00:08
The 65 is on the road here in Santa Clara, the Pure Storage World Headquarters during GTC Dan. What a week? What an event? Yeah, it's great to be here. It's, it's beautiful here in Santa Clara, but this week has been a story that has been building of A I in the future. And we've had so many great opportunities to
00:29
hear from the leadership at NVIDIA. But we also are talking to so many of the partners of those that are driving the business and that's what's brought us here to Santa Clara. That's absolutely right. And the other thing we know that it really takes a village here and you know, GP us uh networking. I think one of the only thing NVIDIA is not
00:47
doing now uh is storage and we, we happen to have uh Charlie here, CEO of Pure Storage. Charlie. Welcome back to the 65. Thank you. Welcome to uh pure headquarters. It is beautiful here. A lot of Pure storage orange, which is amazing and a lot of green as well. A lot of green moss art is in and trendy.
01:08
So it's a beautiful campus. Thank you. Thank you. Very much. I heard it's pretty new. It is. We moved in last June and, uh, the, the, the troops love it. Uh, we, we love it. I feel like we finally got out of,
01:19
uh, you know, the dormitory and have our own place. Yeah, I know. It looks great. It's beautiful here. Um, Charlie, you joined pure storage about seven years ago. You know, you had been, uh, speaking about A I since really you, you joined and you were focused on the three sort of legs
01:34
of infrastructure storage, network compute, talk us, talk to us a little bit about kind of that position you had seven years ago and where it came from and how it's evolved. Now, do I have to go back, you know, we have to go back in the time machine a little bit and seven years ago, the the the general view was that everything was going to the cloud,
01:52
right? You remember that, you know, very well, right? Literally everything was going to the cloud and uh storage in particular was headed for white boxes, you know, open source code, uh you know, fully J Bo and Frank everybody in the market, practically everybody in the market, all the major vendors viewed it that way.
02:12
And so what does that mean? That meant they stopped investing and I had a, you know, sort of a contrarian point of view, which is that, well, if we believe that um uh that computer software and A I was going to continue to change everybody's life. Well, data centers, whether that was in the cloud or in my belief,
02:33
still hybrid, that is to say that uh many of the enterprises were going to be running their own data centers. Uh If that was going to continue to be important, well, and if we continue to see advances in compute, continue to see advances in networking, well, then we have to see advances in storage as well. I was going to leave the opportunity open for a uh uh a challenger to really challenge the
02:57
major players who are no longer investing in it. And so uh when pure came a calling, you know, I I responded, I thought this was the company that actually by investing in storage was going to change, you know, the the trajectory of of storage and maybe win the day and we're seeing that happen. Yeah, the foresight, uh it was, it's easy to look at now,
03:19
right? But it was really a hard call at that point. And, you know, I'm constantly trying to educate as an analyst that, you know, the importance of I call a quadrangle, which is uh you know, processing uh network storage, uh memory and accelerators and uh storage is such a key part of the A I pipeline. And as you said,
03:41
this is not new for you, you had multiple A I deployments for years. But, but I have to ask what's the state of those deployments and, and maybe we can go from, you know, MLDL, generative A I, uh, fill in the blanks here. Well, we have hundreds of, um, of the A I customers. Right.
03:59
And I have to smile a little bit when I now talk about traditional A I. Right, because, you know, there's nothing traditional about A I, but A I has been around, you know, for at least a decade and it was doing things such as protein folding or, you know, simulating stock market reactions, you know, for high speed trading or um you know, just uh uh controlling robots,
04:21
you know, in, in a factory, right? And the, we've been selling to those environments, you know, for many, many years now, of course, uh uh the world has, has changed seemingly overnight uh with uh um with uh chat GP T with uh JA I and now with uh R A and, and other and, and undoubtedly it's going to continue to,
04:42
to um build upon itself over the next several years. Um You know, fortunately we have, we do have a good background, but we also have to run like hell uh sorry for the French to keep up with this uh with this market. It's great to have a great partnership with uh with NVIDIA and others as we go down this path.
04:59
Uh a number of our customers, especially our largest customers are driving architectures and driving the software that demands a lot out of our systems. Uh So it, it's great to be working with these companies. I, I think, and just to, to, um, you know, put a rapid boat around this picture, I personally believe it's not only going to drive the highest end of storage for things
05:23
like large language models and training. I believe it's going to raise the, the uh the level of the needs for all of the storage in customers environments as they start to want to run models against all of their data makes sense. You know, you mentioned uh traditional A I Charlie and you know, it almost makes me think of the term legacy A I that doesn't really exist.
