AI training chips too concentrated to cause grid collapse

Kyle Corbitt, co-founder of AI startup OpenPipe, revealed that he communicated with a Microsoft engineer in charge of the GPT-6 training cluster project, and the latter said that it is very troublesome to configure infinite bandwidth links between GPUs in different regions. But if more than 100,000 Nvidia H100 chips are deployed in a state, it will lead to a local grid collapse.

Commentary: The concentrated development of AI computing power has caused safety hazards to the power grid, and it needs to be planned and deployed carefully.