Rambus Inc. (RMBS) Earnings Call Transcript & Summary

Okay. Hello, everyone. Simon Woo here at Bank of America. I cover semiconductor memory chip companies and industry. That's why I'm moderating this session. Great. We have Rambus' CEO, Luc, here; and Rambus' CFO, Des, here. And before we get started, let me read the disclaimer. Today's discussion includes forward-looking statements that involve risks, uncertainties and then the other factors, which are further described in the Rambus filing, and also, Rambus assumes no obligation and then does not intend to update any such forward-looking statements. Thank you. Okay, let's get started. Thank you very much, Luc, and thank you very much, Des.

So maybe our first question is how to understand the Rambus overall background or business model? Thank you.

Yes. Thank you, Simon, and good afternoon, everyone. Rambus started about 30 years ago and reinvented the DDR technology that's how Rambus started. Actually, the main Rambus comes from the RAM bus, people don't remember this. We started our business mostly as a patent licensing company. That's how we were known. It was 30 years ago. But over time, we have evolved the business to have a much better relationship with the ecosystem and be more integrated with the ecosystem. So today, we go to market in 3 different ways. We continue to have a patent licensing business, that's the bedrock of our business. It generates about $200 million to $220 million per year of 100% margin revenue. And what we've done over time is we invested that cash inflow to the company into the development of technologies and products. So the number one pillar of our company is patent licensing. The second pillar of the company is really silicon IP. So the model for Silicon IP is we actually develop pieces of IP that we test and then we sell to semiconductor companies and those semiconductor companies integrate those pieces of silicon IP into their silicon products. The type of silicon IP we develop is 2 types. One is interfaces because that's what we know how to do, interface controllers and interface PHYs for memory and SerDes. And the other type of silicon IP we develop is security. And I'll explain why security is really, really important. That second pillar of silicon IP last year generated about $130 million of revenue. And the business model is really we sell license to the semiconductor companies every time they use that IP. So that's the second pillar. The first one is patent licensing. The second one is silicon IP. So for example, we may sell an HBM interface to people building AI GPU boxes, for example. That's how we go to this market. And the third pillar of our business is actually silicon chips. And the silicon chips we offer to the market are memory interface chips. And these chips sit between a processor and a DRAM module. It actually sits on the DRAM module. So every processor that uses a DRAM module in a server, on that DRAM module you have an interface chip and we're one of the providers of that chip. That business has done really, really well for us. Last time we met, I think, in 2019, that business was $39 million for us for the year. Last year, it was $227 million. So we saw exponential growth in that business. And we continue to see high-growth business, mainly because the market continues to need more bandwidth and more capacity. And the industry is transitioning to DDR5, which addresses that question of bandwidth and capacity. And that creates additional opportunities for growth for that third pillar. So again, 3 pillars: patent licensing, silicon IP to semiconductor companies and actual products.

Great. Maybe one quick follow-up question should be once you design the whole interface of driver chips for the memory team, who can make -- can fabricate the chip there?

So we have a standard fabless semiconductor model. So we actually design the chips and then we have a supply chain that does the front end and the back end manufacturing for us.

Yes. In Asia?

Yes. Taiwan.

Yes. Okay. Great. So guess what, everyone here at the conference -- is a day 2. We are completing the day 2 conference here. Guess what, people are talking about AI, artificial intelligence. So how Rambus can benefit from all the AI boom, AI story these days?

That's a great question. We also heard about AI in this conference since last few days. The way -- for us we saw AI as an opportunity a long time ago. We saw the fact that people would need more data and would need access to data faster. And AI is a typical example of these type of applications. So we see AI as rising the tide for us in terms of opportunity. We address AI in 2 different ways. The first is, if you remember, the 3 types of business models we have. In the silicon IP business, we have opportunities in AI because these AI processors need very fast parallel memory interfaces, typically HBM. And we have, as part of our portfolio, HBM PHYs and HBM controllers. So actually, we saw that AI wave coming early as when people were designing their AI chips, they were interacting with us on that particular IP that interfaces to these high bandwidth memory interface. Still in the silicon IP layer of our business, what AI brings to us is, it reinforces the importance of security IP because when the data center architecture becomes more complex and disaggregates a bit, where you have specialized GPUs and specialized DPUs and you have standard servers, there's more transfer of data between all of these. And people want to make sure that the data is safe, that it's secure, it's authentic, and it's safe. So the disaggregation of these new applications in the data centers have increased the need for security, and we have a strong portfolio of security IP in our silicon IP business to serve that. The third benefit of AI for us is, AI is an additional workload for data centers. We still have a lot of standard workloads that use standard servers that create the demand for buffer chips at chip level. But if you take an AI box with a training device for example, the training device itself uses HBM memory. It has to have access to parallel data very, very fast. But you need to feed that AI processor with -- typically with a standard server. So every time you have an AI GPU, AI DPU you have a lot of preprocessing, preparing the data, caching the data to make sure that you can feed it to this GPU with high-speed bandwidth memory. That preprocessing requires standard servers. The standard servers increases for standard buffer chips. So we see AI as a confirmation that people need more bandwidth and data. It's a new application that can create demand for buffer chips, but not for the AI box itself. And on the silicon IP business, as I said, it increased demand for HBM interfaces, DDR interfaces and security.

