Redefining Scale and Efficiency for the Al Era with Solidigm

Oh, good morning everybody. Uh, my name's Scott Shaley. I'm a Director of Leadership Narrative and Evangelist for Soy.

And just in case you're wondering why I have such a long title, I decided to get paid by the letters in my title. So it helps with the, with the whole process of up upping my annual income. And alongside me today is Phil, um, manes.

I said that right? Right. Menez Menez, thank you.

Go to Market Execution lead at Vast Data. So we're gonna split this up a little bit for you guys today. He's gonna step off for a bit.

We're gonna give you the, uh, look and feel of the world from a storage provider, IE hardware. And then we're gonna bring, fill up and give you a great rundown of how software and hardware work well together in the ecosystem. So looking forward to talking to you guys soon.

Alrighty. So we wanted to talk to you today a little bit. It's the beginning of the year.

There's lots of fun things going on in 2026. We're gonna talk about the market some drives, 'cause you know, that's what I make the cooling effect of what's going on in the world. And also, uh, a little bit about a lab that we might have, uh, spun up at solid IME as well.

But before we get into it, I wanted to kind of give an analogy and kind of wake up the room. You know, it's eight o'clock on a Friday, day three, we gotta have some fun. So beating the data requirements, how can we talk about storage in a fun different and sugar filled way?

Because I have a sugar problem. So we have this really cool feature known as the glazed donut. You look at it, it has some very interesting things about it, but there's always this problem with the hole.

And the hole in the look of this to me says hard drive and it spins and has a certain feature and there's a, there's a gap. You have to cross with some data processing and things like that. And if you put everything in the same shape and design of a glazed donut, it's always gonna take like a glazed donut.

So SSDs have spent decades following the hard drive mantra, but we've decided to have some fun now in the ecosystem. And now we've brought in the maple bar, kind of looks like the form factor of certain drives that now exist in the marketplace. E one SE three, things like that.

We broke the mold of going round. We've gone off and done our own thing with storage and we make some amazing dough at soy. Soy is a great data houser because you can store all kinds of data in the nice little intricacies of the dough inside.

And we put a nice little glaze on top 'cause we wanna make sure that you have some flavor as well. A little bit of performance kick here, a little bit of capacity kick there. But we also know that at the end of the day, you gotta have a little bit more once in a while.

And I'm not gonna break it open 'cause it's gonna spill jelly everywhere. But there's this thing called a jelly donut. And so if I build a really amazing storage drive, and you guys have lots of fun putting your data in there, what are you gonna do with your data?

'cause I'm not the guy that's gonna sit here and figure out how to manipulate, play and manage your data, but I'm gonna hold it. I'm gonna store it. I'm gonna keep it safe.

So that's why we've brought our partner, Phil, from Vast Data along, because they're kind of like the jelly in the jelly filled donut. We make the right amount of room, we make the right cavity in the space of the data and how it's managed, stored can be transferred in and out. And they fill it up very nicely with their software solutions and take advantage of some of the features that when we partner together, make the products even more effective and useful.

So that's my, uh, morning wake up for the how to bring sugar, because I love sugar into a presentation at eight o'clock on a Friday morning. So when you think about data and you think about next time you pick up a donut, there's so many different varieties of donut, but you know, the right one is always the solid item one. How about that?

So I also wanna ask the room 'cause you guys get to speak up. So there's this whole thing about storage for ai. We've got a lot going on, we're gonna dig into a whole bunch of it today.

But before we kick it off, I'm just curious, anybody awake enough to kind of throw out their thoughts about what's going on with the bubble that we're in today and where we see that bubble going? Is anybody willing to throw out a, a thought on what they think of this? Is it a flash in the pan?

Are we three years, five years? Does it last forever? Just open it up for a quick thought.

We need Is take the jelly donut and do this to it. I was told not to throw things, But you can smash them. You can smash Them.

No, that's true. But then I would make a mess and I'd have to get involved with fame. The cleanup bill.

That's your answer. That's the answer's going to be a mess. It's, I like that.

Yes, it's going to be a mess. So we don't disagree. There's definitely been a lot of challenges and a lot of it has to do with supply and demand and how they cycle.

I've been doing this for way too long. It was fun when I hit 20 years. I'm like, I'm a 20 year veteran in the industry.

