Just Three Days After Launch, Kimi K3 Hits the Brakes as Surging User Demand Pushes Moonshot AI's Computing Capacity to Its Limits
Moonshot AI announced that it has temporarily suspended new consumer (C-end) subscriptions after demand for Kimi K3 surged far beyond expectations shortly after its launch. The company said it is prioritizing service quality for existing paying users while expanding its computing capacity, and plans to gradually reopen subscriptions to new users once additional resources are in place.
Just three days after Kimi K3 went live, Moonshot AI has hit pause on new consumer subscriptions. On July 19, Kimi announced that it would temporarily stop accepting new C-end subscribers in order to protect the experience of users who are already paying for the service. For now, the company says its available computing resources will be directed first toward existing subscribers.
The decision followed a surge in demand that arrived faster than expected. Since Kimi K3 launched on July 16, the model has attracted unusually strong user interest, and in the most recent 48-hour window, request volume climbed well beyond prior estimates. According to the official statement, traffic has moved close to the upper limit of what the current compute cluster can handle.
Until additional capacity is brought online, Kimi will keep new consumer subscription slots closed. Moonshot AI said it is accelerating infrastructure expansion and plans to reopen access in stages as fresh computing resources become available, with a full return to normal enrollment to follow later.
Subscription access is being adjusted along with capacity planning
At the same time it announced the subscription pause, Kimi also revealed a change in how future subscription benefits will be structured. Going forward, the company plans to separate its core Kimi membership benefits from Kimi Code-related benefits for new subscribers.
Kimi core benefits will cover Kimi Web, Kimi App, and Kimi Work.
Kimi Code benefits will be configured independently rather than bundled into the main package.
Kimi said the split is intended to better align compute allocation with actual product usage and to preserve service quality across different scenarios. The move also signals that programming functions, which typically consume more resources, will likely be managed and costed separately in the future.
Kimi K3 is positioned as Moonshot AI’s most capable model so far
Released on July 16, Kimi K3 is currently described as Moonshot AI’s strongest large model. The company says it features 2.8 trillion parameters and supports a context window of up to 1 million tokens. Its main use cases are long-horizon coding tasks and end-to-end knowledge work.
Information published on Kimi’s open platform shows that K3 uses a usage-based pricing model:
Input pricing: RMB 2 per 1 million tokens when cache hits apply, or RMB 20 per 1 million tokens when they do not.
Output pricing: RMB 100 per 1 million tokens.
Strong demand is also showing up in commercial performance
The surge around Kimi K3 has not been limited to consumer attention. On the business side, the product appears to be driving immediate results as well. On July 18, Kimi President Zhang Yutong said the company would open K3 model weights in the coming days.
She also disclosed that, based on internal data, the company recorded the largest single-day increase in annual recurring revenue in its history on July 17, around the time of the K3 release. That suggests the launch not only brought in a wave of user traffic, but also delivered a notable lift to monetization metrics.
At the same time, the rush of real-world usage has exposed the pressure sitting beneath the surface: infrastructure capacity. The popularity of a model can translate into business momentum quickly, but it can also test whether backend compute systems are ready to support that growth at scale.
The next stage of large-model competition is increasingly about infrastructure
Kimi K3’s rapid move from launch to subscription suspension highlights two realities at once. First, demand for high-performing AI models remains intense. Second, explosive user adoption can strain service delivery even for well-funded platforms.
As large models continue to improve and user enthusiasm converts into heavy, sustained usage, the industry’s challenge is no longer just about building bigger models with more parameters. It is also about whether companies can expand computing capacity fast enough to keep those products reliably available.
In that sense, Kimi K3’s first few days on the market may be remembered not only for the attention it generated, but also for what it revealed: in the current AI race, compute availability is becoming just as critical as model capability.