Mandeep Singh

And actually, there's a lot of talk about consumers being cautious.

So I'm curious to see how it plays out.

There's a lot of mixed signals going into this holiday shopping season, even more so than a typical one.

Yeah, it was a little bit of a play with that with David and Goliath.

I mean, they are really going after data center capacity right now.

And the way they are doing it is by diversifying their supplier base.

So it's not just relying on NVIDIA, which everyone does right now for compute, but really leveraging Broadcom, which is a custom silicon maker.

So think about NVIDIA giving you a generic chip where you can run your AI workloads, whether it's training or inferencing.

custom silicon is used just for uh you know the specific workload that open ai has to run for its proprietary model so no one else has any benefit of using a custom silicon because open ai is not looking to sell its own chips to compete with nvidia it's looking to use its chips for its own chat gpt app or any other custom app that it has developed in-house

And Google is a prime example of what a custom silicon looks like because they have their own TPUs, which when you compare it to NVIDIA GPUs is more customized in nature, but it does a terrific job of running YouTube or any other AI workloads that Google wants to run on its chips.

So that's what OpenAI is doing.

And it has a tremendous cost advantage because it costs a lot lower than the NVIDIA price tag of $30,000 on average for a GPU.

I just want to make sure I understand.

Yeah, I mean, look, one gigawatt requires up to 500 to 600,000 accelerator chips.

So we're talking 0.5 to 0.6 million chips for one gigawatt data center.

Imagine if you can save up to $5,000, how it multiplies in terms of cost savings.

The real constraint right now is power.

It's not as if you get a cheaper chip and you are all good.

You still need the performance per watt, which is why NVIDIA is so good, because it gives you 5 to 10x more performance per watt than the nearest competitor.

Exactly.