I hit 30 this year, starting to feel a little old. Um, so I've been doing this for 30 years and my first cycle was the year I started in 1996. 'cause I joined a, a semiconductor company and everybody was raving about the bonus checks and the up cycle and everything like that.

And I started when it hit. And so I saw the big checks, but I never saw the big check. All my coworkers started selling cars and all that kinda stuff because of that first cycle all the way back in 96.

And we've ebbed and flowed through these cycles and I've got a couple of examples of the big cycles that have hit the market. So when we transition from the 2D architectures to 3D architectures, there was a huge hit. 'cause when you change a wave for, from one technology to another technology, there's gonna be a gap.

It doesn't matter if you've got more fabs, more wafers, whatever to bring one online. You have a glu in the old one. But then the new one comes online at such a higher density that boom, we hit this massive oversupply.

So 2D went, 3D we had an oversupply problem. So that was induced by the vendors. So there's a, there's a storyline here.

The pandemic came around, this is jumped, you know, a few decades later. And we started having the hoarding happen. And I threw up here my favorite little object that's been hoarded by all the folks during the pandemic.

And we remember those days, right? They did the same thing with semiconductor products. People were like, oh my word, the world's shutting down.

We've gotta have everything on the shelf desp despair because there's no way we're gonna be able to replace products. Over time, people were locking their engineers in data centers. I had a friend of mine who literally was on a one month cycle.

He got one month, he got to stay in the data center, sleep on a, a cot in a, in a storage room. And then he got a month off and he went back in and he was getting paid hazard pay for the month. He was in the, in the building, but he could not interact with anybody.

And so we, the, the industry hoarded our products. And so we sold everything in 20 20, 20 21. And then it stopped.

Everybody kind of said, whoa, hold back. And then we hit the next big down cycle. And it was a big down cycle because the world shut down.

Innovation had stopped. Everybody was just keeping the lights on. So the, the wells dried out for the innovation on the r and d side to put up new fabs, new this, new that.

And now we come into the current cycle and all of a sudden we've got this situation where we've got a, a good amount of supply. We've got a great situation in the market for the opportunity for amazing upside, but nobody was investing at the back half of the pandemic. 'cause we didn't have the money to do it.

Nobody did. And so we get things like the CHIPS Act, which help a bunch of companies do things and we get a bunch of stuff happening and all of a sudden the AI bubble hits and we're like, oh my gosh. Memory, memory, memory, memory, memory.

Well, soy doesn't make memory, but we use memory in our drives and we have a parent company that makes memory. And it turns out that the memory fab and the nan fab are not interchangeable anymore. That stopped around the 3D era where you had to have dedicated man lines and dedicated memory lines.

All the investment in the last 18 months to two years. Memory fab, memory fab, memory fab. Because the storage was just still not really being the, the, the bell of the ball, if you will.

But now we're starting to see, especially with what's happened, and we'll get into a, a little bit of the demonstration of that with the recent uptick of need of storage. So we've hit the memory wall, we've got memory, they're building more memory and the memory will come online in the next 18 months. But we still have a problem on the storage side.

So we have capacity that's coming online. People have made announcements. We've talked about what we're gonna do to increase supply and how we're gonna manage all that.

But you throw in the AI bubble, the lack of investment and the Moore's Law slowdown of the two, the two on two. And you have this amazing situation we're now writing, which is this, uh, little bit of lack of supply for the amount of demand we have. And the demand is driven by the industry.

Think of all the iot devices. Everybody's watch, everybody's everything. The autonomous car takeover, the humanoid robots.

It's not just the simple fact that we don't have enough wafers to make stuff. It's the fact that people are consuming it in so many different ways. Even more so now than we did 20, 30 years ago.

So the age of AI is definitely gonna keep us moving. You know, you've heard a lot of people talk about being sold out for 26 and going into 27. Our friends in the spinning world just yesterday on the earnings announcement said they've got LTAs through 27 for their spinning products.

So we're gonna see this for a little while longer, but we're excited about it because it gives us an opportunity at solid I to do things that are on the forefront of innovation instead of just making the next great drive. How do we make what we've got, what you can get access to work better, faster, stronger. And that's where partnerships like, uh, with Vast come into play that you'll hear about as we get through this presentation.

