A Harvard professor breaks down GPU competition: AMD is Nvidia's only real rival, but Google's TPUs and cloud providers are building cheaper alternatives for specific tasks. Training AI models still needs Nvidia's power, but inference can run on cheaper hardware, raising questions about whether demand will hold.
Chapters
And paying attention to a lot of different players, because it's not just amd. AMD is the only company that really has a GPU competitor that can compete directly with Nvidia on the core technology. But there are two other things we have to be thinking about. One is that there are a variety of other types of accelerators, whether we're talking about TPUs at Google or some of the other capabilities that Amazon and Microsoft are trying to develop that would allow them to do some of the same things in a more narrowly focused way, but without the expense of a Blackwell gpu. In addition, the big question down the road is how much of this capacity that's being put into place is going to be used for training, which requires a gpu, or is going to be used for inference. And for inference, you don't necessarily need a high powered gpu. You can use a CPU and an accelerator, which is a fraction of the cost. So down the road the question is, will that same demand for Nvidia GPUs that we see today be sustained, or will we start to see a tail off.