Xin Guo, Solidigm — theCUBE + NYSE Wired: AI Factories - Data Centers of the Future

[00:00:00] Speaker 1: Palo Alto Studio Connect in Silicon Valley and Wall Street. I'm John Furrier with Kofi Teeth here with Dave Ollade, my co-host. [00:00:16] Speaker 2: Welcome back to theCUBE Studio here at the New York Stock Exchange. I'm Gemma Allen with NYSE Wired AI Factories. And one thing we know for sure is memory is having a moment. Joining me now to talk about that is Xun Guo, co-CEO of Solideim. Welcome, Xun. [00:00:32] Speaker 1: Hi, I'm happy to be here and I'm very happy to have the chance to talk to you. [00:00:38] Speaker 2: Well, you are operating in what has become a very interesting space very fast. You know, storage, memory. It's gone from being something that was seen as like a back office, back rack component, something that's very front and center of this race for AI, right? [00:00:53] Speaker 1: Right. [00:00:53] Speaker 2: You describe it as a seismic shift. Mm-hmm. Maybe start there, Xun. Talk to us a little bit about what's happening right now and what it means for Solideim as a company. [00:01:02] Speaker 1: Yeah, I think we call it a seismic shift in both architecture and, of course, of volume. Because, you know, fundamentally the infrastructure is changing or need to be changed to meet the of the current AI-driven demand because simple thinking is previously the infrastructure is built for human direct interaction. There's always a hand on keyboard, single thread, back and forth. But now the infrastructure needs to be built for machine interaction. It's the agents talking to the infrastructure. So human is still the initiator, but at the backend, it's agents amplify the activity by a hundred times, a thousand times. That's why there's a lot of higher demand to the infrastructure. [00:01:52] Speaker 2: Yeah. And for SSD, does that mean new addressable market, more and more and more? Like, is it a demand game? What is it driving from the perspective of your business metrics? [00:02:03] Speaker 1: Yeah, because we think this is a volumetric expansion, multidimensional, because there's a, you know, the LLM model, there's a reasoning, a genetic AI, and in the future, there's a physical AI, also multi-model with video processing, for example. So it's, you know, there's expanding on both data richness, complexity, and also persistence, so there's a lot more regulation, you know, security reason you have to store all data. So if you think about this data as one pyramid with multiple tiers, the whole thing is expanding, because it's AI activities. That's why the demand is so huge, and even compared to six months ago, it's many, many fold increase. [00:02:58] Speaker 2: Wow. And I think one thing that has really interested markets like this and analysts around this whole memory space is that suddenly as well, there's a level of predictability on run rate, right? There's an understanding of three years from now, we can almost be guaranteed that numbers and demand will continue to grow. That's somewhat unusual in the tech space, correct? Yeah, yeah. What does that mean for you and your team? Are you, like, thinking, you know, planning three years out, five years out? How are you kind of thinking about that demand internally? Yeah. [00:03:27] Speaker 1: Yeah. So I think, we think of this demand, we're really at the very beginning of the growth of this demand, because I think about how many people are actually using AI, and the activities keep increasing, right? So we need, on our side, we're managing the demand by increasing our production, looking at a faster node transition to the future technology to increase the node, but also we're looking at a higher efficiency, SKUs, for example, higher capacity SKUs, because with everything equal, higher capacity per device is the, probably the single largest lever to, to crank, because you have a certain power envelope, floor space, by increasing the density of your storage, that solve a lot of problems because higher capacity, like, take an inference and take a currently the key value cache issue that NVIDIA brought up, because you want to store all the contacts, so you can, you don't need to spend your GPU power to rebuild all the past knowledge, so you can continue work, and especially in the complex eugenic flow, you can, agents can work in, multiple agents work in the same data set, and the agent can hand off the, to next agent during the flow, and all that are just driving this key value cache expansion, so more capacity means, you know, faster time to first token, minimize the latency between tokens, and increasing your throughput and generating new content. That's what the AI economics is, and you, you want