published benchmark charts.

The competitive positioning of a K3 enterprise API would differ from Meta's Llama 4 series and Alibaba's Qwen3.8-Max — the two most relevant open-weight alternatives at comparable capability tiers. Llama 4, already available on AWS Bedrock, Azure AI Foundry, and Google Cloud, carries no geopolitical complications and is offered under Meta's own commercial license. Alibaba's Qwen3.8-Max, a 2.4-trillion-parameter model, is moving toward the same revenue-sharing licensing structure as K3 and raises similar China-origin concerns for regulated industries. What K3 offers that neither alternative currently matches at the open-weight tier is the combination of a one-million-token context window at commercially viable latency — the product of the KDA architecture — and pricing that benchmarks at a discount to equivalent closed-model API access from OpenAI or Anthropic. For AI developers building applications that require very long document processing or complex multi-turn agentic workflows, those two features together represent a meaningful capability gap relative to US-origin open-weight alternatives.

The Token Auditing Problem and Why Data Access Is the Sticking Point

Among the three unresolved issues Reuters identified in the negotiations — revenue split, data access, and auditing token usage — the last two are structurally linked and potentially the most politically sensitive.

Token auditing is operationally necessary for usage-based billing. In the standard AI cloud billing model, a customer pays per token processed: every character of a prompt and every character of a generated response is metered. For Moonshot to collect its revenue share from Azure, AWS, or Google Cloud, it needs an accurate count of the tokens processed under K3 service agreements — which means receiving detailed usage telemetry from the cloud platforms about how their enterprise customers are using the model.

That data creates an information channel running from US enterprise environments into a Chinese AI company's billing systems. It is distinct from the question of whether user query content transits Moonshot's servers (with managed cloud hosting, it likely would not), but it represents a different kind of data exposure: operational usage patterns, consumption volumes, customer categories, and usage intensity signals that a sophisticated actor could use to understand enterprise AI adoption trends across the US economy. The unresolved nature of this issue in the negotiations suggests the cloud companies are aware of the sensitivity.

The legal context sharpens the concern. Moonshot AI is subject to Article 7 of China's National Intelligence Law, enacted in 2017, which legally obliges all Chinese organizations to support, assist, and cooperate with national intelligence efforts. China's Data Security Law of 2021, as amended effective January 1, 2026, and the Cybersecurity Law of 2017 impose additional data localization requirements and government inspection authority that explicitly extend to AI systems. These obligations apply regardless of where inference runs, regardless of Moonshot's Singapore incorporation, and regardless of the specific privacy policy governing any cloud partnership. If a revenue-sharing agreement requires Moonshot to receive token-level usage telemetry from Azure, AWS, or Google Cloud, the legal framework governing what Moonshot can be compelled to do with that information exists independently of any contractual commitment Moonshot makes to its cloud partners. Trade attorneys advising the hyperscalers would need to weigh whether any data-access concession to Moonshot, however limited in scope, creates a channel that falls within reach of those legal obligations.

A Company Operating Under Active US Scrutiny

The cloud talks are proceeding against a backdrop of unresolved government allegations that would complicate any deal's path to completion.

In July 2026, White House officials publicly accused Moonshot of distilling Anthropic's Fable model to help develop Kimi K3 — and of acquiring Nvidia GB300 servers, Blackwell-generation chips restricted from sale to Chinese companies, by accessing them in Thailand. Treasury Secretary Scott Bessent followed with a warning that sanctions and Entity List designations would be on the table for Chinese firms conducting large-scale distillation attacks. The Bureau of Industry and Security opened a formal investigation. No enforcement action has been taken as of August 27, 2026.

The distillation allegation has attracted significant skepticism from independent AI researchers. Anthropic's Fable 5 returned to full public availability only on July 1, 2026, after a June export-control suspension, giving Moonshot a window of just 15 days before K3 was released on July 16. Researchers familiar with the training timelines required for a 2.8-trillion-parameter model have concluded that building one from scratch through distillation in that interval — without the underlying architecture already being substantially complete — is not technically plausible. Braden Hancock of Snorkel AI, commenting in July on the timeline, argued that the gap between Fable's availability and K3's release was far too short for Fable to have been the primary training source. Nathan Lambert, an independent AI researcher, concluded that while some distillation may have occurred, it was clearly not the whole story.

What is independently documented — and what exists as a separate, legally stronger claim — is an earlier episode: in February 2026, Anthropic published an investigation finding that Moonshot generated more than 3.4 million exchanges with Claude through approximately 24,000 fraudulently created accounts, targeting agentic reasoning, tool use, coding, and computer vision. Moonshot denied those allegations. The legal framework most directly applicable to that campaign is the Computer Fraud and Abuse Act, not copyright law — a distinction that matters because model distillation, without more, does not clearly constitute copyright infringement under current US law, a point legal scholars and independent analysts have made explicit.

