Everpure, Inc. (P) Earnings Call Transcript & Summary

All right. Welcome, everybody. We're delighted here to have Rob Lee with us here today, who is the CTO of Pure Storage to explain a little bit more about flash and kind of the power advantages that you can get from flash in a data center. So Rob, thanks so much for being here today. Maybe you could just kind of give us a sense -- a brief introduction on yourself and your history of Pure Storage for those who may not know.

Sure. Absolutely, and thanks for having me, Meta, today. So hi, everybody. Thanks for tuning in. My name is Rob Lee. I serve CTO here at Pure Storage. I've been with the firm for 10.5 years since really the early days, and I joined as part of the founding team behind what's become our second major product line, the one that we see being put into use the most often, most frequently within the AI environments.

Got it. Perfect. And then starting given this is kind of a generalist audience who may not be familiar with the storage market, can you just give us a sense of what the approximate percentage of data center power drain that comes from storage and how much that's kind of expected to grow over the next few years. So just kind of help give a sense of the scale of the problem being faced today?

Yes, absolutely. So if we look at the -- and there's lots of estimates and it obviously depends on geography, this and that. But if we step back and we look in totality, the best estimates are that data center power on a global basis accounts for about 2% to 3% of the global power grid. So data centers as a whole today are using about 2% to 3% of the global power grid. Now within that data center power usage, it's estimated about 20% to 25% of that is being consumed or being used to power storage. So if you figure that, you're really on the order of almost 1% of the global power grid is going to power data center storage applications, which is a tremendous amount of energy that's being spent.

Got it. And just as we think about things, does that get -- like why would that get worse over time or why would that scale up over time? Just to help in people kind of contextualize how much bigger that problem could get just from a storage standpoint?

Yes, absolutely. I mean I think if -- the biggest driver for that getting worse over time really is the rate of data collection and data growth. And we've all seen the analyst charts projecting and -- well, looking at what's happened over the last 10 years and then projecting forward the rate of data growth, then you combine that with newer technologies such as AI that are giving users, giving customers more value out of the data they're storing, that just drives customers to store more and more data. Now the challenge is that the legacy or kind of the typical methods of data storage aren't getting that much more power efficient. And so you've got this major driver of data collection and data growth, which is just driving that energy consumption higher and higher. Now the other part of this is the compute that goes into these AI environments as well, right? As we all know, GPUs that are powering these AI applications are incredibly power hungry. There's -- we're obviously in a phase of large build-out of those environments. People are deploying more and more of these GPUs. And so in totality, data center power consumption is growing and is projected to continue to grow. I saw one projection, which I think is probably a bit far out there. But I did see one projection coming out of MIT's Lincoln Labs that says by 2030, the data center share of the world's power grid is projected to be north of 20%. Now I think that's probably a bit far. But that just kind of goes to show you how extreme the drivers are in terms of this wave of data collection and data use.

Got it. And so just kind of giving the audience a sense, just how much can moving off of the flash change kind of that equation or the power drain of storage within the data center?

Yes. Well, so there's really 2 elements of that. One is making the actual data storage way more power efficient and just to kind of ground the audience, we're talking about an order of 10x compared to spinning disk. The second element of this is by moving to flash from spinning magnetic hard disk, we can provide way more performance to the data that's being stored, which ultimately means that whatever job is being run, whatever AI application is being run, can be done a lot faster. So the GPU time involved is going to go down, thereby also reducing the power consumed by the GPUs or allowing customers to deploy fewer of them, again, having the same power savings effect on the GPU side. So you really have the 2 sides of it, making the storage a lot more efficient, but then also by way of faster performance, making the GPU power consumption way more efficient. Now we often get the question of, okay, so how is it that Pure's flash is able to reduce the power consumption of hard disk-based solutions by 5x to 10x? It really comes down to 2 things, right? So one is, if you look at hard disks, it's inherently a spinning media, right? You've got -- this is a typical hard disk and there's motors in here. This thing is extremely heavy. It takes a lot of power to run. If you look at Pure's flash solutions, the flash solutions that we provide, component or, call it, drive-by-drive, our's take about, call it, half the power to run. Now that's only half of the equation. The flash module I'm holding in my hand here provides about 3x to 4x as much storage as this hard disk. And so when you kind of multiply those together, right, our flash takes about half the power to run as a hard disk, it provides about 3x to 4x as much usable storage, you're now into that 8x to 10x more power efficient storage solution on the whole. And again, that's not even accounting for all of the ancillary power savings you get by just being a lot more performant, moving data more quickly to the GPUs, thereby reducing the amount of time you have to run those.

