In the Expo Exhibition Hall of the 2026 World Artificial Intelligence Conference, large models and computing chips compete on the same stage. To make computing power worry free, the power grid itself must first become an intelligent agent. AI not only needs to collaborate with humans, but also with the power system - this is precisely another layer of deep meaning that was overlooked in the theme of intelligent partners at this year's conference.

Computing power is becoming an electricity swallowing beast. By 2025, the electricity consumption of computing power facilities will reach 170 billion kilowatt hours, a year-on-year increase of about 30%. It is expected to climb to 800 billion kilowatt hours by 2030, accounting for about 6% of the total electricity consumption in society. A ten thousand card cluster has a chip power of 7 megawatts, and the supporting cooling system requires an additional one-fifth of energy consumption. The utilization efficiency of each kilowatt hour directly determines the survival space of computing power operation.

The power grid itself must first become an intelligent agent. In 2026, the collaboration of computing and electronics will be included for the first time in the government work report. Southern Power Grid has built the first integrated carbon accounting and collaborative operation system in China, compressing the accuracy of carbon emission measurement from monthly level to minute level, and guiding the scheduling of computing tasks through electricity prices and green power signals. At the conference, SenseTime released a collaborative agent for computing, which increased the output of tokens per unit of electricity by 80% after its implementation.

The collaboration between AI and power systems has moved from concept validation to large-scale implementation. When computing power load becomes a flexible resource that can be dispatched by the power grid, tokens are no longer unconstrained consumption, but intelligent production capacity that can be optimized by the computing power collaborative system.Editor/Cheng Liting