The issue has shifted from model capabilities to usage efficiency. Many agent invocation methods turn a simple request into multi-round calls + super long context, rapidly amplifying computational power consumption. This also explains why Claude has started tightening third-party access. Moving forward, agents need to focus on solving a few key things: fewer calls, more precise calls, and more efficient cache utilization. Otherwise, it’s just continuous consumption of computational power, and the growth of computational power itself has an upper limit. The key to future competition lies in who can use models more efficiently. Also, now I understand why Musk wants to build power stations in space. Musk approaches problems from first principles. I used to think he was just painting a pie-in-the-sky vision for humanity, but after the large-scale explosion of agents, people are realizing the severe energy limitations and massive power shortages. Many are buying hydroelectric stations, and if Musk’s space power station succeeds, it’ll be like NVIDIA in the AI era—selling shovels. He’ll become the guy charging Earth for electricity.