Moonshot AI Freezes New Kimi K3 Subscriptions and Cites Unprecedented Computational Challenges
Three days after launching the 2.8 trillion parameter model, Chinese company Moonshot suspended new users due to GPU limits. Current subscribers remain unaffected as the company moves toward a listing in Hong Kong.
Three days after bringing Kimi K3 online, Chinese company Moonshot AI announced on Sunday, July 19, that it has temporarily closed new subscriptions for the model. In a statement published on the company's official channels, Moonshot stated that it is facing "unprecedented computational challenges" after the volume of requests in the first 48 hours approached the ceiling of available GPU capacity. Current paid subscribers were not affected. The company indicated that it will reopen spots in batches as it manages to free up resources, without a defined timeline. The API for enterprise developers remains available, but with aggressive throttling.
What is the Kimi K3 Model?
Launched on July 16, the Kimi K3 is a model with 2.8 trillion total parameters in a mixture-of-experts architecture, featuring 896 experts, of which 16 activate per token. This results in a cost per token equivalent to a dense model of 30 to 50 billion parameters, which explains the aggressive pricing: $3 per million fresh input tokens, $15 per million output tokens, and $0.30 per million cached input tokens. The context window extends to 1 million tokens, and the model has native vision capabilities.
In the Artificial Analysis ranking, the K3 placed fourth among leading models, behind Claude Fable 5 and GPT-5.6 Sol, and ahead of Claude Opus 4.8. For an inaugural launch, this position led several data architecture teams in the U.S. and Europe to reconsider the code purchase pipeline for the latter half of the year. The publication The Decoder referred to the K3 as "the end of ultra-cheap Chinese AI": Moonshot had been offering rates 30% to 50% lower than Western peers in 2025 and adjusted to be comparable to Anthropic's Sonnet 5.
GPU Limits, IPO Approaches
The issue behind the freeze is physical. The American restrictions on the sale of high-density Nvidia accelerators were only relaxed on July 8, when Washington authorized the release of the H200 for the Chinese market. Prior to this, Moonshot primarily relied on the H20, a reduced-capacity version, and limited allocations within Alibaba Cloud, its main sponsor and strategic investor. With the K3 generating volume organically, the ceiling became apparent quickly.
The timing is sensitive. Moonshot AI is in the preparatory phase for a listing in Hong Kong, with a valuation that could reach $4 billion, according to a July 15 report from the Financial Times. The viral success of the K3 has squeezed the company between commercial demand and the need to present a scalability narrative in discussions with institutional investors. Refusing customers, even for 72 hours, risks undermining that narrative.
Two Geographic Readings
The impact of the pause feels different in each market.
In the United States, the freeze is an involuntary showcase for Anthropic and OpenAI. Data architecture teams in banks and Big Tech that were evaluating replacing Sonnet 5 instances with Kimi K3 in text analysis pipelines returned to internal meetings this week. It is unlikely they will reverse their interest, but Sam Altman gains fresh stability arguments in the annual renewal cycle, and Anthropic reinforces the premise that "a cheaper model is only valuable if it is actually available."
In South Korea, the effect is the opposite. The KOSPI fell 4.5% in the sessions following the launch of Kimi K3, with investors interpreting the model's success as a sign of reduced demand for HBM3E from Samsung and SK Hynix. The pause in subscribers, counterintuitively, detracts from that thesis: if Moonshot needs more chips in the short term, the Nvidia-SK Hynix-Samsung corridor becomes a bottleneck again, not a victim. For the CIO who buys memory and cloud under a three-year contract, the lesson is always the same: infrastructure prices cycles that the popular model later capitalizes.
One metric that weakens the Moonshot narrative: the K3 was trained with preferred credits from Alibaba Cloud, and reproducing equivalent capex in private cloud would cost hundreds of millions of dollars, according to estimates from SemiAnalysis. Cost democratization is on its way, but it depends on who pays for the subsidy this time.