Moonshot AI Wants 30% of Hyperscaler Revenue to Host Kimi K3 — and It Might Have the Leverage

The most audacious licensing negotiation in AI is happening right now, and it runs in the opposite direction from everything the industry is used to. According to a Reuters exclusive published August 26, China’s Moonshot AI is in early talks with Microsoft, Amazon, and Google over agreements to host its Kimi K3 model on Azure, AWS, and Google Cloud — and the Beijing-based startup is seeking up to 30% of the revenue generated from K3-related services on those platforms.

If any of these deals close, it would be the first major revenue-sharing arrangement between a Chinese AI lab and US hyperscalers — and it would formalize a business model that turns the open-source playbook on its head: give the weights away for free, then charge the world’s richest technology companies a toll to serve them.

What’s on the table

Reuters’ sources describe the discussions as early-stage, with Moonshot demanding as much as 30% of revenue from K3-powered services running on the three American cloud platforms. The figure is identical to the cut Moonshot already imposes on large commercial users under its own K3 license — the terms that made headlines in July when the model shipped.

The timing is not subtle. Moonshot is IPO-bound: Bloomberg reported on July 29 that the company closed a $3.5 billion funding round at a $35 billion valuation, exceeding its original target, and it has told investors it is preparing a Hong Kong listing that could arrive within six months. A signed distribution deal with even one US hyperscaler would be the kind of anchor revenue and legitimacy that underwriters dream of. The company’s annualized revenue had reportedly reached roughly $300 million even before K3’s arrival supercharged its consumer business.

For the hyperscalers, the calculus is equally naked. Cloud customers increasingly want the best frontier models available, and K3 — with 2.8 trillion parameters, a one-million-token context window, and native vision — has benchmarked alongside the leading American systems, trailing only Anthropic’s best on several evaluations when it launched. Hosting it would sell GPU hours. The question is at what margin.

The model that started a market correction

To understand why Moonshot believes it can charge Microsoft a toll, rewind to July 16. That day, Moonshot released Kimi K3 as a hosted service, billing it as the world’s first open 3T-class model, with full weights following on July 27. On benchmarks it matched leading US models — trailing only Anthropic’s frontier system in several suites — at a fraction of the serving cost of its American rivals.

The market reaction was immediate and violent. Within days, semiconductor stocks shed a cumulative $3.3 trillion in market value, as investors confronted the possibility that frontier-quality AI could be served cheaply on open weights rather than exclusively through expensive proprietary stacks. The selloff recovered, but the point was made: open-weight models from China had become price-setting forces in the global AI market, not curiosities.

That leverage is exactly what Moonshot is now trying to monetize. An open-weight model cannot be locked behind an API, but it can be made legally expensive for the largest distributors — and the big three clouds are, collectively, the largest potential distributors on Earth.

The Alibaba precedent

Moonshot is not alone in redrawing the boundaries of “open.” On August 7, Reuters reported that Alibaba is preparing to require large commercial users of its next open-weight flagship, Qwen3.8-Max, to share revenue at rates that could also reach 30%. The two Chinese giants have effectively converged on the same doctrine: open weights for researchers, hobbyists, and small companies; a commercial toll booth for everyone else.

This is a genuinely new species of licensing. Traditional open source — the GPL, MIT, Apache world — constrains what you must give back, never what you must pay. The Chinese labs’ approach instead treats openness as a distribution strategy: maximal adoption first, monetization of scale later. Critics argue this makes “open-source AI” a marketing term. Defenders counter that a 30% cut charged only to trillion-dollar platforms, with free access for everyone below that threshold, is a more honest bargain than closed APIs ever offered.

Either way, the definition is now being negotiated deal by deal — and the Moonshot-hyperscaler talks are the highest-stakes test yet.

Why the hyperscalers might actually say yes

Thirty percent of revenue is a painful ask by any standard — app stores, the classic platform tax, take 15–30% of developers’ revenue, not the other way around. But three factors make this negotiation less lopsided than it looks.

First, demand-side pull. Enterprise customers evaluating K3 will run it somewhere. If Azure offers it and AWS doesn’t, Microsoft captures that workload’s compute revenue — GPU hours, storage, networking — which is the actual product the clouds sell. A 30% share of model revenue in exchange for 100% of a growing workload’s infrastructure spend can be rational arithmetic.

Second, competitive dynamics. All three clouds host each other’s rivals’ models already — Azure serves OpenAI’s competitors, Bedrock aggregates dozens of labs. No one wants to be the platform that sat out a frontier-class Chinese model while its rivals bundled it.

Third, the IPO clock. Moonshot needs at least one signed deal more than it needs the full 30%, which suggests the number is an opening position rather than a walk-away floor.

The open questions are the hard ones: whether US regulators will tolerate American clouds deeply commercializing a Chinese frontier model amid ongoing technology restrictions, and whether serving K3’s 2.8-trillion-parameter architecture at scale is even economical for clouds that didn’t train it. Early benchmarks suggest inference efficiency is competitive with the best proprietary systems, but hyperscaler engineering teams would be optimizing someone else’s giant.

What to watch

The negotiations are early and could still dissolve — Reuters’ sourcing carefully notes that all three companies may simply decline. But the direction of travel is clear. Open-weight frontier models from China are becoming distribution assets with real pricing power, their creators are heading to public markets with billions in fresh capital, and the American clouds that once regarded open-source AI as a commodity sideshow are now negotiating to pay for the privilege of serving it.

Whoever blinks first will set the template for how the next decade of AI distribution economics works.