Moonshot Pauses New Kimi K3 Subscriptions as GPU Capacity Maxes Out in 48 Hours

Days after Kimi K3 launched, Chinese lab Moonshot stopped selling new subscriptions, saying demand had pushed its GPU clusters to the limit — and split its plans into two tiers to ration compute.

Jul 20, 2026Source: Reuters

Moonshot AI has stopped selling new subscriptions to its Kimi assistant, telling users that demand for its freshly launched Kimi K3 model overwhelmed its computing capacity within two days. In a notice posted July 20, the Chinese lab said requests had "sharply exceeded forecasts and were approaching the limits of existing clusters."

Compute, not appetite, is the constraint

The pause is a capacity story, not a demand problem. Moonshot said that over the previous 48 hours usage had "pushed close to the limits of our current capacity," adding bluntly that "our GPUs are feeling it." Rather than let service degrade for paying users, it is redirecting spare compute to existing subscribers and will reopen new sign-ups gradually, in controlled batches.

Two tiers to ration the load

Alongside the freeze, Moonshot restructured its plans into two separate products: Kimi Membership, covering web, app and office workflows, and Kimi Code Membership, aimed at programming. The split lets the company steer heavy coding workloads — the most compute-intensive use — onto a dedicated tier while keeping general access stable for everyone else.

A model that arrived hot

Kimi K3, released July 16, is a 2.8-trillion-parameter open-weight model that vaulted into the global top three on public benchmarks. Its arrival has been read as a second "DeepSeek moment" — a competitive Chinese open model landing just as Western labs ration their own frontier systems. Moonshot's annualized recurring revenue reached roughly $300 million in June, up from about $200 million in April, and the company raised more than $2 billion earlier this year.

Export controls in the background

Serving a 2.8-trillion-parameter model takes clusters of high-end accelerators, and Moonshot — like its domestic peers — is doing so under US restrictions on advanced chips. The capacity crunch is the visible edge of that squeeze: a lab with clear commercial momentum forced to turn away paying customers because it cannot add GPUs fast enough. Weights for K3 are still slated for public release by July 27.

Why It Matters

A lab pausing revenue because it physically cannot serve demand is the clearest sign yet that the bottleneck in Chinese AI is silicon, not users. It reframes the Kimi K3 launch: the competitive threat to Western labs is real, but China's ability to scale it is capped by the same export controls meant to slow it.

Filed underAI Models Open Source AI #China AI #GPU capacity #Kimi K3 #Moonshot AI #open-weight models