Moonshot Halts New Kimi K3 Subscriptions as Demand Swamps Capacity Days After Launch

Moonshot AI has suspended new subscriptions for Kimi K3, the Chinese flagship model that triggered what analysts called a fresh "DeepSeek moment" for markets, after a flood of demand overwhelmed the company's serving capacity within days of launch.

Victim of its own launch

K3, released on July 16, is a new-architecture mixture-of-experts model with roughly 2.8 trillion total parameters and a 1-million-token context window, aimed squarely at long-horizon coding and agentic workloads. Its benchmark showing against frontier US models at a fraction of the price sent Kimi Code sign-ups vertical — and Moonshot's inference fleet could not keep up.

The company said existing subscribers retain access and that it is racing to add capacity, but new paid sign-ups are paused indefinitely. For a lab that was fighting for relevance against DeepSeek and Alibaba's Qwen less than a year ago, it is a strange kind of crisis: too many customers.

The July 27 wildcard

The suspension raises the stakes for Moonshot's commitment to publish K3's weights on July 27 under a modified MIT license. With first-party access rationed, the open-weight release becomes the primary way for developers worldwide to get their hands on the model — via third-party inference providers, sovereign clouds and self-hosted deployments that don't depend on Moonshot's own GPUs.

That timing collides with DeepSeek's V4 stable release, expected July 24, setting up a single week in which two frontier-class Chinese open models hit the market while US labs keep their best systems closed.

Compute is the constraint

The episode is also a reminder of the binding constraint on China's AI surge: inference compute. Export controls have made serving capacity, not model quality, the bottleneck — one reason Beijing is pushing gigawatt-scale domestic-chip data centers like the one Zhipu just completed, and why Washington is now debating whether Chinese open weights themselves should be restricted. Moonshot's overloaded servers are, in their own way, the strongest evidence yet of how much global demand those models command.