Masters in Business · Friday, September 18, 2026
Glen Kacher explains AI compute involves training models using powerful chips like NVIDIA's GPUs, while inference uses these or specialized chips to execute tasks. He sees both as investable themes, with inference potentially becoming a larger market than training.
“Sure. So the training compute or the chips, the AI accelerator chips and Today, NVIDIA dominates that still with their graphics processor chips. And that, you know, those chips originally were made for gaming, for doing very rapid mathematics that have to do with calculating physics and lighting, shading in video games. It turns out that the same kind of mathematics are incredibly well positioned to do the math around AI. And then, so you're training a model, an AI model that will be able to make judgments.”
“And then when you're actually using that model to ask questions and, you know, or have it solve problems and actually execute those problems, that's called inferencing, right? And so inferencing can be done on a more simple chip. So people have kind of used a phrase XPU to X out the graphics and let's say this is the next generation of chips that can be used to actually solve the problems with those models that are built.”