There will never be enough GPUs! The Kimi K3 circuit breaker is just the beginning: every generation of stronger models in the future will repeat this.

On July 27, according to CCTV News, GPUs are once again in short supply. As AI applications rapidly penetrate various sectors, computing power shortages are becoming a norm in the industry. Within 48 hours of the launch of Kimi K3, Moonshot AI announced the suspension of new user membership subscriptions due to user request volumes exceeding expectations.

Some users have noted that not only do they need to pay for using AI, but membership access is also starting to feel like grabbing concert tickets—hard to secure. Professor Shen Yang's team at Tsinghua University analyzed that this is the Jevons paradox of the AI era, where increased efficiency actually unleashes more demand, leading to continued growth in overall computing power consumption.

Industry insiders point out that Kimi K3's computing power overload is just the beginning. In the future, each generation of more efficient large models may repeatedly follow this expansion logic. Computing power shortages are not a short-term phenomenon but a structural challenge for the AI industry.

Li Kaimeng, Deputy General Manager of Inner Mongolia Zhongke Supercomputing, stated that the company's current data center is already 80% full, with computing power reaching about 8,000P. Building data centers, purchasing chips, and ensuring power supply are the three thresholds that must be crossed for computing power expansion.

He noted that inference computing now accounts for 90% of the company's business, with training only 10% remaining. At the same time, computing power leasing prices have been rising. Monthly rents for some GPU devices have increased from 8,000 to 9,000 yuan last year to about 12,000 yuan, an increase of over 30%. Prices for server components such as CPUs and memory have also risen significantly.

Equipment procurement is only the first half of computing power expansion. Once the hardware is in place, energy supply becomes the key factor determining the upper limit of scale.

Public data shows that in 2025, the electricity consumption of national computing centers reached 170 billion kWh, accounting for 1.6% of total social electricity consumption. The electricity consumption growth rate of the eight major hub nodes is about 39.5%. To actually run chips and data centers, round-the-clock stable power supply is essential.