05:48
Although Pat and I love to talk about how we have four decades of the earliest algorithm. So it's it's also not that new but storage. I think there is sort of a a well understood legacy architecture for storage. And of course, here has been very focused on flash right now with this A I inflection that we're seeing talk a little bit about why flash and why you see that it's really important for
06:07
driving this A I innovation, right? Well, there are several reasons and and not only are we uh highly focused on flash, we're only focused on flash, right? We have no dis uh whatsoever flash is important for multiple reasons. One is that, you know, there's a lot of data in fact, the world's the majority of the world's data is on hard disks.
06:25
Ok. Which is hard to believe that when we think about computers and modern computers, we don't think about mechanical systems. Remember, you know, go back to EAC and had vacuum tubes, right. Uh Well, there's still mechanical systems in most, most data processing today and they're called hard disk.
06:40
Um, those hard disk systems just barely have the performance levels necessary for whatever application they're hiding behind. Right? And because of that, if you also want to leverage that data for um for A I, well, you have to copy it out of there and put it in something more performant. But we now have the ability to just replace those hard disk systems with a similarly priced
07:03
flash system which will have 54 to 5 times the performance. And at 1/10 the space power and cooling of that hard drive system. So that's another thing that 1/10 the, the um the uh the uh uh power and the cooling, well as you start adding GP us, what do you need? A lot of, you need a lot of power and cooling and data centers tend to be limited in terms of
07:30
the amount of power that they have there. You know, data centers are not sold in square feet anymore, they were sold in megawatts. Um And if you're pressing up against the edge of your power envelope, you're stuck, you know, there's no more power to be had in most locations of the world, including in the US right now.
07:48
And uh and if you are forced to expand to another data center or to bring in more power, you're talking about millions of dollars in years of activity and expense. So by being able to reduce the, the power and cooling footprint of your storage, you could save something on the order of 20% of your total data center power, which that you could then reuse for, for GP US. So you know whether it's power or whether it's
08:13
performance, you know, flash just has it all over hard disk. And you know that the press loves simple uh simple projections and you know, we're now on track to eliminate all uh disc based systems in the next four years. Yeah, it's interesting. Uh You put the performance Capex and an opex together that combination and you know, you talked about millions of dollars,
08:37
we're talking about hundreds of millions if not billions. I know there's a difference between hyper scale or, and maybe a co O or, or enterprise data center. But, but you're looking, you're in the tens, if not hundreds of millions of dollars now and, and also accelerating that uh and easily I you know, you see factoids of uh
08:59
how much next generation data centers, how much power of the entire globe. Oh yeah, but we're looking at doubling the data center power draw in two years. So it's not just an economic Capex or an economic op X or even performance uh when you have all three like, yeah, with Flash, it's very compelling. And by the way, I have heard that uh Flash was going to get rid of hard drives uh each year
09:24
for, for the year we've seen that. But, but it's like we were getting closer and closer there. So, uh GTC NVIDIA, UN A, uh you and NVIDIA have been partners for a long time. Can you talk a little bit frame, the, the relationship, uh what have you worked on? What are you working on now together?
09:46
Well, it's a strong technical relationship. Um You know, we have delivered together some of the largest, if not the largest A I um you know, supercomputers in the world, right? Um When you are pressing the boundary of performance, you know, there's always another bottleneck to get through, right?
10:05
A lot of times, a lot of times it's software, uh by the way, you know, it's drivers uh or it is uh actual, the performance of the actual application software that's running and together, we've run up against a number of those hurdles and we work together to be able to eliminate them. So that's been great. Now, uh the, as you know,
10:23
one of the latest um areas of focus for NVIDIA is what they call Rag or Retrieval augmented uh generation um in R A, what you want to be able to do. And this is this is something that we're very focused on together. You need to be able to access a, a large fraction uh if not all of the data inside an enterprise, right?
10:43
And again, that, that means that you have to be able to get access to it. There are two reasons why you can't uh why it's very, very difficult to get access to all the data in an enterprise right now. One is, as I mentioned before, data is largely hidden behind the application it serves today, right? It is not a first class citizen from a network
11:01
server, er P with er P. And that's right. So if you want to copy the data, you often have to do it, you know through the application, if you want to access the data, you have to do it through the through the application. Uh And if the performance level is is just enough for the application,
11:17
again, that's another barrier. So between the fact that the data itself is not networked and that it is um and it doesn't have the performance necessary for R A. As I mentioned, this is an opportunity to raise the level. We we pure help that in two different ways. One we mentioned which was the performance level. The second one is because all of our systems uh
11:38
operate on the same operating environment uh in data storage today, even if you have a single vendor that's not pure. Generally, they're supporting that uh that full environment with four or five different hardware software combinations, right? They're, they're not unified, we have one software environment, we call it purity.
