THIS WEEK IN AI: ChatGPT Built its Own Chip, Micron Soars, a24's $75m Google Deal

**Ejaaz** (0:00)
OpenAI announced they're building their very own AI chip, custom designed for OpenAI models, Codex, ChatGPT. But the craziest part was, it was designed by their AI models themselves. Everything from the software stack to potentially even the hardware design itself was made by OpenAI's internal model, as well as Codex, their coding model. It's going live in nine months, which is absolutely crazy. Typically, these things take one to two years to at least design and build the first couple of prototypes. So the fact that OpenAI is entering this market is the first real example of an AI lab owning the entire stack, going from hardware all the way to frontier AI models. Now, in other news, we've seen a few rumors. Fable 5, the most powerful model from Anthropic, might be coming back a lot more on that, and Micron absolutely killed their earnings report, blowing every bare expectation out of the water and expecting to make more revenue than NVIDIA did in Q3 of last year. A lot of stuff to go through today.

**Josh** (0:54)
Spicy little Jalepeño, look at that. This is exciting. It's fun to see the OpenAI team move into, I guess, enterprise hardware is what we would call this. They're making their own chips. And who was the last company that made their own chips? We have Google, who has their TPUs, incredible. Amazon has Tranium, and now OpenAI has their own Jalepeño chips. And it's important to note that they're going the same route as Amazon and as Google. This isn't a GPU, it's an ASIC. It is a chip specifically designed for a very narrow task instead of something general purpose. Traditionally, when OpenAI was to use NVIDIA's GPUs, the bottleneck isn't actually the compute, but instead it's the memory that sits on the chip. So at any given time, they're paying for 100% of the GPU, but they may be using 70 to 80% of the capacity because they're limited by that memory bandwidth. What these accelerators are going to do with this ASIC chip, whatever this Jalepeño chip is, the idea is that it's going to be used for inference. It is going to be purpose built to supply AI tokens very, very fast and to be fully utilized. And that utilization, when vertically integrated, creates magical things. We always talk about this on the show, but Apple and their M series chips and how it was such a step function improvement in every single aspect of the devices. It was like a day and night difference overnight when they released this chip. There's an opportunity for OpenAI to do that now because owning the SAC allows you to squeeze out so much more efficiency than you have anywhere else. And it's very exciting to see them build a chip that looks something like this.
I liked the idea that they used ChatGPT to actually accelerate the creation of this chip. I think it was nine months, which is pretty damn fast. And you have to assume that these two things are going to exist in parallel. Right? It's like as ChatGPT gets better, it is able to help with the tape out and the design of this chip. As the chip gets better, it's able to process tokens more efficiently and better. And it's this like kind of cycle that this flywheel that they're starting. And I think this is step one. I mean, it's very exciting to see them getting in the game.

**Ejaaz** (2:47)
I think the entire story for this has got nothing to do with the chip, and more so that the models were used to design the chip. So the way I see it is, the most scarce resource that every single AI lab that's fighting out for the number one position has right now is compute. And where you apply that compute defines whether you're going to win the race or not. Now, for the last nine months or for the last year at least, Anthropic has made it very clear that it's coding. So use the compute to train a better model that's better at coding. If you own the best coding model, it can build pretty much the entire software stack. Now, emphasis on the software stack. What good is AGI or an AGI-like model if it runs on suboptimal hardware? So seeing OpenAI make this move, realizing that they've used all their compute to train a model that can then build or design better hardware that then potentially makes sure that they spend less money to run the same level of intelligence means that they have more compute in the future to build whatever model that they want. And it runs on optimal inference or whatever hardware stack that they build. So this, to me, is that kind of mind-blowing moment that I had a year ago when Claude Code went live, where I was like, oh my god, coding is the future. OpenAI has strategically made a very cool decision here, where they're like, no, we want to own the hardware as well. And we don't want to rely on NVIDIA. And we're going to partner up with Broadcom. We're going to partner up with MediaTek, which, by the way, these are companies. Broadcom is publicly traded. So these are companies that could potentially rival NVIDIA in the future to build their own chip. And the fact that they're releasing it by the end of this year is just a sign of the times that OpenAI, that doesn't have any scaling hardware experience, has been able to accelerate this using their own model. So it's the first real proof that you can do this, not just on the software scale, but on the hardware scale. It's just very cool.