So it's an opportunity for us to actually engage more and do more with our customers than we've already been doing. 'cause one of the big things is, um, there's a lot of recent headlines that come from a whole bunch of different places. Tech Crunch, Bloomberg, Fords, Benzinga, that are all talking about what needs to happen for the AI data center.

And we managed to throw up a headline from our site about all the work that we've already done to address some of these problems today and that we can move forward into the next world. So talking about, you know, went from back end to center stage from a data center perspective. So we've talked about economics, we've got a great bunch of stories on, uh, TCO including a great one with our partner, vast achieving, uh, the scale.

We're very close to several of the Neo Cloud players in the marketplace. Core weaves and others that have been very good about being partners with us in this industry. Drive as we move things forward.

And then, you know, the market that never existed, we'll get a little bit more into that, but Jensen at CES made this comment that storage is now important. And I think everybody at solid on kind of took a, you know, fall off the chair moment and said, yay, somebody finally talks about storage. So, but the, the idea is we've been thinking about this from day one.

We've always been that. We've talked about it. We were here at field day two, AI infrastructure field day two talking about the AI cycle, how we impact it, how we work with it.

It's not about building a ECI Gen six drive that operates at seven gigabytes per second and 200 million iops. It's about how do you use that product in a system and make it most effective for the customer. And that's what we spend most of our time.

That's all that I'm doing, is focusing on this concept of how to drive customer success with the portfolio of products that they need, not necessarily the big shiny object in the room. So this is the one product pitch slide I have for you in the entire presentation. And you can see we basically got two categories of products for, uh, people to consume from a, uh, performance perspective and a capacity perspective performance.

Our flagship product here is the PS 10 10. I've highlighted it because we're gonna talk about it a little bit more in the deck with some of the stuff we're doing around our cooling infrastructure. And of course the, uh, P 53 36 is our bell of the ball.

It's the 1 22 drive. If you were lucky enough to join us last year at, uh, the, the second field day, you all received your own personal version of a 1 22. Unfortunately it was Legos, but you did receive one.

Um, but some of the metrics around that. So high capacity is a thing that a lot of people were like, we're never gonna see it, we're never gonna be able to ship it because it's, you know, it's just too big, it's just too expensive. But at the same point in time, you can see the metrics here.

Four and five exabytes of a high capacity drive in 2025 were sold by solid I 30, 61, 22, 1 in two exabytes of anything over 30 terabytes was shipped by solidi in the market. And we had 100% of the 1 22 market in 25. So a lot of people talk about delivering a product and a lot of people talk about, look at my, my pretty little object and some announcements were made last August FMS, but we're out the door shipping it, delivering this product.

And you'll hear how it's being consumed by my co-presenter fill a little bit later on. So we don't just build a product, build a product, we build a product that customers want to consume. Because if you look at the market and the mix, PCIE Gen 3, 4, 5, and we've got these pretty graphs in our marketing decks that we can send out if you guys are interested, that show the transitions of who's consuming what form factor, what generation of PCAE, what capacity, we track it all and we make sure that we're always in the sweet spot.

Our products and our solutions are always gonna be where people can get them, use them and deploy them effectively. And at some point the TCO asked to work out too the cost structure. So when we were here last time, I talked to you guys a little bit about liquid cooling and I brought a cool little um, uh, display.

That display is currently running around at many different places and actually getting in the process of being deployed. But we thought about this and we said, you know what, cooling has three aspects. First of all, we got the fans, everybody knows the fans, we hear some fans in the room from the projector, things like that.

Everything was cooled by a fan. And that dictated very much like the round, uh, circle of storage, a specific form factor and deployment model for customer server solutions. So the only way to get around that is to either slow them down by adding bigger fatter fins on our drives and making airflow adjustments and all these conversations about linear feet per minute or LFM, not the nice LLMs and L and ai.

But then we get this AI bubble and we get these massive GPUs consuming hundreds of watts of power and we'll talk about that as well. And they started putting chi cold plates on them. But it's interesting, if you look at a server configuration, the one product in most servers today that is considered the must still be 100% human serviceable.

It's the storage product because the history of certain types of round media had lots of problems. They were, they're great products, don't get me wrong. I love our friends in industry, but they just have mechanical issues.

And when we switched over to the solid state drive market, we put 'em in the same boxes 'cause it had to fit in the same slot that made it serviceable. But if I look, if I show you failure rates for these drives, the the solid state drive, there's so much less. But yet people still don't trust them the same way they do the GPU in the back that's mounted to the board with a cold plate.

