$RMBS Rambus Q3 2024 Earnings Conference Call

Welcome to the Rambus third quarter fiscal year 2024 earnings conference call. At this time, all participants are in a listen-only mode. At the conclusion of our prepared remarks, we will conduct a question and answer session. If you would like to ask a question, you may press star one on your touchtone phone at any time. If anyone should require assistance during the conference, please press star zero at any time. As a reminder, this conference call is being recorded. I would now like to turn the conference over to Desmond Lynch, Chief Financial Officer. You may begin your conference.

Thank you, Dez, and good afternoon everyone. In Q3, the company delivered strong sequential growth fueled by a double-digit increase in product revenue, driven by continued market leadership from our DDR5 memory interface chips and sustained execution across our growing portfolio of products. While we invest in our product and technology leadership, we continue to generate excellent cash from operations and deliver consistent return of value to our stockholders. In Q3 alone, we generated $62 million of cash from our operations and repurchased $50 million of stock, marking our fifth consecutive quarter of share repurchases. In continued demonstration of our product leadership, earlier this month we were the first to introduce complete chipsets for industry standard DDR5 MRDIMMs at 12,800 megatransfers per second and at 8,000 megatransfers per second, which are commonly referred to as the MRDIMM 12,800 and RDIMM 8,000 respectively. Designed to meet the insatiable demand for higher memory performance in advanced data center and AI workloads, these new chipsets represent a significant expansion of our addressable market and long-term growth opportunity. In Q3, memory interface chips delivered revenue of $66 million, up 177% sequentially and 27% year-over-year. In Q4, we anticipate our third consecutive quarter of double-digit sequential growth. As expected, we are pleased to see that data center demand growth in the second half of this year is driving increased sales of our core DDR5 RCD products, where we continue to gain share and early contributions from new products. We are also very excited by the positive customer feedback on our new products, with qualification shipments on the way for both server and client applications. The rapid development of new leadership memory subsystem products is critical to the sustained advancement of powerful server architectures, and the growth of AI and other data-intensive applications are accelerating the cadence of new products. As a fundamental pillar to our growth strategy, we will continue to leverage our strong balance sheet to support our strategic investments in new products to expand our market opportunity and to drive the long-term growth of the company. The recent introductions of our complete chipsets for the industry standard DDR5 MRDIMM 12,800 and RDIMM 8,000, which I mentioned earlier, are great illustrations of this strategy in action. These new chipsets enable the next wave of DDR5-based systems and incorporate the advanced clocking, control, and power management features needed for higher capacity and higher bandwidth modules. As a recap, our newly announced offerings for DDR5-based systems include the Gen 5 RCD enabling next-generation standard RDIMMs operating at 8,000 megatransfers per second, the multiplexed RCD or MRCD and multiplexed data buffer or MDB enabling industry standard MRDIMMs running at 12,800 megatransfers per second, and finally the second generation server PMIC designed for both the RDIMM and MRDIMM. Our SPD hub and temperature sensor are also utilized in both the MRDIMM and RDIMM memory modules, rounding out our comprehensive chipset offerings. These chipsets are designed to intercept future generation server platforms currently targeted for 2026 and beyond that will use the next wave of DDR5. As part of this next wave, the RDIMM 8,000 and MRDIMM 12,800 are designed around a common architecture that allows an end user to populate a server with either module type, enabling flexible and scalable server memory configurations addressing a broad range of traditional and AI servers. A DDR5 RDIMM 8,000 utilizes the Gen 5 RCD and next generation server PMIC to deliver more than 67% greater bandwidth versus the first generation of DDR5. This represents an important new benchmark in server main memory and supports the ongoing growth of AI and other compute-intensive workloads. As AI continues to scale with larger and more complex models demanding ever greater memory bandwidth and capacity, the industry standard MRDIMM will be key to extending the DDR5 roadmap to meet these needs. MRDIMM does this by employing a novel and efficient module design that boosts data transfer rates and system performance by multiplexing two ranks of DRAM. In order to support the increased complexity of this functionality, each DDR5 MRDIMM 12,800 requires one MRCD and 10 MDB chips to multiplex the memory channel and support more memory devices. This doubles the bandwidth and provides for greater capacity while using the same physical connections of DDR5 RDIMMs. As I previously mentioned, our second generation server PMIC is part of the chipsets for both the RDIMM 8,000 and MRDIMM 12,800. This new server PMIC supports higher speeds as well as more DRAM and logic chips per module. This second generation server PMIC, along with the Gen 5 RCD, MRCD, and MDB, are further demonstrations of Rambus product leadership and are exciting new additions to our chip portfolio that significantly increase our addressable market in the years to come. Turning to silicon IP, AI continues to drive long-term momentum across our IP business. We further expanded our portfolio of leading AI-focused offerings with the introduction of the industry's first HBM4 memory controller IP. This new HBM4 controller IP provides a vital building block for cutting-edge AI accelerators, graphics, and HPC applications. Looking ahead, the requirements for power, performance, and security will only intensify, spurred by generative AI and other data-intensive workloads. Across all of our businesses, Rambus is strategically focused on advancing system memory bandwidth and capacity through innovative memory connectivity and power management solutions. In closing, Q3 was a strong quarter for the company, driven by excellent execution and double-digit product revenue growth from our chip business, consistent with our view that we would see our performance strengthening throughout the year. We expect another quarter of double-digit product revenue growth in Q4 and are on track for greater than 30% product growth in the second half compared to the first half of the year. Through our continued leadership in DDR5 RCDs, growing momentum in new products, and the introduction of our industry-first MRDIMM and RDIMM chipsets, we continue to expand our addressable market and are well positioned to capitalize on secular trends in data center and AI. Our strong cash generation enables us to continue investing in innovative products and to pursue strategic initiatives that drive the long-term growth of the company while consistently delivering value to our stockholders. As always, I'd like to thank our customers, partners, and employees for their ongoing support. With that, I'll turn the call over to Dez to discuss the quarterly financial results. Dez?