Yes. So that's a great point because investors were saying because of the very powerful GPU, AI training, AI inference related server, traditional server demand could be down, could be weaker. But you have very opposite comment.

We have a different view. I think they are very complementary. I think, as I said, a lot of workloads are standard workloads. We still continue to use Zoom. We still continue to use email to serve the Internet and a lot of these type of applications. AI processing, AI algorithms are parallel algorithms. They are different algorithms. So you need to have access to a lot of data in parallel. Most of the workloads in the data centers are not parallel algorithms. They are serial. So we see this as complementary, which is like you bring the data for this type of specialized application. A standard processor cannot do a proper, I will say, AI training algorithm on a very large data load. But the server needs to bring the data to this AI layer.

That's great. So it sounds that the -- let's go back to the traditional server, the CPU and then we have a DRAM. DRAM as a technology improved continuously. Today, the most advanced DRAM which supports the CPU is DDR5. And then the Rambus generates revenue selling the interface driver for DDR5 [ DIMM ] which supports older CPU. Okay, that's the Chapter 1. However, recently, the U.S. GPU guy provided a very [indiscernible] guidance and great result, which is based on the GPU, graphic processor unit, beside the CPU and the GPU requires high bandwidth memory, HBM memory. But your revenue generation is -- your maybe GPU guy needs Rambus patent IP to design powerful GPU. That makes sense?

Yes, it makes sense. So I would say everyone building processors and memory need access to our patent portfolio. That's what we do. So we have licensed literally the whole industry. Now if you talk about us providing technology, the type of technologies we provide to people building GPUs are actually silicon IP HBM interfaces, HBM controllers and HBM PHYs that we have tested, and they integrate this into their GPU. So it's a different model than DDR4, DDR5, what you call Chapter 1. On DDR4, DDR5, we deliver a chip, the chip goes on the memory module and that memory module goes into a traditional server. If you take an AI box or an AI processor, we don't deliver a chip. We deliver a piece of IP that they integrate into their processor and their processor with that piece of IP can interface to a different type of memory, which is HBM memory. But typically, with an HBM memory, they don't sit on the module, and they don't need an interface chip. That's why the interface sits on the chip, if that makes sense.

Yes. That's great. And then let's look at the overall macro trend. 2 years ago, people were talking about data center CapEx increase led by U.S. hyperscaler, so Internet companies. But since the last year, macro very shaky and then the CapEx spend may be down. You don't feel any near-term further correction over the server data center area except -- we can talk about AI separately, GPU separately, but how about a traditional server over the industry trend?

So on the traditional server, what we have observed is, last year the industry was capacity constrained, and it was hard for everyone to actually find chips to sell to customers, not only Rambus, everyone. So last year was really capacity constrained. The industry thought that DDR5 will start faster than what it did. So what happened is last year, in a capacity-constrained environment, people bought a lot of DDR5. Then DDR5 slipped a bit in time. So the same customers have to come back and buy DDR4 to bridge the gap. And then that created a situation where there's a little too much inventory in the channel. So what's happening this year is that people are digesting the DDR4 inventory in the channel, while they're restarting the DDR5. So we've seen a lot of volatility, if you wish, in the demand in the short run. I think in the second half of this year, we are going to continue to see some digestion in DDR4. But we also see DDR5 starting in earnest. So if we look at the longer term, we'll continue to see growth in our DDR business. DDR4 will continue to exist. It will go down while DDR5 will go up. We expect the crossover point to happen in the first half of 2024. But between now and the end of the year, it's going to continue to be lumpy, if you wish, as people digest the inventory and build their DDR5 modules.

Maybe towards the end of this year, maybe...

Things will stabilize.

I see. That's great. And then you don't see any -- I know it's maybe indirect area, but you don't see any shortage of the new CPU, which support DDR5 these days. So because last year, people were talking about CPU, which support DDR5 not enough for the supply chain.

What we saw -- as explained a few minutes earlier, what we saw is that there was an expectation in the market that the DDR5 CPU would start faster than they did. And one of the reasons we have this DDR4, DDR5 transition challenges is that those launches of DDR5 capable CPUs were delayed a bit. But we do see now them starting to ramp in earnest. So we are much more confident when we look at the second half of the year.

That's good. Okay. So then let's go back to the AI server. I know the AI server sounds very broad. But for me, as an analyst, we split AI server, one is a training purpose and the other AI server is mainly inference how to use to train the data. So the question is, so Rambus IPs can be used in both AI training, AI inference related chip?