And if it decides to fail, you replace the server. My little drive out front still has to come in and out because you might just wanna replace it. And so how do you do that?

You put a liquid cold plate on the drive and we showed that off to you guys, uh, last time and I have a slide of it as well. But then we gotta go one step further. Okay?

If we're gonna talk about cooling, there's this really cool thing called a dunk tank. And when I first started selling enterprise drives into the immersion market, it was 2008 and we had to pull the drive apart conformally code it, sell it as NCNR and pray and hope it never failed. Because once you conformally code it, there's no rework, no nothing.

Mm-hmm We can't do that and make it a mass scale at the way an AI data center is going. We have to find other ways around it. So what does solid Im do we tell our quality guys go figure out how to make a dunk tank work.

You know, for, for for example, so we showed off the liquid cold plate architecture last time. This is the cool little sliding uh, instrument where we take our E one s and we slide it in against a cold plate. We actually had to take that cold plate design that we showed off last year to SNA because we're an active and leader in SNA standard body around form factors.

Because when you put up drive against a cold plate and if you look at my nice broken pinky fingers, they don't touch and if they don't touch, you don't get cooling. Mm-hmm. So we actually took the specifications we designed around that E one s drive to snea to redefine the form factor requirements for a cold plate for the entire industry.

We gave that information away so that we can enable this market moving forward. It's not just about us, it's about making sure the customers get what they want and they're not always gonna buy us, they're gonna buy someone else. We want them to have a working solution at all times.

And so we, we, we stiffened the flatness, we made the, the tolerances different and we made sure that you had to put the sticker on the appropriate side. 'cause if you've looked at drives from all the vendors, we tend to put stickers on whatever side we want. And of course a sticker is a dielectric so you lose your cooling and effects and things like that.

We had to chafer the corner of a drive. We have these nice square blocks 'cause it just works. And when you come from a mechanical pressing perspective, a square corner is actually easier than a round corner.

But when you're sliding hard metal on soft metal, you really don't wanna scrape it and it has to be in search and friendly. So we actually cut the corner off the drive, things like that. Those are the kinds of work that went into just doing a liquid cold plate solution.

So now we take a look at full immersion. We're sitting and we're dunking the server in a tank. We have two versions of that that exist today.

Single phase and two phase. I've grayed out two phase 'cause we're kind of not playing in that space right now. It's a little bit more of a, an interesting ecological and environmental beast and not as much work is done in that space today.

'cause it's a constrained, confined and somewhat caustic environment depending on how you look at it. So we're focused on single phase, which is an open air tank, and we worked with partners like Hypertech around the server design, how to help them make sure their server solution and our drives work well together. And we work with a company Doug, who does portable edge friendly data centers.

It's literally a container that you can take out and drop anywhere and plug it in, fire it up, and you've got a portable data center in an immersion tank. And they've been doing that for quite some time. But it's fun to talk about it.

But you guys like data technical presentations are always a good thing. And so we wanted to give you a little bit of detail on this. So we were at Super Compute in, uh, November last year and we had Hypertech as a server vendor, Valvoline as a fluid vendor, and Midas as a tank vendor in our booth together showing a combined architected solution for our customers.

So we are one little cog in that wheel and we're not the biggest cog in that wheel, but we bring these people together intentionally to show the innovation and capabilities of these technologies. And Scott, Yes, Ray there. Sorry.

I heard a, I heard the voice of God. I'm trying to understand where NAND requires that much cooling. It's not like it was a big thermal, you know, heat sink in the, in the past.

I mean, is it because of the capacity? Is it because of the, uh, density, maybe the speeds of your logic or, or or what's going on? So interesting question.

It they do get, they have gotten warmer over the years. PCI Gen four to gen five, just nature and thermals. Just like with going from hopper to grace to Vera.

Yeah. Yeah. Gets hotter and hotter and hotter.

We're not the hottest thing in there. But in order to cool us, you still have to cool us in the right environment. And if everything else in the box has a fan or has a liquid and I have to have a fan, you've put something together.

