Moonshot AI paused new subscriptions for its Kimi K3 large-language model on July 19 [2] after surging demand exhausted the company's GPU capacity.

The move highlights the critical bottleneck of hardware availability for AI firms attempting to scale massive models. As demand for high-parameter systems grows, the gap between software ambition and physical compute power becomes a primary operational risk.

The Beijing-based startup reported that the Kimi K3 model, which features 2.8 trillion parameters [1], pushed the company's GPUs to their limit within 48 hours of launch [3]. This rapid surge led to what the company described as "unprecedented compute challenges" [4].

"We have experienced 'unprecedented compute challenges' and will temporarily focus on ensuring we can serve our existing paid users," a Moonshot AI spokesperson said [4].

To manage the load, the company paused new sign-ups and split its existing plans into two tiers [3]. Moonshot AI said it temporarily paused new subscriptions after high usage "pushed close to the limits of our current capacity" [5].

The company's growth has been rapid, with annual recurring revenue reaching $300 million [2]. However, the immediate strain on its data centers in Beijing suggests that even well-funded startups struggle to predict the exact scale of user adoption for next-generation models.

Moonshot AI representatives said the pause was necessary because demand for Kimi 3 pushed GPUs to their limit within 48 hours [3]. The firm is now prioritizing the stability of the experience for those who already hold paid accounts.

Demand for Kimi 3 pushed our GPUs to their limit within 48 hours

This incident underscores the 'compute crunch' facing the global AI industry, where the ability to deploy massive models is limited not by code, but by the physical availability of GPUs. For a company with $300 million in annual recurring revenue, the inability to onboard new users due to hardware limits demonstrates that infrastructure scaling remains the most significant hurdle for AI startups attempting to compete at the frontier level.