Thank you. Ladies and gentlemen, if you have a question, please press star followed by one on your touchtone phone. Our first question today is from the line of Kevin Cassidy with Rosenblatt Securities. Please go ahead. Your line is open.

Thanks for taking my question and congratulations on the good results. Just on the results, maybe to help clarify the guidance for the contract revenue was $17 to $23 million and you came in at $15 million, and now you did much better on the licensing. But maybe can you just talk what some of the moving parts might have been there?

Okay, great. Thanks for that explanation. And maybe just for further explanation, you announced having the HBM4 controller. That's I guess CPUs using that is expected a couple years out. Do you get paid for this now? Is this a license that you would sign and get paid currently, or do you have to wait for the device to the CPU to be in production?

Hi Kevin, thank you for your question. Typically, we get paid when we deliver the IP to our customers, and we've done this consistently over the years. We've been actually leading the HBM roadmap, where this IP we announced our HBM3 in '21, our HBM3E in '23, and this year we're announcing our HBM4. Every time, we have a wave of silicon companies that buy that IP in the form of a license. So we do not wait for these products to go to production to be paid for these licenses.

Okay, great. I'll get back in the queue and let other people ask questions.

The next question today will be from the line of Blayne Curtis with Jefferies. Please go ahead. Your line is open.

Hey guys, thanks for my question and congrats on the product revenue growth. I maybe just want to follow up on Kevin's question in terms of the silicon IP. I'm not sure I followed the math, but I guess for the guide, it's kind of lumpy in contract revenue and you have it back up. You had talked about kind of the silicon IP growing at like 10-15%. I don't think it's been growing that fast, so I'm just kind of curious is this kind of a base that it grows off of, or is it kind of lumpy and comes back down? It's just hard to kind of reconcile on a quarterly basis there. But maybe you can talk about if there's anything in December and then if that 10-15% growth is still the right number.