to generate the new content. [00:05:16] Speaker 2: Yeah, and higher capacity on SST, does that mean more reliability, more centralized control? What does it mean in terms of the actual product offering itself? [00:05:25] Speaker 1: Yeah, the product, you know, is just from the device capacity, you know, and the Soladan way, that was a 61 terabyte, and 122 terabyte. Now we're, this year, we're going to launch 245 terabyte, and we're going to continue pushing the MLB on that side, but also we're looking at how do you further optimize the power, to improve the power efficiency, because every watt you save on the storage, is the power you can give the GPU to use. Again, that increases your overall infrastructure [00:06:00] Speaker 2: performance. You mentioned NVIDIA. I met some of your team at GTC this year, had some great conversations. I asked one of your team members, you know, what makes you competitive, especially you think about new developments or, you know, developments that have been promised for a long time, like liquid cooling across the rack, right? Yeah. And they said, well, we co-developed that with NVIDIA. So, Solidime has a pretty strong and pretty innovative co-led partnership strategy. What other sorts of things are you guys working on, working toward, partnering with? What else is happening in the broader ecosystem for you? [00:06:33] Speaker 1: Yeah, for the broader ecosystem, you know, there's multiple vectors on the technology innovation, like lethal node, like you said, the liquid cooling, and we're also looking at immersion cooling, and also how many bits per cell, because we use a unique floating gate technology that has advantage. We could explore five bit per cell to further increase the density, but also there's other research directions by understanding the whole stack problem. Because, you know, at Solidime, we have built the industry first, I mean, among the SSC vendors, we're the first to build an AI central lab. Over there, we actually acquired, you know, the latest GPU system, working with ISV partners. We built a GPU rack, storage rack, network attached, the storage rack. So, we use that to study the AI workload. And the way to do that is actually nominal, with a nominal fee, we rent it out to a potential customer. So, any customer can log in through the standard API to run their whatever they want to do on our system. And then we can, from the back end, we can look at, oh, where's the bottleneck? Is the network switch? Is the software layer? Is the hardware layer? Is the hardware layer? So, we can optimize and tune the system. And then, on the customer, now, they actually know, oh, this configuration works for me. I'll just go buy as a reference design. Wow. So, there's a lot of new players find that very, very valuable because they don't have a hardware department. They don't have a software department. They just need to do AI. So, what do I buy? So, they come to us, try it out, and they go buy. That's why we want to see us as moving toward the AI infrastructure company by solving customer issue at the system level. Wow, that's so interesting. It's like SSD as [00:08:40] Speaker 2: a service almost, right? Like Amazon-type model for whatever your needs are. Yeah, because continue [00:08:45] Speaker 1: focusing on this, you know, pretty much a commoditized SSD itself. It wouldn't solve the AI problem. That's why we build this capability so we can understand where the pain point in the entire stack. Then, we [00:09:01] Speaker 2: optimize and solve their issue. So, if we look at Solid Dime three years ago and Solid Dime now, there has certainly been some new developments from the perspective of TAM, right? Yeah. You guys are continually growing that, it seems. We know demand is super high. We've covered that. We hear a lot about different elements and different kind of macro trends across the AI race, right? So, we hear about AI on the edge, and obviously, there's a huge race in the data center business, NeoCloud, hyperscalers. There's all sorts of competitive forces at play here. From a SSD perspective, how do you kind of fragment that market approach? Like, is it one approach fits all ideally? Or do you have kind of different market segments and verticals trying to meet new demands all the time? Yeah, we, you know, interesting. Yeah, [00:09:52] Speaker 1: just yesterday, we had our internal staff meeting. We actually went through our marketing strategy with the different verticals because there's a constraint, right? But we want to keep engaged with all the important verticals. Like you said, hyperscaler, NeoCloud, and the key OEMs. And so, we separate them into verticals and define our strategy. And, you know, there's no other