If you look at the C as presentation from Jensen, the 2 million parts down to, you know, 15 parts, yeah, they got there because we worked with them to define the liquid cold plate for the storage. So we're not, it's not necessarily about how, how much heat we generate or how cool you have to Keep us, but you're in this ecosystem that has to require, we Enable the rest of the ecosystem to move forward with net innovation. So the, the Vera Rubbin compartmentalized liquid cooled system exists because we worked with them to define how the cold plate can work with a storage device.

And so when we talk to people that want to do immersion, you talk about supercomputing like Cray and other companies like that, that have been immersing things for decades, that's the next logical step. So it isn't about just the thermal of our drive, but how our drive impacts the ability to deploy the system, if that makes sense. And so to do this, as I mentioned, you used to have to conformally coat these things and all that kind of stuff and it creates a whole new product category, cost category.

People want off the shelf products going into an immersion tank. And so I had the luxury of going up to Montreal and that's, that's I'm, I'm the hand can. So that's why it's not a pretty hand, um, dragging my drive out of a live server in an immersion tank, pulling it up, putting it back in.

I would show a video, but videos don't tend to work great. So you get a screen cap. Um, but there's interesting challenges again for our product in this solution.

Like it used to be front accessible. Well now it has to be top accessible, right? They still have to architect an immersion server where I can reach in and do that without causing problems for the rest of the system.

Now it's interesting, Valvoline, Castrol, all these companies that do the, the fluids, it's classified as a food grade lubricant so that they can ship it from around the world and it comes in these massive tanks that are being shipped back and forth. And now the problem when you go to display one of these at a trade show isn't about physical footprint, it's can this floor support the weight mm-hmm. Of the solution.

Mm-hmm. Because these things are now laying down and think of a giant bathtub full of electronics. So it's, you know, thousands, tens of thousands of pounds being put on floors.

So now going to deploy immersion gets even more entertaining, things like that. But what we found out in doing this research is we started playing with a bunch of these different fluids. There's a whole bunch of different types I'm not gonna get into today.

If you wanna learn a little bit more about that, go to OCP Global 20 fives. Uh, a history file in our hardware engineer does an actual deep dive where my little graph comes from. But one of the benefits that we found for solid I was that when you have an A fan, you have certain amount of nuance how the LFM works.

When you have a cold plate, you've got, is the water temperature always the same in a tank? You have the most succinct sustained temperature of any environment. And what is a semiconductor like consistent temperature?

The hardest problem with doing things like endurance and reliability on a, on a semiconductor is going really warm and going really cold and keeping the data accurate. The flatness of an immersion tank is amazing from the temperature And the throttle limit is when you would start throttling the performance of the drive because of the thermal. Yeah, because it's getting too hot.

We, if we program too hot on a four bit per cell product, you go cold for whatever reason, like you decide to turn it off for a while, it's a little harder to read type of thing. So we throttle it intentionally to make sure that we don't overstress the semiconductor. But in the fluid you can see it doesn't really matter.

It's always very subtle, very comfortable at those temperatures. So are those drives in targeted for storage solutions or, or also for servers? Because I mean, CPUs will have liquid cooling GPUs.

Yeah. You know, memory. So does everybody do their own little thing then Everybody has to play well in the ecosystem.

Right? And so for us, what the big focus was is validating our products are are warrantable and usable in these environments for long term people have been putting them in tanks without quote permission forever. Right?

But now we're saying it's time to be part of the party and make sure we understand what's going on. So the, the middle picture of the drive, the reason it's showing it open is one of the biggest challenges with these food grade lubricants is the Tim acts like a sponge and or can start to melt away. Now you've still got the fluid running through the drive 'cause it flies in its ways and the nooks and crannies, but it has a potential to create too much, uh, liquid connection to a single part because the sponge just sits there and is constantly pressing one temperature dot on a mini transistor inside the drive that could literally cause the drive to fail.

So we've done this research to make sure what's going on and our partners at Hypertech actually take an off the shelf air cold ready GPU, tear it apart and remove the Tim and put it back together to put it in their servers for these tanks. Because the, Tim is one of the biggest problems in these drives. So when I, I was there in Montreal, I said, how do you guys test these products without going into this giant suber or Doug or whatever tank and validate whether the products work or not?

And he said, he walked me over to the corner, he is like, can't take a picture because there's too much other stuff going on there. But it's an industrial sized deep fryer from like a food truck. And I'm like, we used to joke about can I get my tater tots in the tank?

