Chinese artificial intelligence startup Moonshot AI temporarily halted new subscriptions to its 2.8-trillion-parameter Kimi K3 model on Sunday, July 19, saying GPU capacity had been overwhelmed within 48 hours of the open-weight system's release.
Demand outruns compute
"Kimi K3 has received far more love than we expected, and our GPUs are feeling it," the company wrote on X, noting that traffic had "pushed close to the limits of our current capacity." Moonshot said existing subscribers would not be affected and that new sign-ups would reopen in batches as capacity is added, according to reporting from PYMNTS and Reuters. Executives split membership into a general Kimi plan and a dedicated Kimi Code plan for coding workflows, an attempt to match compute more precisely to how the model is being used.
An open-weight challenger to Claude and GPT
Launched on July 17, Kimi K3 is billed as the world's first open 3-trillion-class model, with native vision and a 1 million-token context window. Moonshot said the system "consistently outperforms other tested models" while still trailing proprietary frontier systems from Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. Its lower serving cost and open weights briefly hit US AI stocks on Monday, echoing the reaction to earlier Kimi releases.
China's IPO push meets its GPU wall
The pause lands as Moonshot is gearing up for a Hong Kong IPO. Analysts told Reuters the outage shows Chinese labs can now ship models that pressure US pricing, but that scaling inference remains hemmed in by export-controlled GPUs. Rival Chinese release Alibaba's Qwen 3.8 preview a day earlier compounded the market's attention on open-weight competition.
Reporting based on coverage from PYMNTS, Reuters, and the Associated Press.