way to say it, but to get allocation. We have a lot of strong long-term customer relationships. We want to continue to support them. But we kind of support all of them. But, you know, so we decide, you know, all the key customers, and we want to keep long-term relationships, and then we support them with different verticals. [00:10:41] Speaker 2: So, wow. Yeah. So, in terms of what's ahead for you and the team, I know you're also in a co-CEO role, right? Maybe it was too big a job for any one man or woman. You know, talk to me a little bit about how you guys divide the company, and your time, and your, I guess, you know, priorities. How does that [00:10:58] Speaker 1: play out? Right. So, I'm an engineering background. I am a career flash guy. I started from North Flash the device engineering, and then NAN and the SSD. So, I was appointed co-CEO in March. Before that, I was acting co-CEO for a few quarters and while running the data center engineering, basically our entire engineering group. So, going forward, my focus is really on driving the business, global business, and execution, development, technology, and operations, right? In May, we have a new co-CEO. His name is Richard Chen. He joined us from SK group. He has very, he's very experienced and long track record in corporate development, finances, you know, business strategy. So, he'll be focused on, you know, elevating our overall performance, looking at the future growth potential options, and maximizing our growth trajectory forward. So, we kind of have a very complementary skill, and I think that's the best [00:12:12] Speaker 2: for the company. And for a business as technically sound as this one, I think that's very important too, right? Having the technical leadership combined with the business leadership. Yeah. You guys can [00:12:22] Speaker 1: double-team it. Yeah. I think that the team is very happy with the combination. I think so far, we worked really well together. I think together, we're going to take the company really to its next [00:12:33] Speaker 2: phase. So, Shin, final couple of questions. What keeps you up at night from the perspective of meeting the demands of this moment, right? We hear so much about compute constraints, energy constraints, all of the things that affect the macro outputs of this era. From your perspective, in the world of SSD, what is the one thing you wish you could fix tomorrow? [00:12:54] Speaker 1: It's definitely the amount of supply available. This is always on my mind. So, working very hard of looking at how do we increase production, make investment to increase production. But, you know, in the net industry, we've been burned many times. So, we also need to calculate the risk, be careful, because we don't want to so hot-headed and then, you know, spend all the money on capex now. And so, we're constantly doing this calculated risk and staging the capex. And so, we make sure we want to support our customers' needs, but not to drive the industry into an oversupply situation. [00:13:40] Speaker 2: And then lastly, when we look five, ten years out in this industry, we hear a lot about quantum computing, all of the potential opportunities that that brings, right? I think everyone thinks it's five years away every year. From the perspective of your industry and the quantum industry, what are your thoughts there? Is this like an accelerated development process? I imagine we'll still have, you know, GPUs will still be part of that side process, right? The QPUs. What would the, what would SSD mean in the world of QPUs? [00:14:14] Speaker 1: Yeah, I think that the quantum computing is going to happen. There's no doubt about that. Whether it's five-year or not, it's been always saying this five-year. I think it's getting closer and closer now. And on the SSD side, I think our main focus is mostly how to support quantum computing security. So, all this post-quantum computing security features are being developed and integrated into the product. And also, how do we work with the software stack to make sure we support quantum computing. I think that's [00:14:47] Speaker 2: currently our R&D's focus. Well, it's certainly no easy feat. I imagine, you know, the whole, this world of tech right now, it's so busy, but also so opportunistic, right? Yeah, especially with the [00:15:00] Speaker 1: AI capability, these things are changing, evolving really, really fast. Yeah. So, we have to keep [00:15:05] Speaker 2: up with it. For sure. Well, Shin, thank you so much for coming on the Qube MRC Wired. Great chat with you. Yeah, it's great talking to you. Happy to be here. I'm Gemma Allen here at the Qube Studio at the NYSC. This is NYSC Wired, AI Factories. Thanks for watching.