Moonshot’s Kimi K3 drew so much traffic after launch that the company paused new subscriptions within two days, spotlighting the AI industry’s shift from training limits to inference capacity.
Moonshot launched Kimi K3, and demand shut down new subscriptions in 48 hours. The Chinese AI lab paused sign-ups after usage surged close to the limits of its available GPU capacity, prioritizing existing subscribers while it races to add infrastructure. The freeze is a clear signal that the industry bottleneck is moving from training frontier models to serving them at scale — especially for long-running coding and agentic workloads.
At roughly 2.8 trillion parameters, Kimi K3 is among the largest open-weight models planned for public release, with Moonshot scheduling a weight drop for July 27. On Arena.ai’s Frontend Code Arena it topped OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5; on the broader Artificial Analysis Intelligence Index it trails slightly. Strong coding scores matter because agent-style tasks keep GPUs busy far longer than one-shot chat: models continuously generate, read, and process tokens across multi-step workflows, which can re-convert lower per-token prices into higher total resource use and shift pressure onto server memory as well as raw compute.
Pricing is aggressive — about $3 per million input tokens and $15 per million output tokens, roughly 40% cheaper than Anthropic’s Opus 4.8 and about 70% cheaper than Claude Fable 5, according to Bernstein Research. That combination of capability and cost helps explain the rush of developers. Moonshot said it will reopen subscription spots in batches as capacity comes online. For Chinese AI labs, the crunch is tighter still under U.S. export controls on advanced Nvidia chips, which push greater reliance on older silicon, domestic alternatives, software efficiency, and cloud capacity from providers such as Alibaba, Tencent, and Huawei. Meanwhile, Chinese hyperscalers are pouring tens of billions into AI data centers.
For developers, the episode is an architectural warning: cheap, unlimited API access can vanish overnight when inference demand spikes. Capacity planning, multi-provider fallbacks, and efficient agent designs matter as much as model choice. Read more on The New Stack.
Original Sources
Disclaimer
This is an AI-summarized article from authoritative sources. If you want your content removed, please .