11:58
Um It exists all of our systems that allows us to then be able to leverage that with something we call fusion to network the data storage, you know, underlying the applications that it serves. And so this allows data to be accessible to things like A I even when that data is supporting its primary application. Yeah, it's really interesting. We, we, we've entered a world now where A I is
12:21
sort of infused in every, every product. And then of course, every one of our businesses are using A I I have to imagine, you know, storage has evolved in a lot of ways. You talked about purity, you've talked about your, your software layer. How are you thinking about a pure using A I like storage oper for your,
12:41
you know, for your practitioners that are using your system? Is A I becoming part of that story. It is we just saw a demo of this uh last week at our sales kick off. Um you know, simply stated, uh you know, one of the things that pure was founded on was the idea of simplicity. Uh you know, we are our systems and individual systems.
12:59
Very simple. What fusion is doing is making um a full scale deployment of lots of different systems able to be accessed as if it's a single pool of data. And now we're looking at using A I to basically give it a natural language interface to customers. So customers could say, listen, I want to be able to provide X number of terabytes of storage to this application that I've just
13:25
deployed. And instead of having to, you know, define um you know, define lungs or to be able to actually write uh the, you know, the code necessary to get it to happen. It would be generated automatically or let's say you wanted to analyze your retailer and you've been running uh pure storage for, you know, the last year or two,
13:46
you could say, look, I'm expecting a 30% increase in traffic or uh through the Christmas holiday, uh will the system be able to operate or are there things I need to enhance and it would come back to, you know, that kind of language in a prompt and it would be able to come back to you and tell you, you know, what you might need to do differently or order or uh change in order to be able to meet that surge in demand.
14:11
It's been really fun. I in my heart, I am a product person and seeing kind of point of differentiation, incremental value to customers. I like the way that you were software first and it was really about the experience. I like the architecture that enabled. Hey, you need new storage. Let's just pull it out and, and put a new um an upgrade in very sustainable
14:38
that you also have, have evergreen. How does evergreen fit into? I can almost hard for me to say this Straight face Legacy A I and, but how does it play into this newer flavor of generative A I? Well, the great thing about evergreen and now especially evergreen one is it gives the customers optionality for the future.
15:02
If they had purchased the product, they kind of get stuck on what they purchase and have to utilize that, right? With Evergreen one, they can basically sign a contract and then as they go forward in the future, as their needs change, they can change out different parts of their storage for other parts of their storage. So, you know, where as a customer in the past decided, well,
15:24
I I got this environment that I need storage for, for the next 10 years and they would buy it. How, how much can we predict life? Five years from now? 10 years from now in the it environment hard. Evergreen one gives the customer the benefit of, of not having to decide today, what the next five or 10 years uh looks like it gives them all that flexibility that they can
15:46
change underneath. Yeah, I mean, we've seen just crazy now, MLDL Generative A I um mixture of experts, I mean, it, it, it seems like every two years, we're coming up that next new thing and sure researchers are working something so somebody can buy evergreen one and have some sort of a guarantee that they can protect some of their investment even though you might be putting new tech in new
16:13
tech or different tech, right? Or, or higher performance or frankly, I need less of X I do that. I mean, pure really was building the evergreen one model around looking and feeling a heck of a lot like cloud and how people consume. Well, and actually I will tell you that it is cloud now. It is a Yeah, thank you. No, it is a SAS model.
16:34
The only difference is the infrastructure for that SAS model. You know, it could be on a hyper scale, it could be on a colo, it could actually be on the customer's premise. But to, you know, really wrap a low around that SAS model. If it's on the customer premise, we pay them for the space and power because we're hosting
16:52
it there. It looks, but it very much feels like how consumption law enterprises want to consume. And you've given them the flexibility to do that, stay up to date, stay upgraded with the newest technologies, you can continue to evolve it and then deliver it to the and this as I see it, you know, it's interesting that so many people at GTC are talking about helping both is in
17:12
avenue and I both, you know, we endlessly like to talk about, you know, there's, there's the chip kind of mercenaries, right. Oh, you've got a GP U, you've got an A, you've got, and then there's the systems and then there's the whole stack and in the software, if you look at that, that's sticky with one.
17:30
I think there's an argument that you've put some years time energy effort into building a pretty big mode. How big is that? Moat? It's a big mode. We have four things that we think it's, you have to start your software from scratch in order to be able to mimic what we do. Evergreen is a good example of that,
17:45
right? This whole idea where we can do disruptive upgrades literally forever, you know, of every component in the system without causing any application downtime with our cus scanner, right? If you haven't built your software from scratch to be able to do that, you're not going, I don't think our, our competitors can,
18:02
by the way, we've had it for 10 years, no one's caught up yet. So then we have our direct to flash technology, right? Direct to flash means we don't use S SDS which were designed to mimic hard disks. Um You know, since when do you design a semiconductor to mimic a mechanical device? So we have our directive flash, it's what's going to allow us,
18:20
in my opinion, to be able to penetrate even the hyper scalars with our technology, third, we have what we call our cloud operating model. So this is the idea that the customer interacts with us entirely through and with the systems entirely through the web, uh you know, and with evergreen one, they never have to touch a system, we manage that uh entirely right.