Got it. And I just wanted to -- I didn't say it upfront, but if anybody has any questions, just put them into the Q&A portal, and we'll get to them at the end. So Rob, what you just described sounds amazing that you could get kind of these 8x to 10x savings on power. So why have clouds primarily stayed with this for upwards above 90% of the storage need?

One word, cost. If you look historically, there's really nobody out there that says, "Hey, cost aside, there's some reason why I prefer to be on disk." Historically, it's really just been about where is flash cost effective to deploy and where is it not yet cost-effective to deploy. And so that's why you've seen flash really over the last 10 years start to proliferate in the data center, whether it's in the enterprise data center or the hyperscaler data centers beginning with the highest performance applications because that's where the benefits justified the additional cost. Now that equation is changing. And we would like to think that we at Pure are driving that equation to change. Because of our technology, we're now able to offer not just the enterprise, but we believe the hyperscalers as well, a cost-effective alternative to hard disk based on flash that gives them all the power benefits that we just spoke about, significant space savings. I neglected to mention that, that our solutions, not only are they 5x to 10x less power consuming, they're also 5x to 10x less space consuming. And so when you think about this from the hyperscaler lens, as they're building out data centers, as they're building out footprints to deploy GPUs, to deploy additional compute, they're running into physical limitations, right? Physical limitations in terms of how much can they fit in the buildings, how much power can they get into the buildings. Now all of a sudden, the significant power savings that they're able to get on what they have historically been spending on the disk storage start becoming very meaningful, and so that's a large part of the driver behind our current conversations that we're having with many of these hyperscaler firms really to modernize and refresh what has historically been tier point 80% to 90% of their storage deployment, which is sitting on nearline disc today.

Got it. And so maybe you kind of alluded to it, but for probably the last 5 to 10 years, it's very much been a cost equation and the cost equation just didn't make sense necessarily to move to flash except for kind of some high-performance needs. Are you seeing that conversation with the cloud [changed] to be around the power saving conversation? Or is it still just kind of -- we need to kind of take it in with the power equation or the performance equation?

The conversation is definitely shifting to incorporate power savings for sure. But I would say, it's still primarily focused around the TCO, the total cost of ownership. Power, obviously, is a big component of that, especially where there are places where you're just up against logistical limitations, like you can't -- there's a certain amount of power you can get from the utility. But at the end of the day, with the hyperscalers, it's very much driven and focused around the total cost of ownership. And when we look at it, it's a significant savings of infrastructure surrounding the storage media, whether it's the hard disks or Pure's flash, the fact that we can reduce that considerably reduces their costs, it's the reduction of power and the costs associated with power, which, again, as you'd imagine, at their scale is quite significant. It's the reduction in costs associated with lower reliability, significant lower reliability with the disc compared to Pure's flash solutions where compared to disc, about 20x, 30x more reliable. And so when you think about that from the operational cost of a hyperscaler needing to have somebody go out and wheel around the card of hard disks every day and change components, that's significant cost savings as well as overall longevity. And so the conversation very much is around TCO. I would say, the big change is that if you think about the TCO, if you think about the total cost of ownership, you've got the cost to acquire and you've got the cost to run and operate, right? And when you think about the cost to acquire, it's typically dominated by hardware costs, things that nominally sit on a deflationary scale, right? Bit by bit, hardware, whether it's flash, whether it's disc, generally deflates in costs over time over the long, long period. When you look at the cost to run and operate, whether it's power, whether it's space, whether it's labor, those generally operate on an inflationary scale, right? And so I think one of the things that's kind of tipped over, if you will, is that we've reached a breaking point where the significant operational cost savings that our solutions can deliver to customers, especially hyperscalers in the way they operate are shifting that TCO equation significantly.

