Name: Nvidia CEO Huang: DeepSeek incident underscored the substantial demand for AI compute power
Uploaded: 2025-02-27T07:58:35.000Z
Duration: 5 min 14 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

The guide was also above, as you mentioned.

Participle clauses

Now, earlier today, speaking of demand, I was talking to Amazon CEO Andy Jassy.

The second conditional

He told me that as of now, if he had more AI resources to sell through AWS, he could sell more.

That's kind of the short-term signal of demand that you talked about on the call.

Tell me more about the mid-term signals that investors should be aware of that give you confidence in the continued demand.

The scale-outs of data centers, AI factories, relative to what you've historically seen.

The short-term signal are just our POs and the forecasts.

And on top of that, the things that are not forecasted are new startup companies that are spinning off.

And at the risk of forgetting any of them, I won't mention any of them, but there's some really, really fantastic startups that have come out as a result of new reasoning AI capabilities and artificial general intelligence capabilities that they have breakthroughs in.

And several of them, there's several of them that are related to agentic AIs, really exciting companies.

And there's several of them related to physical AIs.

There's just handfuls of each one of them, and each one of them needs additional compute.

And that's the type of things that Andy talks about, because they need to go to AWS and they have urgent need for more compute right away.

And so that's on top of what we already knew to have POs and forecasts and such.

The midterm comes from the fact that this year's capital investment for data centers is so much greater than last year's.

And of course, we had a very large year last year.

It stands to reason that with Blackwell and with all the new data centers going online, we're gonna have a fairly great year.

Now long-term, the thing that's really exciting is we're just at the beginning of the reasoning AI era.

This is the time when AI is thinking to itself before it answers a question, instead of just immediately generating an answer.

They'll reason about it, maybe break it down step by step.

It'll do maybe some searching in its own mind before it creates and composes a smart answer for you.

The amount of computation necessary to do that reasoning process is 100 times more than what we used to do.

So if you could imagine, we thought computation, the amount of compute necessary, was a lot last year.

And then all of a sudden, reasoning AI, DeepSeq was an example of that, ChatGPT 4.0 is an example of that, Grok 3 Reasoning's an example of that.

So all of these reasoning AI models now need a lot more compute than what we used to, used to, used to, we're expecting.

Because some people took DeepSeq to mean actually that you need less compute, right?

Because the initial report was that they were doing more with less.

But you're saying, in fact, some of what came out of DeepSeq was the opposite, that there's gonna be more compute demanded.

There are three phases in how AI works, how AI is developed largely.

It's kind of like us going through high school.

A lot of basic math, basic language, basic everything.

That basic understanding of human knowledge is essential to do what is the next step, which is called post-training.

In post-training, you might get human feedback.

You know, it's like a teacher showing it to you.

We call it reinforcement learning human feedback.

You might practice and do thought experiments.

We call it reinforcement learning AI feedback.

You could either also do tests and practice, and we call it reinforcement learning verifiable reward feedback.

So now, basically, it's teaching AIs how to be better AIs.

That post-training process is where an enormous amount of innovation is happening right now.

A lot of it happened with these reasoning models, and that computation load could be 100 times more than pre-training.

And then here comes inference, the reasoning process.

Instead of just spewing out an answer, when prompted, it reasons about it.

It thinks about how best to answer that question, breaks it down step by step, might even reflect upon it, come up with several versions, pick the best one, and then presents it to you.

So the amount of computation that we have to do even at inference time now is 100 times more than what we used to do when ChatGPT first came out.

And so all of a sudden, the combination of all these ideas largely related to reinforcement learning and synthetic data generation and reasoning, all of this is just causing compute demand to go sky high.