18:41
And the fourth one is the fact that we alone have a highly consistent portfolio by which I mean, we have one operating environment which we call purity and one management environment, which we call pure one to manage all of our products. Whereas the the uh the uh the storage industry has been characterized by having different software, hardware combinations for every storage niche that exists out there.
19:05
What flash has allowed us to do is really unify to have one environment. Imagine we had a half dozen different networks in every customer. How would you be able to network those applications? You wouldn't be able to do that, right? With a half dozen or more different storage systems, you can't really network that data,
19:23
um you know, conveniently or efficiently we have one operating environment across all our products. Makes, makes it easier for our customers to be able to manage, gives us the ability to have fusion, whereby the data is able to be networked. And it just reduces the complexity for a data management story in there, there's a data management story in there.
19:41
Exactly. And by the way, I have to hit on this. So I've been known to be the person sitting in these and I'm usually not, I'm never the first analyst to ask the question. But one thing I've definitely seen in the storage industry is storage companies recasting themselves as data companies. And then it's like, OK, well, what about the data management companies?
20:01
How do these two come together? So there's a little bit of marketing going on, but is there a little bit of reality going on where that data is right there? Let's do something. So there's, there's a little bit of both as, as, as you point out and, and the difficulty is because there's not the right words yet to describe this.
20:20
So for example, the data storage companies generally don't know the exact data that's in the system. But what we will be managing is not the individual bits of data, but the data sets themselves the whole data set. And if you think about how data is exploding and how there are different data sets that need to be used and need to be managed because when you make a copy of a data set,
20:42
don't you want to know where that copy is? And when you make and when you modify that copy and yet make another copy of it, you want the provenance of all these different data sets and how they relate to one another. That's what we can manage. Now, um there's still going to be you know, the whole ETL chain uh somewhat disrupted by A I, but there's a whole ETL chain and there's
21:04
data management that goes on there as well. So if I were to be, you know, even more exact as to what we do when we talk with that is we pure do when we manage data, I'd call it data set management trust kind of a, you know, file versus object. And there's some complexities in that. But the way, eventually and by the way, Pat you were, you were kind of nice,
21:24
but I'll give Charlie the last word on this. But I mean, there are some companies that are storage companies that are kind of proclaiming to be data management company and that's going on and, you know, we're doing a lot of work internally to, to sort of, you know, there's a certain amount of that to your point that can be done,
21:39
right? But there is a cut over, there's a cut over between data sets and data itself. That's right. So wrap us up here. Um give us the outlook for the next year. I mean, jeez, it feels like asking you doesn't give you enough runway, but realistically how much innovation happens in a year now backing you.
21:56
What's the next year for pure? Well, we've done, I mean, really our product um uh proliferation over the last year has been nothing short of tremendous, you know, we introduced just not even a year ago, our first uh uh uh E family product, OK, which was slash slave to e uh we've had tremendous success with the E family. This is the first product line that can now
22:17
address um cheap dis, you know, at a similar price point, but 1/10 of the space power and cooling and it's what allows us now to claim with the one operating environment that we can satisfy all the customer's needs, right? Because while it was just what flash was before, which was high end, we weren't addressing the majority of needs.
22:37
Now, we can address the majority of needs. Um you know, fusion is going to allow us to network that data uh or those data sets so that customers can get access to it regardless of where it is inside the entire environment. And you're going to see, of course, um us really addressing some of the highest performance needs for even training in the A I environment.
22:58
So full range on one operating environment, one set of products that are consistent from A I uh all the way down to archive right block file and object and everything from small scale to large scale to, you know, exabyte scale. So it's a pretty broad, you know, an amazing uh uh product line to be, you know, fully integrated with one operating system.
23:24
Charlie, thanks so much uh for sitting down with us here. No, it's really my pleasure always. It's great stuff. Great. Uh Great to see you. Great, Janine. It's fun stuff. Yeah, excited to follow the journey and I'm
23:35
sure we'll be talking really soon. Thank you very much for the time. All right, everybody hit that subscribe button. Join us for all of our episodes here of the 65. We're on the road at pure storage at their beautiful headquarters here in Santa Clara during the GTC 24 con uh conference we gotta go for now. We'll see you all really soon. Bye bye.
