[llm_trw]

It's matrix multiplication being done as fast as possible. On local machines the best proxy for useful gpu load for training is power draw.

[grandma_tea]

Perhaps my question was worded poorly. I'm saying is: the premise of the article is wrong. It attributes the 29% figure to AI but the source attributes the 29% to new datacenters. It seems unlikely to me that 100% of new datacenter usage would go to AI.