“Within 2 to 3 years, AI compute clusters will be large enough to deploy millions of instances of a single model, with each instance operating faster than a human.”