Kimi K3 Demand Overloads GPU Capacity, New Subscriptions Halted

Following the release of the Kimi K3 model, popularity has far exceeded Moonshot's expectations. The surge in demand over the past 48 hours has pushed existing GPU computing power to near capacity limits. To avoid impacting existing users, the company has temporarily suspended new user subscriptions, prioritizing computing resources for current members. The core of this event lies not in model feature updates, but in the supply-side computing power and service carrying capacity, which has sparked widespread external attention regarding its stability.

Official Response and Membership Adjustments

Moonshot stated that the demand brought by Kimi K3 is approaching the limits of current GPU capacity. Therefore, new subscriptions have been temporarily halted and will be resumed in batches after more computing power comes online, with existing subscribers unaffected. Meanwhile, Kimi plans to split its membership into two more focused plans (including a general Kimi plan for Web/App/Work) to allocate resources more precisely. Additionally, some existing users reported being unable to upgrade their computing power packages. Alongside the tight computing capacity and soaring popularity, Moonshot is simultaneously advancing its IPO process.

User Experience and Market Reaction

Users reported trying to use K3 on the Kimi web client for two consecutive days without receiving a response, asking others how to access K3 and about API stability. Screenshots showed some tasks being paused due to system peak times, with Agent credits refunded. On the market side, Polymarket gave Moonshot a roughly 29% probability of becoming China's top AI company by the end of the month, reflecting the increased attention brought by this surge in popularity. Some commentators believe that halting new subscriptions is a more appropriate approach compared to reducing quotas for everyone or limiting access speeds. Furthermore, the low cost and high performance demonstrated by K3 have sparked heated discussions; a case showed that the model generated a 3D WebGL landing page with scrolling animations and individually modeled leaves in a single pass for only $2.37, highlighting its immense application potential.

2026-07-19 ~ 2026-07-21 · 10 related posts

Primary sources

2 near-duplicate retellings: SuggestionMission516 · mark_k