Victor Peng0:09
Yeah, AI chips are designed purposely to accelerate AI and there are different forms of AI chips. I think what most people think about is GPUs, which are used in the data center primarily when people think about training large language models, foundational models. And they're purpose built for that. So they're different than a chip like a CPU, which is a more general purpose jack-of-all-trades, if you will, that runs many different applications. And you can run AI on it, but it would not perform nearly as well as a purpose-built chip, like a GPU that accelerates AI.
So we, like most semiconductor companies have the faba model, and what that means is we do the design and we take responsibility for the overall product. However, we outsource the manufacturing to very specialty companies that are called foundries, okay? And we work very closely with our foundry partners, and we have a great relationship so and that includes, if there's concern around geographic concentration, we do work with them when they have different sites as alternatives. And so that's the main way we mitigate any concern about a specific location. And of course, we do look at what the state of the art from other suppliers. We've had an excellent relationship with our current supplier, but if it made sense, we could look at other opportunities and options as well.
Let me first explain the AI in terms of the full stack, because one of the things about AI that's so interesting is that it's a pretty broad stack from a software perspective, and also is a pretty broad stack from a hardware perspective, and you need both to interoperate well. And also, there's many markets, some of which are in the early stages of adoption, some of which are going quite vigorously already, even though it's still relatively early, right? So in that regard, let me focus on the data center. AMD plays in the data center both in data center GPUs, our MI instinct products, as well as our server, epic CPU servers. We also play in clients even embedded. But again, let me just focus on the data center right now, and that's whether it's in a public cloud or private enterprise deployment on prem and so forth. Again, I think we're participating. We have a very strong product and inference. We also are quite competitive on training. But we're relatively compared to the competition, penetrating earlier. And we expect a big ramp, in terms of what could be choke points from an innovation and competition. Again, it's both software and hardware, and you need to interoperate with CPUs, with GPUs and even the any one of those could be your bottleneck to the end delivered performance. And then also the software that enables that end-to-end performance with AI. So, there's a lot of place for many companies to add value. What you do have to be careful of is if whatever their blockages in interoperability, I would say the lack of customer choice at different levels of the stack, both the software and the hardware. So those are potentials, but AMD participates in a lot of it, and we also have an approach where we're working with partners and an ecosystem, and we also support the open source.