Okay. Perfect. And so there are competitors of yours that have flash solutions. Can you just give some background on Pure Storage and kind of what gap you saw in the market that informed kind of the formation of the company or kind of what gap you saw in the market?

Yes, absolutely. So if we go back to the beginning of Pure, we started in 2009, started shipping product in 2012. We weren't the only flash storage provider attempting to break into the enterprise. We weren't even the first. What I would say is we were unique in really the industry and still are and that we recognize that for flash head as a media had tremendous promise, tremendous promise of performance, high reliability, a number of attributes, which we deliver through our solutions. But we realized in order to exploit and deliver the best of those properties, you really needed to build a system entirely designed for flash. You had to build a software designed to treat flash as it was meant to be treated, that it behaves very differently than disc. You can't just pretend that it's disc. And so we embarked on a very unique and I would say, a long engineering road to build our entire software stack and now our hardware stack designed to work with flash at the most native level. And we did this because we recognized that if flash was going to succeed long term, if Pure was going to succeed long term, it was going to be largely driven by the consumer market, right? If you look at what's driving flash, the semi market for flash, flash chips, it largely has been and largely continues to be driven by consumer. Well, it stands to reason that you want to be on that road map, you want to be on those economics, you want to make the consumer stuff work at enterprise levels. In order to do that, you've got to go do the hard work of building software, building hardware to make the consumer flash work. Every other competitor in the market, even today, relies on packaged SSD or solid state drive, the same type of solid state drive that you might have in your laptop or on your desktop device. And if you look at what an SSD does, essentially, an SSD is an industry's coping mechanism for the fact that flash behaves very differently. The industry didn't want to retool all the software it had to work with flash in the most native way. We had a tremendous amount of software that knows how to talk to a spinning disk, and SSD is essentially a translation layer that makes flash behave and look and feel like a spinning disk. Well, as you can imagine, making a semiconductor that is potentially capable of lots of great things, performance, efficiency, reliability, making it behave like a hard disc, well, you lose a lot of those benefits. And so that's really the distinction between Pure's approach and the rest of the industry. We've gone and done the hard work to work with flash in the most native level that allows us to get way superior efficiencies, reliability, performance and densities out of the flash modules we ship versus the competitive set and really the rest of the industry which doesn't have that IP and is really reliant on very inefficient SSDs in order to deliver their flash. And when we look at the hyperscalers, the hyperscalers largely are in that same position.

Got it. And so I mean just maybe putting that in terms of TCO or how we -- you've kind of talked about the TCO or the power advantage that you can get from moving from disc to flash. But as we think about kind of your solutions versus kind of competitor flash solutions, is there a way to kind of contextualize what additional TCO or kind of additional savings you get out of Pure solutions?

Absolutely. We've done a number of comparative studies really third parties have looking at our flash solutions versus the competitive sets all-flash solutions, really looking at, customers considering Pure, they're considering an alternative from a competitor, what are the power envelopes and power footprints of solution A versus solution B? And what we typically find is that our solutions are about 3x to 5x more power efficient than competitors' all-flash solutions. Again, because they're based on SSDs, the SSDs really limit them in a number of dimensions that all lead to worse power efficiency.

Got it. And just in terms of -- we always kind of get a question of how permanent some of those advantages that you have are -- as we move or as we're moving into kind of different drive capacities, et cetera. does that kind of -- does it get harder? And so does your lead kind of expand? Or just how do you think about how your advantages persist as we move down the cost curve?

Well, our advantages are going to widen at least for the next several years, for sure. We've -- a large part -- as you mentioned, a large part of what's driving our advantages is our drive densities. And for the audience where -- who may be less familiar, the more flash that we can put on these modules, if you think about it, I put a certain number of modules into an array, that array generally consumes same amount of power. If I can put more flash -- if I can double the flash into -- I put into a system, well, I've pretty much generally about have the power footprint per amount of storage I'm delivering. And so drive densities, right, the more flash I can deliver at the requisite performance levels, improves our power footprint considerably. And when you look at our drive density road map, we're really at the beginning of a rapid pace of improvement. As we've spoken about, we shipped 48-terabyte modules. Last year, we've shipped the 75s in November of last year. We're on track to ship 150 terabytes. So again, doubling. Year-over-year, we're actually 1.5x or almost 3x, a little early...