Thanks. And I just had a question on product. You know, DDR4 had been quite small. I was just kind of curious within this pickup, is it still small levels or has it recovered any?

DDR4 is still at small levels, as expected. We've seen six consecutive declines of inventories in the channels. However, the sales of DDR4 remain modest, as indicated in the last quarter at about $5 million a quarter, and we see this continuing. We see a long tail but with slow sales. The good news for us is that most of our sales that are growing this double-digit quarter over quarter are coming from DDR5, and we were expecting this given the good design win footprint that we had on several generations of DDR5. So that translates now into the growth that you are seeing in Q3 and that we are expecting in Q4. But DDR4 plays a very modest role in that growth. Most of our sales are in DDR5.

Thanks so much.

The next question today will be from the line of Mehdi Hosseini with SIG. Please go ahead. Your line is open.

Yes, thanks for taking my question. First one is for Luke. When I look at your product revenue, you've had a very nice incremental increase from Q1 to Q4, from $50 to $75 million, $25 million of incremental increase. What I want to learn from you is how much of this $25 million of incremental increase is driven by companionship and how much of it is just a migration to DDR5? And if we were to look into next year, should we assume kind of a scaling from the $75 million that you're exiting 2024?

Thank you, Mehdi. In terms of the second half of this year, you are correct. Our product revenue for the second half of this year is going to be about 30% higher than the first part of this year. There are two contributors to that. The main contributor is, as I said earlier, DDR5. We've worked over the last few years to make sure that we're leading the DDR5 race by introducing DDR5 products first in the market to secure the footprint, and that footprint is translating now into market share. So that DDR5 market share is explaining the growth in the second half compared to the first half. We also have the initial sales of companionships. We know we missed the first generation of products, but we're starting to ship those companionships. Q4 should see an increase versus Q3, and we are also very excited with the traction that we have on our first wave of PMIC products, which are in qualification with our customers and where we see qualification volumes into Q4. So the main contributor is DDR5, followed by a modest contribution of companionships in the second half, but growing with a growing qualification footprint. I think we're going to see that pattern over and over again. That's why we announced the MRDIMM products and the second generation of PMIC as well, because being first in the market with these leading-edge products eventually translates into market share. So that explains the growth that you saw between Q1 and Q4 of this year. We expect continued growth next year on our product revenue, probably with a similar pattern as this year. Granite Rapids has been announced. Typically, when a product is announced, it takes six months plus before it's in full production. So we're going to see qualification volumes and pre-production volumes into the first half, and then full volume production in the second half. So we're going to see a similar pattern as we saw with Emerald Rapids. Our companionships will continue to go through their qualification process with our customers' customers, and we continue to contribute growth quarter over quarter, with them being in full swing in the second half of next year.

Great. And just a quick follow-up for Dez. You're exiting the year with the product gross margin of low 60s. Should we think of product gross margin trending towards the higher end of your guided range of 60 to 65?

Great, thank you.

Thanks, Mehdi. The next question today is from the line of Nam Kim with Arete Research. Please go ahead. Your line is open.

Hi, thank you for taking my question. Can you give us an update on CXL controller IP development and timeline? Recently, I feel market sentiment on CXL is very strong compared to a few years ago, maybe due to HBM. What's your view on the CXL market in general and your opportunity there?

Hello, thank you for your question. First of all, we do see CXL playing a role in our silicon IP growth. We continue to see people buying from us CXL IP, PCIe for that matter, which is a sign that a lot of chips are actually using CXL interfaces. But at the same time, what we see is there's a fragmentation of that market. A lot of very different chips, custom chips, are using the CXL interface, and that's good for us from our silicon IP business. On the product side, we do have a product out with customers in the hands of our customers, and they are testing those products, and more precisely testing the use case of those products. We believe, however, that the market will consolidate at the Diamond Rapids node at CXL 3.0. Until then, it's going to be very fragmented. That's our view. The other thing is that there are some use cases, such as memory expansion, that are going to be addressed by alternative solutions like the MRDIMMs that we've just announced, which actually allow using more bandwidth and more capacity on existing infrastructure where the whole ecosystem has agreed on the same solutions for the industry. So this is a place that we understand, we are monitoring. We are enjoying the growth on the silicon IP side for CXL IP. We are ready for the CXL 3.0 node on the product side, but we don't believe it's going to be sufficient volumes on a standard type of product until CXL 3.0 is out. But we're also investing in MRDIMM for applications such as memory expansion.

