“Within 2 to 3 years, AI compute clusters will be large enough to deploy millions of model instances simultaneously.”