It's early in California. Yes.

It is. It is. And then we're projected to -- we're looking at 300 terabytes beyond that. So we've got a couple of doublings ahead of us, each of which will drive commensurate power savings. Beyond that, there's a number of other things that we have access to -- techniques that we have access to, to further improve the power footprint, again, because we're able to work with flash at the most native levels. We can be very intelligent in a large system about what parts of the system we're powering, when we can power them down way more than competitors who, again, don't have that visibility because they're relying on SSDs. So I would expect over the next several years, our gap in terms of power advantages will increase, and then we'll have to see where we go beyond that.

Great. I mean, just given the theme of the day, I think we've spent a lot of time talking about power and kind of the advantages that flash and Pure Storage can provide. But just can you give us a sense of what are the other reasons with AI workloads that it makes sense to move to flash, kind of the other piece, the performance piece of the equation?

Well, I think that there is -- if we look at AI as a whole, I think it's going to change and we are starting to see this with our enterprise clients. It is going to change how customers are looking at their entire data footprint. I think historically, customers have definitely looked at their more operational data, their more online data with the lens of, yes, I've got to have a certain amount of performance here. It's my transactional database. This is where customer transactions are happening or customer support interactions are happening. Okay, there's a sense that, that has to be a perform environment. But if you look at the typical enterprise, not dissimilar to the hyperscale, there is another whole pool of data that is being kept cold or barely lukewarm where people haven't historically been thinking about performance considerations. Now all of a sudden in the world of AI, where there's value in those data sets, either to drive training or to provide context for inference applications, well, now all of a sudden, if that data may, in some cases, go back decades and decades, if that's locked away on slow, inefficient spinning disc, well, A, we've talked about the power and space considerations, it's getting more expensive to store, but B, if you can't provide the performance to connect that data to these AI models, well, that's going to be a big limiter as well. And so I think we're starting to see customers incorporate their forward plans and thinking around AI into how they're keeping these environments up to date, how they're modernizing them. And I think it's going to be a big tailwind for shifting a lot of this cold disk footprint to warmer tiers powered by flash.

Got it. And then just how do we think of the changes and kind of storage needs as we move through training, refining, inference use cases?

That's a great question. I'll preface this by saying we're still, I think, early days in cycle. I mean, just look at -- I don't know if you were at GTC last week, but you look at the tremendous rate of change and technique improvements, I think there's -- a lot of these environments and techniques are still evolving. Now that said, I think there are some basic principles or foundational principles. When we look at training, which perhaps has had the most focused, I think there's a couple of data storage needs there. One is, well, certainly to provide data very quickly into the GPU servers so that, that data can be consumed by the GPUs to build and train those models. But what people often miss is there is a tremendous amount of -- kind of data preparation and kind of a data workflow that goes -- that happens before that data is made available to GPUs. You get raw data in, it's got to be indexed, it's got to be looked at, it's got to oftentimes be transformed, so on and so forth. Well, that all drives storage consumption and storage needs. As we look at refining and inference, I think that's where the world gets a lot more dynamic, right? And I think a large part of that is that we're going to start to see those environments connecting models to both operational real-time data as well as historical contextual data, right? So think about it. If you're deploying a model to do fraud detection on credit card transactions, as an example, right? Well, you want to make that decision in line with somebody swiping their credit cards. So you've got a real-time constraint. You don't want somebody standing at Target for 5 minutes while you're trying to make that decision. But at the same time, in order to make that decision well, you probably want to provide the model that's driving that decision with the context -- the historical context, hey, what was this individual consumers' purchasing history, what were the purchasing histories of like cohorts, so on and so forth. And so I think you're going to start to see a wide variety of data demands that are all pulling in different directions that are all going to converge in these inferencing environments. And I think that's something that certainly we see as a significant opportunity for us at Pure because when we look at the Pure Storage platform, we hit on all of those elements, right? Certainly, the latency element, certainly what we're doing to modernize these -- historically cold pools of data, bringing them into warmer tiers of flash and then providing the bandwidth and really the fast data movement to connect it all together.