But they really do use a industrialized 'cause it's the only thing they can put a smaller component in and keep it at the right temperatures. Right? They're not frying it, but they use the fryer to get the temperature, have the basket and all that kinda stuff.

I thought it was absolutely hilarious. So the number one problem that we actually have with our drive in an immersion tank system is the sticker, the food grade lubricant eats the sticker glue and the label falls off. It doesn't cause any problems in the server, it just is now unidentifiable.

So we actually, the the current solution is to put a layer of uh, nail polish lacquer over the stickers for all the components, not just ours. Any label on any drive falls off in these fluids. So they literally to get that drive to stay intact for the label, for the pretty uh, you know, Photoshop, we put nail polish on it.

Scott, in play real Quick. You mentioned, I think you said Tim. Tim, What was that Tim?

Uh, the, the blue goo on the drive is thermal interface material. Tim, sorry, I used an acronym without defining it. I'm not really, don't tend to forget that.

Thank you very much. So that's an example of how we're working that. And so what we have to do is now we have to validate our drive in all these different fluids.

We have to get our own version of an industrial fryer, if you will, put our products in it, test it, make sure they work. And there's, as you would with any kind of fan system, there's multiple vendors that we're playing with. So we're doing a lot of work right now to validate, uh, the immersion cooling architectures for our products At this point.

Are you doing any of the kind of new stuff like Microsoft is doing where they're actually sending the fluid to, uh, to with a, they have cold plates and they send the fluid to directly to the component that's overheating when it needs it. Have you seen that? Um, not specifically if it's cold plate involved, it's gonna be, it's gonna be a different type of fluid.

It can just be a glycol or even water. The, these are different types of things. These are oil based derivatives.

Yeah, I think they're doing it without water. That's the whole initiative we're doing. So, so if not, I thought you knew about it, you had to tell Me, but that, not that specific version of it.

No. Okay. But yeah, the, the whole process here is also to avoid the, the water loss, right?

Because we've talked about one of the things that we're actually working on in our efficiency stories that you'll be seeing coming out from solid I over the, over the rest of the year and we'll introduce a little bit at our next field day event that we're coming to is the idea of A WWE. So we have a PUE, which is power utilization effectiveness. Now there's a water utilization effectiveness and we actually help customers understand, we've gone to the point where some of our TCO models take into the amount of carbon footprint com, uh, impact of the concrete used for the footings underneath the data center that our drivers are going in to help customers understand how to use these products.

So we're, we're definitely very much thinking outside the box on help on how we help customers solve, uh, their net problems with these products. Now Scott, you mentioned something about changing the interface. So it's top versus uh, side.

I mean I don't see that in the drive here. Well, Uh, so if the server's inserted in a rack right? You call it a front loaded drive, right?

But if the server's vertical, it's now a top loaded drive. Well it's the same, It's semantic. Ah, I gotcha.

It was part of a fun conversation. There's a video I can send you that week two of that. But we were discussing is it really front loaded now because it's coming out of the top.

But no, it, it is effectively the same architecture for our cage, right? But even in a liquid cold environment, the cage is 100% sealed around the drives and it kind of has its own little baby ecosystem. But we, we have to work with them to build the slots for the drives to allow the fluid around it.

So we go from making sure air can flow, getting it as tight as we can with liquid coplay and then we open it back up again. Mm-hmm And then it making it stay effective for the different uses of the drives. So as you can see E one s and this thing, we've done the UDOT twos as well.

Um, but it's all about helping our customers figure out how to play well with these products in these ecosystems. So customer innovation, there's this wonderful thing called AI in the AI factory and we started talking to a whole bunch of our customers and we were like, so I can give you a drive. Where are you gonna put it?

Whatcha gonna do with it? We wanna help you. And they're like, well we like the help you part, we're not sure where to put it 'cause we don't actually have one either because nobody can get 'em 'cause they're all getting deployed by a neo cloud here, this, that or the other thing.

Soy built the solid AI central lab. We actually went out, acquired a whole bunch of architectures, footprints, partner platforms and we've built an independent lab that can allow customers to remote in and test our architecture because lack of infrastructure, we fix that problem. High cost of a lab, we're taking on that burden limited real world storage data generated on these tools in our environment.

