YouTube Distilled
HomeVideos

AMD CEO Lisa Su: for the first time, more AI compute is being used to run models than to train them

Watch on YouTube

Now Playing

AMD CEO Lisa Su: for the first time, more AI compute is being used to run models than to train them

Transcript

13 segments

0:00

We can say that in 2026, for the first

0:02

time, the world is using more AI compute

0:06

to run models than to train them.

0:08

Compute demand is growing at an

0:10

incredible pace. We're actually seeing a

0:13

step change in comput demand because

0:16

when you ask the agent to do something,

0:18

it actually has dozens of steps and it

0:21

has to reason and it has to call tools

0:23

and it has to access data and it has to

0:25

keep doing it over and over until it

0:27

solved the problem. And so you need lots

0:29

of GPUs to do all that reasoning.

Interactive Summary

In 2026, the world will reach a significant milestone where AI compute used for inference and running models surpasses the compute required for training. This shift is driven by the increasing demand for agents to perform multi-step reasoning, tool usage, and data access, which requires extensive GPU resources.

Suggested questions

2 ready-made prompts