Moonshot rejected the July 2026 allegations through its head of enterprise, Huang Zhenxin, who told China's National Business Daily that K3's performance gains came from original changes to underlying architecture. Moonshot's proprietary architectural innovations — Kimi Delta Attention and Attention Residuals — were present in its earlier Kimi Linear model released in October 2025, predating any availability of Fable 5. That timeline gives K3's novel design elements an independent origin that is inconsistent with the distillation-as-primary-source hypothesis.

China's Ministry of Commerce responded to the US threats by branding the position "AI hegemonism" and warning of countermeasures. Beijing also separately held meetings in July 2026 with Alibaba, ByteDance, and Z.ai about potentially restricting overseas access to China's most advanced AI models — a mirrored response that would constrain Chinese AI's global distribution from the other direction if implemented.

What the Cloud Companies Risk and What They Stand to Gain

For Azure, AWS, and Google Cloud, the calculus involves competing pressures. On the revenue side, K3 represents a credible frontier-tier model that Artificial Analysis independently rates as comparable to the leading US models on demanding tasks, at pricing that undercuts closed-model equivalents. Adding a model of K3's capability to an enterprise API marketplace expands the options available to developers who are already evaluating Chinese open-weight alternatives through smaller cloud providers like DigitalOcean.

Azure has operational precedent for this kind of commercial relationship with Chinese AI companies: ByteDance, the Chinese parent of TikTok, is Foundry's largest non-OpenAI spender and was on track for more than one billion dollars in annual Microsoft AI and cloud spending as of mid-2026, purchasing access to OpenAI models through Azure for its own services. A K3 agreement would be structurally different — Moonshot would be the model provider receiving a revenue share, rather than a customer purchasing compute — but the broader pattern of Chinese AI companies relying on US cloud infrastructure is well established.

The regulatory risk, however, is immediate and concrete. The Bureau of Industry and Security's investigation into Moonshot remains open. If it results in an Entity List designation before any deal closes, hosting K3 through a US cloud platform could constitute a violation of export-control regulations — a consequence the cloud companies would need to model explicitly. The Huawei precedent from 2019 illustrates what Entity List designation does in practice: it severed Huawei's access to US hardware, software, and cloud services, and created downstream liability for US companies that continued providing services without explicit authorization. A Moonshot designation under similar circumstances would place any signed revenue-share agreement in immediate legal jeopardy and could require unwinding.

The trajectory of DeepSeek's presence on US cloud platforms offers a relevant recent reference point. DeepSeek's models are currently accessible through several US cloud providers and API aggregators, with varying levels of official marketplace integration, despite the company facing similar national security scrutiny. No formal restriction has been imposed on hosting DeepSeek open-weight models on US infrastructure as of August 2026 — a legal gap that reflects the absence of a model-layer framework in existing export-control regulations. Moonshot is attempting to establish its cloud distribution agreements in the same regulatory window, but with the additional weight of a named government investigation and named Treasury threats that DeepSeek has not faced at the same level of official specificity.

The export-control framework that currently targets chips does not explicitly restrict model hosting, but a fall 2026 interim final rule that trade attorneys and policy analysts have widely anticipated would extend controls to model access and remote inference. Timing a deal against that regulatory trajectory requires precise legal navigation.

The Export-Control Gap That Moonshot's Strategy Exposes

The deeper significance of the K3 cloud talks is structural. Since 2022, US policy toward China's AI development has centered on restricting access to advanced semiconductors — Nvidia A100s, H100s, and now the GB300 Blackwell series. The logic was that compute bottlenecks would slow Chinese labs' ability to train and run frontier models. The K3 situation reveals the limits of that strategy when applied to the model layer.

K3's weights are already public. No chip restriction can recall them. Any US cloud provider that hosts K3 would be running inference on hardware it owns, in data centers it controls, under its own security framework — without Chinese hardware. The model's origin becomes operationally irrelevant to the infrastructure that serves it. Export controls designed around chip flows cannot prevent frontier-quality Chinese AI from reaching US enterprise customers once the model weights are openly distributed.

Regulatory analysts and trade attorneys have identified precisely this gap, noting that the forthcoming US policy revision is expected to target model access and remote inference rather than chips — the first time the framework would reach the software layer. Beijing's own discussions about restricting Chinese model exports represent a mirror-image recognition: that open-weight releases are now a strategic variable that both governments need to manage.

For the three hyperscalers in the middle of this conversation, completing a K3 revenue-share deal before that regulatory framework is finalized would give the arrangement an established commercial history that could complicate future restriction. Completing it after would require explicit regulatory clearance. The open question — whether any of the three will close a deal, and when — will determine which side of that timing equation the transaction falls on, and how much precedent it sets for the next Chinese frontier model that follows K3 through the same door.