It's remote to you, it's secure. We've got all those data sovereignty problems solved and then we can help you work on how to optimize it if you want us to play nice with you, if you just wanna play with it, let us know how you wanna work together. And so the first version of the lab, um, looks something like this.

So we've got E three two petabytes in granite rapid servers. We've got four JBoss full of 122 terabyte drives and we've got another petabyte and storage servers the whole north, south, east, west. This is built, this shows all the hoppers and the B two hundreds, H two hundreds that we have.

We've got Grace, uh, Blackwells coming in and we are, we will be getting the Vera Rubin platforms as well to put into this lab. So it's an ever evolving ecosystem and that's why it says quote exciting new infrastructure updates. So this lab has the ability to create whatever architecture you want.

You wanna play with performance drives in a storage server. I've got 'em in my E three S platform. You wanna play with my 1 22 drives and you want two petabytes of storage to play with but you don't wanna buy it and I really don't wanna ship it to you for free.

Guess what? I can let you remote into it here in the AL lab. So we've actually got a wait list right now of partners that want to come in and do work in this lab on these infrastructures.

And we found out that some of our partners have even wanted to deliver their hardware solution that taps into this to go in the lab so that we can actually do it real time in our lab that way as well. We've, we've changed the game on the concept of sampling a customer. 'cause sending drives out to everybody is always a lot of fun.

People like to see the real product and sometimes do those direct compares. But a one-to-one never really helps sell the value of a solid day drive. It never has.

Maybe in a laptop environment sure, send them one drive, but in a server environment you need to send a rack of data to make it work. And this is a way that we solve that problem is give you direct access remote in, load your tools up, we'll help you design to develop around it. We actually partnered with a Neo Cloud Farm, GPU and run pod to help us run it.

So it's not just us, we're not doing it in a vacuum. We're playing with the proper players in the ecosystem to drive this lab environment. Just because you have so many customers who are generating 122 terabytes of data every day.

It will, they're generating lots and lots of data and I don't wanna ship drives that I can sell for no good reason either, if you will. 'cause again, you can't sample one or two and make it work for an AI software solution involvement. It seems to me that you could actually generate some data to fill this up by just having Q chat JPTs talk to each other.

Uh, we've we've played around with that actually. Yes. Uh, our, our good friend, uh, John Michael Hans has done a lot of work with his lab environment and playing with different ways to utilize it and we're using a lot of some of those resources actually when they're not active with customers.

We're using them internally due to exactly what you're talking about, optimizing even internal soy systems using stuff from our own AI lab. So drinking our own Kool-Aid. Um, so when we were here, uh, last time we introduced our friends from Metro AI and we talked about a little demo that we put up about how to do rag offload and that's the 65% lower um, DRAM with no performance loss that we talked about.

And that white paper is out on our website. We can get you the link for that if you're interested. Phase two of that was introduced at Super Compute and will be published very shortly.

But with the whole KV C environment, we started looking at it again before the CES announcement. We were already like, KV cache is something we gotta play with. We gotta understand it how it works.

It's been slightly redefined and we'll get into that in this next little section. But we already have been able with this AI lab working with Metro in this environment showcase that we can get 27 time faster time two first token because we're using offloaded KV cache to our SSDs to help solve the problem. So we basically have pre-pro what was announced by what's coming soon in the Vera Rubbin platform using the existing architectures.

So that time to first token, those really big karabots that you have, we've already been able to validate because of the access of this lab to the, to the infrastructure and to those architectures. You seem to be getting to be getting more and more involved in the system functionality rather than just plain storage. We are, we are a, we're storage for AI and to be able to be storage for ai, we have to know how people use it.

That's been one of the biggest gaps in a lot of the storage interactions. And one reason we bring a software partner to an event like this is it is about making sure you know how your product's being used. I can throw a FIO test at something, I can throw a customer test at something that's not real world data.

Yeah. And if anything it helps improve my next generation product. 'cause I can tweak my designs, I can tweak how I do the over provisioning, how many flash channels I have, how I utilize the DRAM in the drive because I know how the customers are actually using the product.

KV cache is a lot different than having tweaking, you know, over provisioning and numbered channels and things of that nature. Yeah, it's a scale thing. Right?

Exactly. So that's, that's what we focus on. We, most of our current customer base is scale customer, whether it's hyperscale, neo cloud partners, all that kind of stuff.

It's about how we can scale our product most effectively. So.