Moonshot AI has suspended new consumer subscriptions to Kimi K3 after demand overwhelmed its computing infrastructure, less than a week after the model’s release. In a statement, the company said that “over the past 48 hours, demand has pushed close to the limits of our current capacity,” adding that user requests had sharply exceeded forecasts. Available computing power will go to current paying subscribers, who Moonshot says are unaffected by the shortage.
The pause covers new consumer signups for Kimi K3, which became the largest open-weight model ever released when it launched on July 16. When signups reopen, Moonshot plans to split memberships into two plans, including one dedicated to coding, so compute can be allocated by use case rather than served from one pool.
The capacity crunch arrives at a sensitive moment for the company. Moonshot is seeking up to $2 billion in fresh capital after its valuation reached $30 billion in June, and is preparing for a potential Hong Kong initial public offering, holding discussions with Goldman Sachs and China International Capital Corp. It has begun unwinding its offshore holding structure ahead of a listing. A subscription pause driven by excess demand is a usable story for that process, and a genuine operational problem at the same time.
It is also the clearest demonstration yet of the constraint that sits under China’s model boom. K3 posted near-frontier benchmark results at a fraction of Western prices, but serving a 2.8-trillion-parameter model is expensive: Moonshot itself recommends clusters of at least 64 accelerators for efficient inference. US export controls limit the company’s access to exactly that class of hardware, and the White House alleged on July 22 that Moonshot has been routing around those limits with Nvidia GB300 servers deployed in Thailand. Domestic rivals face the same squeeze from a different direction: Z.AI answered its own blacklisting by building a gigawatt-scale data center on Chinese chips.
The pause complicates the cheap-frontier-AI narrative that took hold after K3’s launch. The model’s aggressive pricing helped trigger a selloff in US chip and AI stocks on the theory that frontier capability was being commoditized. Prices only commoditize what can actually be served, and Moonshot is currently turning away paying customers because it lacks the compute to take their money. The full model weights are still promised for July 27, which would let enterprises run K3 on their own infrastructure, moving the serving problem, and the hardware bill, onto the customer.
Sources: South China Morning Post, Reuters via The Globe and Mail, The Standard
–
By the Control Plane Editorial Team