Open weights, 30% cut: Moonshot's Kimi K3 ask puts a royalty inside cloud inference pricing

Thirty percent of a sale price is the number that decides whether managed K3 sits below or above the proprietary endpoint listed next to it in the same catalog ^[1]. A provider that agrees keeps at most seventy cents of every dollar billed, and out of that has to fund the hardware, the customer relationship, the security controls and the billing system it brought to the arrangement ^[15][11]. Whatever the split settles at, it settles inside the per-token price.

The escape hatch is narrower than the phrase "open weights" implies. K3 is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion activated, roughly 3.7 percent of the network per token ^[4][14]. Sparse activation trims the arithmetic each token requires; it does not make a 2.8 trillion parameter system something a platform team stands up next to its existing services, and the runtimewire account, which credits Reuters, says few customers have either the infrastructure or the appetite to try ^[5]. That gap is the whole business. The three US providers already hold the enterprise contracts, the metering and the capacity that turn downloads into recurring spend ^[10].

The sequencing shows the intent plainly enough: K3 launched on July 16 and its weights followed on July 27, eleven days later ^[13][16]. Availability first, then a toll on the channel where the volume actually lands.

Collection is the load-bearing part, and it is also the contested part. Moonshot needs enough visibility to verify usage and compute what it is owed, while cloud customers may resist giving a Chinese model provider broad access to operational data, according to the same account ^[8]. Auditing token usage is one of three items the reporting lists as unresolved, alongside data access and the split itself ^[2]. So the mechanism that makes the deal collectable is the one enterprise buyers have the strongest reason to refuse, and it is most likely to be refused by exactly the regulated accounts worth billing.

One thing the reported list does not include is licensing or enforcement ^[12]. If the weights are downloadable and modifiable, what brings a hyperscaler to the table is distribution economics and the wish for a catalog entry it did not have to train, not a legal obligation the reporting establishes ^[11]. That is a weaker position than a proprietary vendor's, and the 30 percent request is priced as though it were stronger.

The capability picture does not help the ask either. Moonshot's own K3 report says the model still trails the strongest proprietary systems on the authors' evaluations ^[9]. The pitch to a cloud is therefore cost and audience rather than frontier performance, and a levy of this size works directly against the cost half of that argument.

What keeps the number from reading as an opening bluff is that the pattern already exists at smaller scale. Reuters reported the 30 percent is a ceiling consistent with terms Moonshot has put to other large customers, the lab has similar arrangements with smaller cloud platforms, and Chinasoft International disclosed a revenue-sharing and product collaboration in July without publishing the financials ^[3][6][7]. Teams that assumed an open-weight model meant a cheap hosted endpoint should read the ask as a template rather than an anomaly.