Okay, thank you.

Our next question today will be from the line of Gary Moy with Loop Capital. Please go ahead. Your line is open.

Hi guys, thank you so much for allowing me to ask a question. Sorry I missed the last earnings call. Luke, you spent a lot of time in your prepared remarks talking about MRDIMM, and I think it's an important topic and a consequential market evolution over time. But the way I think the technology first evolved was specific to Granite Rapids from Intel and then maybe some joint development between Intel, Hynix, and Renesas initially. And if I'm not mistaken, MRDIMM has moved its way into the JEDEC standard set and thus has become more open for folks like you to develop product. Correct me if I'm wrong in that review of the market evolution. But my question for you is, do you think given how that market evolved, that you will have equal to or greater market share in MRDIMM that you have in current generations of DDR5 RDIMMs?

Thanks, Gary. Welcome back. You are correct on your first statement that there were some point solutions in MRDIMM that were not JEDEC compliant, that were probably good proof of concepts for the ability to increase bandwidth on existing infrastructure. The one we are announcing is the first JEDEC standard MRDIMM at 12,800, which is a notch higher than the current speed, which is also adopted by the whole industry. While we were not participating in the custom solution in the first generation, we were working on that JEDEC solution from day one because we think this is the one that is going to get traction in the market, as you said correctly, with the Diamond Rapids node. For this generation, not only are we developing the RCD chipset and the DB chipset, but as I said in my prepared remarks, we've got very good feedback on our first generation of PMIC. The PMIC is becoming more complex for MRDIMM, so we are also developing the PMIC for the MRDIMM solution. Our objective there is over time to get a similar market share as we have in the RCD market. This is a market that is going to show early volumes in 2026 for us, so we have next year to introduce the products to the market, to go through the standard qualification with our customers' customers. But we're going to have a complete chipset including the PMIC, and our goal is to reach the similar market shares we have on the standard RCD. The other aspect of MRDIMM is that the content increases just by the nature of what it is. We have one RCD chip per module and 10 DB chips per module. We will also have a complete power management and companionship offering there. So the content per module is going to be at least four times what you see on the current RDIMMs. So that's what we talk about when we say it's going to be a TAM expansion for us.

Thank you for that clarification. Seeing the product revenue up 30% second half versus first half, one thing you didn't mention that probably is factoring into that is a general market recovery and general server demand, but as well some inventory depletion in that specific channel. So to what degree is that impacting your outlook here and your trends in the second half of the year for product revenue? And how can you be sure that this isn't maybe another situation of an inventory rebuild and setting up for a situation like we saw back in maybe 2022?

Yeah, that's a good question. Yes, you are correct. We did see a recovery of the standard server market in the second half, so that's helping us. Our view of the server market is that this year it is growing about mid-single digits. If you take the midpoint of our guidance for Q4, our product business should grow low double digits. So just from that standpoint, we believe we are gaining share on DDR5. On DDR4, we do see a long inventory depletion. We were starting from a very challenging point in early 2023. On DDR5, we are not too concerned about the levels of inventory. They are in line with what we would expect for people to start production or pre-production depending on what generations they are in. So the levels of inventory on the DDR5 side of things are reasonable at this point in time.

Thank you, guys. Thanks, Gary.

As a reminder, to ask a question, please press star one on your touchtone phone. Our next question today will be from the line of Kevin Cassidy with Rosenblatt Research.

Yeah, thanks for taking my follow-up question. I appreciate the details on the MRDIMM being four times the content of the RDIMM. What are the expectations for the number of CPUs? Or will this be the new standard? Will most servers use the MRDIMM architecture versus RDIMM?

