David Gurra

And then the memory wafers, there's the big three, which are SK Hynix, Samsung, and Micron.

And then there's a whole range of providers

across the rack and manufacturing side.

One of the big things that if you want to manufacture in very large volumes, we hear about these multi-gigawatt deals that are coming out.

These require billions of dollars of manufacturing and then actually setting up hundreds of millions of dollars put into setting up supply chains in advance of delivering that.

That's a big part of what we're excited to be able to do now.

You left Google in 2022, and the goal was creating a better chip from scratch, Reiner.

But have you been impressed by the leaps that TPU has taken?

It seems to have impressed the market.

What is it that you felt wasn't at Google for you that you now can build better?

Yeah, so I think what is really required is if you want to absolutely nail the LLM workload, you have to be willing to break compatibility with previous chips.

And so one of the strong guarantees you see all of the existing players providing is you can take a program that was written on my previous generation chip or my generation of chips five years ago and it will run on my next generation chip.

And so a lot of what that means is there are constraints on, my chip has to support all of the previous number formats I supported.

It has to support all of the different programming model.

The way I communicate between cores on the chip, all of those have to be the same as each other.

We felt that it would be necessary, like if you really want to just absolutely nail this workload without regards for backwards compatibility or other workloads or anything like that, you need to,

something of a blank slate design is required.

For us, this means very large matrices, very low precision support, and then, in fact, an ability to split your very large systolic array into small pieces.

You name-checked Grok, I think, with some admiration a second ago.

I mean, when NVIDIA acquired Grok,