Perfect. And so obviously, you guys are pretty innovative on the R&D front. Just how do you think about kind of the steps that you can take to kind of refine your software or further take a lead as we move into some of these AI use cases and kind of expand it to not only just -- kind of some of the SSD versus flash advantages that you talked about, but really expand that kind of software advantage as well?

Yes, it's a great question. I think this is part and parcel with a lot of the integration work and partnership we have with NVIDIA, where we're working with them on both the training and the inference side to really optimize those environments and whether that's adopting high-speed networking protocols, RDMA, GPU direct storage, essentially making the data transfer even faster than it is today. Whether it's really digging into the application sets that sit behind these inferencing environments and really optimizing all the way down and through the application, the data storage and what we're providing and then ultimately, the GPUs, I think that's the biggest thing that we can do is really work with industry-leading partners such as NVIDIA really understand where the space is going from a tool chain and tool set and workflow point of view and then tailor our products performance and really the best attributes that customers are looking for from the data storage to those tools.

Got it. And maybe kind of circling back, you did a great job kind of laying out those 3 phases. Is there a different power intensity? Or is it just relational to like where the compute intensity is of kind of the different phases of AI?

It's a great question. And I'll go back to -- a lot of this is still evolving early in cycle. What I would say today is the dominant consumer of power as it relates to AI really is the GPUs. Storage is significant, but it is dwarfed by GPUs. I think that will probably change over time. Right now, we're -- again, being early in the cycle, the focus really has been on time to market, training larger and larger models and the results more so than the optimization. Well, that will kind of correct itself in time. And so if you look at where the dominant consumption of power is being in the GPUs, today, that really is much more on the training side. Now I think as inference applications get deployed, that balance will shift mostly because per unit work or per job, if you will, training is way more power intense, but you're going to do inferencing a lot more. You're going to train a model once, and you're going to apply it millions of times. And so that balance will shift and I think that as inference really takes off and people are deploying that in larger volumes, it really is going to put more of a focus and a spotlight on optimizing those environments for a couple of reasons. One is, logistical power limitations is significant. But then number 2 is also cost, right? You're starting to see a lot of companies in the AI space are really digging in and trying to understand, hey, so what's my marginal cost model? What's my marginal business model? It's a little bit -- well, I'll leave it there.

Got it. So I mean the clouds have traditionally maybe built a lot of their own solutions for storage as well as a lot of other pieces of the data center componentry. Just why would that decision process change as we move towards AI or why does that equation -- why do you guys think that, that equation could change?

Well, I think that equation could change regardless of AI or not AI. And a large part of it is that kind of what we spoke about before. The vast majority of data storage deployed -- still being deployed by hyperscalers today, the large clouds is sitting on nearline disc, they absolutely recognize that it's not the most efficient media to deploy. The issue is -- and they know they need to get to flash. That's where they want to get to. The issue is that they are also sophisticated enough to realize that SSDs are not the answer. For all the reasons that we spoke about, they're not going to get the efficiencies they need. They're not going to get the reliability, a whole number of factors. And so they're looking to kind of make that jump, but without the IP, without the direct-to-flash software and hardware, IP that we have, that's a big jump for them, right? And so I think they're really looking at, hey, how do I get over that hump? At the same time, that -- and you can look at public statements by pretty much all the hyperscalers that are tightening their belts in terms of additional hiring, additional headcount and shifting more of their resources over to AI, it's clear that it's an important factor, but at the same time, it's not really something they're likely to dedicate a bunch of engineers to go build in-house. I think that's -- we look at that as a significant opportunity for Pure, right, to go and work with these hyperscalers to provide our technology to help them kind of make that jump from disk to flash, get the best properties out of it and to do it in a very cost-effective manner.

Got it. And I know you guys have highlighted in the last quarter a big win that you had with kind of a cloud AI service provider. Just as we kind of translate the last answer to that win that you guys had, just any of the reasonings or rationale different behind kind of why you were able to kind of win that customer or what they were looking -- what their objectives were?

