[adam_arthur]

Model inference can be done with any GPU, AMD, Nvidia, Intel. Can even be done with CPU in many cases.

Training is largely done on Nvidia cards, but there's nothing mandating that.

Google trained Gemini on their own in-house TPUs, and according to their published stats it exceeds ChatGPTs performance.

There's a collective delusion right now that somehow CUDA entitles Nvidia to a forever monopoly on GPU compute. There's no way they will maintain ~90% gross margins on their hardware sales. It's far too economically inefficient for purchasers in the long run.

The ones who figure out how to use competitor cards at less than half the cost will have a huge advantage

Right now it's fevered money pouring into what they see as the fastest way to get their feet into the game. Nvidia will do well, but that's already more than priced in right now.

[versteegen]

Their gross margin for Q4 is 76% (was 63% a year ago).

[adam_arthur]

Across all products. Margin is likely higher for the newer product lines.

But anyway, good to pull an exact figure!