The number is 30%. That’s the revenue share Moonshot AI is reportedly seeking from Microsoft Azure, Amazon Web Services, and Google Cloud in exchange for the right to offer Kimi K3, Moonshot’s flagship model, to enterprise customers on their platforms.
It’s the same number Apple takes from every app sold through the App Store.
That’s not a coincidence. It’s a negotiating position. And if it works, the economics of AI deployment won’t look anything like they do today.
What Moonshot Is Actually Asking For
Moonshot AI is a Chinese AI startup that’s currently IPO-bound and quietly in talks with all three major U.S. cloud providers. The deal structure it’s proposing is straightforward: Moonshot lets Azure, AWS, and Google Cloud offer Kimi K3 to their enterprise customers. In return, Moonshot gets up to 30% of what those cloud providers earn from services tied to K3.
Kimi K3 is worth talking about. It’s a 2.8-trillion-parameter open-weight model, the largest publicly available of its kind, and it’s not a niche research artifact. Arena.ai placed it at the top of its leaderboard for web interface construction. On demanding multi-step reasoning tasks, Artificial Analysis found its results comparable to OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8.
That’s not a model you ignore. That’s a model enterprise customers will want access to, and cloud providers know it.
None of the companies involved would comment. The talks are described as early stage and may not produce agreements. The sticking points aren’t small: exact split percentages, data access terms, and the mechanics of tracking token consumption all remain unresolved. These are genuinely hard problems when you’re trying to divide revenue tied to a model that processes text in chunks.
But the talks are happening. That’s the part that matters.
The Inversion
Here’s the standard cloud AI playbook of the last three years: an open-weight model gets released, the cloud providers pick it up, host it at scale, and offer it to enterprise customers who can’t run a 70-billion-parameter model on a laptop. The model maker gets visibility and adoption. The cloud provider gets a new revenue line. Nobody writes a check to anyone.
Moonshot is proposing something different. It’s saying: your enterprise customers want this model specifically, and you can’t give it to them without us. So we want a percentage of what you make.
That’s the App Store logic, inverted. Apple built the platform and charged developers 30% for access to its customers. Moonshot built the product and is charging the platform for access to its technology. The direction is different. The number is the same.
What makes this notable isn’t just that Moonshot is asking. It’s that the ask is plausible. Kimi K3 isn’t a commodity model that any of the three clouds could replicate tomorrow. It’s a specific, well-benchmarked product with documented performance advantages in certain task categories. Moonshot has genuine leverage.
If It Works
The downstream effects are significant.
The cloud providers built much of their current AI advantage partly by hosting open-weight models at no licensing cost. Meta’s Llama family, Mistral’s releases, various fine-tunes and derivatives: these were free to pick up and add to a cloud catalog, and they gave cloud platforms a way to offer AI capabilities without paying for them.
Moonshot’s ask changes the calculation. If a major Chinese AI company extracts a 30% revenue share from one or more of the big three U.S. cloud providers, the template is set. Every lab with a model that enterprises want will have a reason to try the same. Llama 5, whenever it arrives, won’t be free to commercialize in the same frictionless way if Moonshot proves the alternative is viable.
This isn’t hypothetical. It’s the natural endpoint of open-weight models becoming commercially valuable enough that the organizations that created them want compensation for that value.
If It Doesn’t Work
The talks could collapse. The 30% figure could be a starting position designed to land at 15%, or the sticking points on data access and token tracking could prove genuinely intractable, or the geopolitical dimension of a Chinese AI company extracting revenue from U.S. cloud infrastructure could cause one or more parties to walk.
But even a failed negotiation at this scale is a data point the industry doesn’t have yet. Before today, the conventional wisdom was that open-weight model makers didn’t negotiate distribution terms with the cloud providers who hosted them. That assumption is now gone regardless of what happens next.
The question isn’t really whether Moonshot gets its 30%. The question is whether the era of “we’ll let you host our model for free and call it a partnership” is ending.
Based on the shape of this negotiation, it looks like it might be.
What It Means for Everyone Else
If you’re building on top of cloud-hosted AI, the most immediate effect is that the models you’re depending on could get meaningfully more expensive to access, depending on how these licensing costs get passed down.
If you’re watching the AI industry from the outside, this is the moment where AI model development stops looking like academic publishing and starts looking like software distribution. The platform dynamics that shaped mobile, and before that PC software, are arriving. They’re arriving fast, and they’re arriving from a direction nobody quite expected.
Moonshot AI probably doesn’t get everything it’s asking for. But it’s asking. And the cloud providers are at the table.
That’s new.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
