And it’s a constant evolving process even for the same customer we deal with. I mentioned that in the last call. So it takes years to really understand or to be able to basically reach a point where you can say that, hey, I’m finally delivering production worthy — and it’s not because silicon is bad. It’s because it doesn’t work well with the foundation models that the customer put in place and a software layer that works with it, the firmware, the software layer that translates into it. All that has to work you are almost like creating an entire ecosystem on — in a limited basis, which we are recognized very well in x86 CPUs, but in GPUs, those kind of AI accelerators is something still very early stage. So it takes years. And for our two customers, we have engaged for years.
With one of them, we have engaged for eight years to get to this point. So it’s something you have to be very patient, persevere and hope that everything lines up because ultimate success, if you are just a silicon developer, it’s not just dependent on you, but dependent as much even more on your partner or customer doing it. So just got to be patient guys. I got the two only so far.
Vijay Rakesh: And on the peers getting into that market?
Hock Tan: Who is getting to the market, please repeat?
Vijay Rakesh: You talked about some of your peers like I think NVIDIA has been talking about entering the custom silicon market. Just your thoughts around that, yes.
Hock Tan: Oh, custom silicon market. I have no comment to be made on it. All I do say is I have no interest in going into a market where — we have a philosophy in running our business, Broadcom. And maybe other people have a different philosophy. Let me tell you my simple philosophy, which I’ve articulated over time, every now and then, which is very clear to my management team and to the whole Broadcom. You do what you’re good at. And you do — you keep doubling down on things you know you are better than anybody else. And you just keep doubling down because nobody else will catch up to you if you keep running in of the pack. But do not do something that you think you can do, but somebody else is doing much better job than you are. That’s my philosophy.
Vijay Rakesh: Thanks, Hock.
Operator: Our next question comes from the line of Matt Ramsay with TD Cowen. Your line is open.
Matt Ramsay : Thank you very much for squeezing me in guys. Just kind of a 2-part thing on the custom silicon stuff. I guess, Hock, if some of the merchant leaders in AI who are interested in some custom networking stuff from you either in switching routing would you consider it? And the second question is for Kirsten. The business model around custom silicon for most folks is taking our e-payments upfront and sell the end product at a lower gross margin, but a higher operating margin and you guys have wrapped this massive custom business with no real impact to gross margin. So maybe you could just unpack the philosophy and the accounting about the way that you guys approach the custom silicon opportunities just from a margin perspective? Thanks guys.
Hock Tan: I’ll take that, because you’re asking business model, you’re not asking really number crunching. So let me try to answer in this way. No, there’s no particular reason shot of what constitutes an AI accelerator. An AI accelerator, the way it’s configured now, whether it’s a merchant or its customer has a — for AI accelerator to run foundation models very well needs not just a whole bunch of loads of floating point multipliers to do matrix multiplication, matrix analysis on regression. That’s the logic part, compute part. It comes — you have to come with access to a lot of memory, literally almost cash memory tied to it. The chip is not just a simple multiplier. It has — it comes attached to memory. It’s almost a layer 3-dimensional chip, which it is.
Memory is not something we — any of us in AI accelerators are super good at designing or building. So we buy the memory from very specialized high-bandwidth memory, you all know about that, from key memory suppliers. Every one of us does that. So you parted combine the two together, that’s what an AI accelerator is. So even if I get very good net corporate silicon gross margin on mine compute, logic chip on multipliers. There’s no way I can apply that kind of add-on margin to the high-bandwidth memory, which is a big part of the cost of the total chip. And so naturally, by simple math, it will — that hold an entire consolidated AI accelerator brings a gross margin below on a traditional silicon product we have out there. No going away from that because you are adding on memory, even though we have to create the excess, the IOs that attach in, we do not and could not justify adding that kind of margin to memory, Nobody could, for us.
So it brings a natural lower margin. That’s really the simple basis to it. But on the logic part of it, sure, with the kind of content with the kind of IP that we develop cutting edge to make those high-density floating point multipliance [ph] on 800 square millimeters of advanced silicon we can command the margin similar to our corporate gross margin.
Operator: Our next question comes from the line of Edward Snyder with Charter Equity Research. Your line is open.
