With agentic [AI] you're doing multiple actions, multiple calls to APIs, opening up applications like an Excel or a Bloomberg or a CNBC website or SEC website, etc. and doing all of these things. So you need an orchestration engine. You need something that goes ahead and organizes all of that. That's what CPUs do very well. If you go back to training, you're just doing the same thing over and over again where you are training these AIs. That's what a GPU does incredibly well, and that's why you saw an Nvidia absolutely explode higher because CPUs weren't really valuable in training. You move to chat-based AI, again, that's pretty decent for a GPU, but now when you move to agentic, that's great for a CPU. So you went originally from about 8 GPUs for 1 CPU. When you move to agentic, that ratio gets closer to 1-to-1.
[...]
Agentic [AI] requires 10 to 100 times more tokens than chat-based AI.
~ Dan Niles, "Dan Niles: Be Nimble - 30-50% AI Crash By 2027," The Master Investor Podcast with Wilfred Frost, 11:40 mark, May 12, 2026
No comments:
Post a Comment