That's a good question, Kevin. It really depends on the use cases. Typically, the MRDIMM will provide large capacities and very high bandwidth. So at a first approximation, one would say that if you take an AI box, for example, for the processes, standard CPUs that do all the pre-processing, that's the typical type of application that would use an MRDIMM because in that AI box, the CPUs require a lot of memory and very high bandwidth. So that's a natural place for MRDIMMs. You'll find it in other applications where the loads are important, but they will not cannibalize the standard RDIMM market. There are a lot of applications that don't need that type of bandwidth. So that's why we see MRDIMM as an addition to our current sandbox in 2026, and it will start in the very high-end, data-intensive type of applications such as AI boxes.

Okay, great. Thank you.

Our next question today is from the line of Mehdi Hosseini with SIG. Please go ahead. Your line is open.

Yes, just a couple of quick follow-ups for me. On HBM4, thanks for highlighting opportunities, but is there any way you can help me better understand and compare Rambus's content IP in HBM4 and how it compares to HBM3E? And then one follow-up for Dez: how should I think about operating leverage looking into next year? I'm not asking for a specific 2025 OpEx guide, but just want to get a feel for the leverage given all the investment that you have made in the business.

Thank you, Mehdi. On HBM, first of all, we've been developing HBM controllers for many years. What we provide is actually the HBM controller that sits right before the PHY. We've had a history of providing HBM ahead of the market. We announced our HBM2E in 2020, HBM3 in 2021, HBM3E last year, and HBM4 this year. The difference between HBM3E and HBM4 is really a question of speed on that interface. Typically, when we release a new version of our HBM controller, we try to beat the speeds that are defined by JEDEC so that our customers have room for their designs. So the main difference is the speed, but the basic offering is the same: an HBM memory controller that CPU vendors, NPU vendors, DPU vendors, ASIC vendors can use to integrate into their chips.

Thank you.

Thanks, Mehdi. Our next question is from the line of Nam Kim with Arete Research. Please go ahead. Your line is open.

Hi, thank you. Sorry, another question related to the MRDIMM. I understand the industry has been developing two new DIMMs like MRDIMM and the other one MCRDIMM, if I recall correctly. Now I'm not sure if those two standards emerged recently, but I understand Rambus is not part of the MCRDIMM initiative. So can you explain why and what's your view on MRDIMM versus MCRDIMM?

To my knowledge, we are participating in every JEDEC standard that is related to MRDIMM and MCRDIMM. We are, like many other companies, part of JEDEC. We are participating in those definitions, and our objective as a company, just like we did for the MRDIMM, is to be present for every JEDEC product that is out there.

So are you also involved with MCRDIMM? I think MCRDIMM is not a JEDEC standard, but I'm curious.

We are developing JEDEC standards. That's our strategy because, as we said earlier, the investments for these types of solutions are huge, not only on our side but also on the part of our partners in the ecosystem. It's important for us to make sure that there's traction in the whole ecosystem for the development of high-volume solutions, as opposed to custom solutions. When you take custom solutions and add them up, the market may be large, but every single player in the market might not have revenues that justify the investments. By focusing on JEDEC standards, whether it's for RDIMM or MRDIMM or all the products we make, we are sure of the adoption of the whole industry, which drives revenue growth as opposed to depending on a custom solution for a set of different customers. So our strategy as a whole is to develop products that are specified by JEDEC so that we know we have the momentum and traction from the ecosystem, and we address a market that everyone is participating in, allowing us to have the growth that we expect to have, as we did in the DDR5 generation of products.

Oh, thanks. That's pretty clear.

At this time, there are no further questions. This concludes our question and answer session. I would now like to turn the conference back over to Luke Cifin for any closing remarks.

Thank you to everyone who has joined us today for your continued interest and time. We look forward to speaking with you again soon. Have a good day.

Thank you. This now concludes today's conference.