Yes. I would say, the similarities, but I would put maybe that in a different bucket, right? When we talk about the hyperscaler opportunity on mass, that would be inclusive of AI environments, but I would say, it's actually largely focused on the cooler kind of data archive almost environments. The general pool of storage that's 80%, 90% served by hard disc today. Most of the AI environments, whether it's cloud, whether it's on-prem in the enterprise, are going to be on flash today. That's just -- again, it sits in that kind of high-performance bucket. With the GPU cloud provider we spoke about last quarter, I would say that -- well, a couple of things, one was unlike our broader opportunity set with the hyperscalers, which is really more about incorporating our IP around flash management into the hyperscalers designs, the win we spoke about last quarter was really a sale of existing products, right, existing product, FlashBlade in particular, but sale of existing products to serve their AI customers' needs. And if I look at what drove that win, I would say, it was a couple of things. One was the flexibility that our offerings gave them to deliver high levels of performance to a number of the phases of data workflows associated with AI training. We spoke about this a little bit before, that everybody focuses on AI training thinking about how fast can you move data into GPUs and provide it for the models to go and crunch on. What they often miss is all of the preparatory work that happens to transform the data, to pre-index it, to label it. Well, this cloud provider saw demand from their customers to capture all phases of that workflow. And they found that FlashBlade was an ideal platform to provide high performance to all phases of that workflow. Secondly, remind you that -- and the viewers perhaps that this was an Evergreen/One deal, right? The customer chose not just the FlashBlade platform, but really the subscription SLAs backed by the FlashBlade product to give them that flexibility, flexibility in terms of how fast customers ramp and being able to deploy and really not have to have capital expenses get way ahead of revenues, let's say, but also the flexibility to shift between SLAs as their customers' workload mix might shift in the future. And so when we look at the large GPU cloud win, those are really more of the driving factors, high performance across a wide set of the data flow steps associated with AI training and then wrapped in the flexibility and optionality that Evergreen/One gives them both from a consumption perspective, but also as a -- in terms of workload mix changes.

Got it. Okay. That's super helpful. And maybe just kind of final few questions. On just like how do we contextualize maybe -- we spent a lot of time talking about the cloud opportunity and clearly, given some of the power constraints, that's a big concern. But just how do we think about the enterprise opportunity as different from the cloud opportunity? And where are you kind of seeing customers most interested today?

Yes, absolutely. So similarities but also big differences. So certainly, we spoke about how the hyperscalers are laser-focused on TCO or enterprises are focused on cost as well. But I would say that enterprises are focused on just total risk reduction and costs being one element of risk, reliability and operational considerations being another important element of risk, but then also optionality in future proofing, right? As enterprises are modernizing their environments, especially as they're preparing for rapid pace of new technology moving to a data storage platform that gives them a lot of optionality so that if the way that AI inferencing environment changes 6 months from now, I haven't locked myself into a rigid infrastructure set that's now hard to get out of. That type of optionality and risk reduction is something that is absolutely top of mind in the enterprise. I would say that power considerations, it varies geo to geo and situation to situation. Certainly, in Europe and parts of Asia, it's a top consideration because of a variety of reasons, whether it's regulatory, whether it's just ability to get power. I would say, in the U.S., there is definitely a focus on it from a cost perspective. But in places where it becomes an absolute inhibitor -- we spoke with a client in Midtown Manhattan that was out of power and I mean just imagine what is involved in trying to add 50%, add 100% power to data center in Midtown Manhattan. It's like you can't just snap your fingers and make that happen. Well, now all of a sudden, power becomes a huge issue. So it's a little bit more binary we find in the U.S. But yes, overall, I would say, the considerations are much more about overall risk reduction cost being one of them. And I think that makes sense because if you look at the hyperscalers, they are they're looking for the best technology. They're always -- they're highly technically sophisticated. They're designing their environments. It's really about what are the best technology pieces for them to go consume in their environments. Well, the enterprise clients are looking to us to provide a much more complete platform and much more complete solution because that's -- designing data centers is -- and data center technology isn't generally their core competency in their core business. They're really looking to a partner such as Pure to go and solve those needs for them.

Got it. Okay. Well, that has been perfect, Rob. I appreciate you coming on today to kind of give everybody an overview and breaking it down in very simple terms, so we could all understand and bring in physical demos, which make even better. And so if anybody has any follow-up, either a follow-up with myself or the Pure Storage IR team, we'd be happy to answer any of you guys with questions. Thanks so much.

Thanks for having me.