theCUBE + NYSE Wired: Physical AI & Robotics Leaders | Avi Shetty, Solidigm & David Flynn, Hammerspace & Hao Zhong, Scaleflux

Clips
More from theCUBE + NYSE Wired: Physical AI & Robotics Leaders

Avi Shetty

Sr. Director—AI Enablement & Partnerships

Solidigm

David Flynn

CEO

Hammerspace

Hao Zhong

CEO

ScaleFlux

play_circle_outline Introduction to theCube's Robotics and AI Leaders series, featuring industry experts.

play_circle_outline Need for faster, smaller, and cheaper components in large scale AI infrastructure.

play_circle_outline Revolutionizing Storage: Solidigm's Advancements in Low Latency, High-Bandwidth SSDs with Innovations in Liquid Cooling and Capacity

play_circle_outline Hao Zhong addresses increasing demand for AI infrastructure supporting robotics and general AI.

play_circle_outline Convergence of Robotics and Edge AI: The Rise of Flash Technology in Data Center Storage Solutions

play_circle_outline Discussion on evolving data platforms and the importance of storing and processing data efficiently.

play_circle_outline Summary of each expert’s focus areas and breakthroughs in the data storage and AI landscape.

Info
Transcript

Avi Shetty, Solidigm & David Flynn, Hammerspace & Hao Zhong, Scaleflux

Avi Shetty

Sr. Director—AI Enablement & Partnerships Solidigm

David Flynn

CEO Hammerspace

Hao Zhong

CEO ScaleFlux

Avi Shetty, senior director for AI marketing enablement and partnerships at (SK hynix NAND Product Solutions Corp.) Solidigm (d/b/a); David Flynn, chief executive officer at Hammerspace Inc.; and Hao Zhong, chief executive officer at Scaleflux Inc., join theCUBE’s John Furrier during theCUBE + NYSE Wired: Robotics & AI Infrastructure Leaders 2025 event to examine how data platforms are powering next-gen robotics and AI. The conversation spotlights infrastructure upgrades from storage to edge and how innovation is reshaping the data center landscape.

... Read more

explore Keep Exploring

What is the significance of the term "data platform" in the context of AI and robotics infrastructure? add

What challenges and pressures are currently being faced in large scale infrastructure deployments, particularly in relation to components becoming faster, smaller, and more cost-effective? add

What are the key factors that influence the design and optimization of solutions in the context of AI data centers? add

What are the key considerations and challenges in developing technology for AI infrastructure, particularly regarding power consumption, data storage capacity, and computational performance? add

What is the current trend in the merging of physical and digital technologies, particularly in relation to robotics and data storage? add

What are the key considerations for data management and storage in AI data pipelines, particularly for edge applications? add

What are some recent breakthroughs in data storage technology and how are companies positioning themselves within the AI ecosystem? add

bolt Powered by CUBE AI

Avi Shetty, Solidigm & David Flynn, Hammerspace & Hao Zhong, Scaleflux

search

>> Welcome back to theCube here in our Palo Alto Studio. I'm John Furrier, host of theCube for our Robotics and AI Leaders series. This is our second annual event, theCube plus the NYSE community together, breaking out all the trends as AI goes next level, as the infrastructure continues to be invested in and extracting that value. We've got a great panel of experts. Going to talk about the data platform, all of the underlying components that are powering everything we see in autonomous vehicles, robotics, and certainly the data centers that are now becoming super computers. We've got Avi Shetty, Senior Director, AI Marketing with Solidigm, David Flynn, CEO, founder of Hammerspace, Hao Zhong, co-founder and CEO of ScaleFlux, all Cube alumni. Gentlemen, great to see you again. Thanks for coming on. Appreciate it. First of all, storage is a bad word. David, you always school me on this one. You say it's a data platform. It is a data platform because you're now seeing kind of an operating system. We've discussed many times individually and collectively that the systems themselves are becoming larger scale. So the pressure is on you guys to get the components to be faster, smaller, inexpensive cost-wise. So this is the pressure point. Networking's hot. Databases are hot. Everything's hot for the agents. David, I'll start with you since you've got the fresh funding. What's your view on this? You've seen the movie before with Fusion-io. You've seen the evolution. Where are we right now in these large scale infrastructure deployments and the explosion at the edge, which really robotics to me represents?

>> Well, I think we're still at the beginning innings. People think, well, maybe we're on the backside, the trough of... I don't think we've even hit the earnest part of the curve yet with operationalizing AI systems and using them in ways that can extract real value for businesses. I think everybody sees the potential in them just working with LLMs and seeing the advancements that have gone in there, but I think we're only in the beginning innings when it comes to realizing the potential that these things have to offer.

>> Yeah. Avi, you guys have been such great performers on the business front, on the product side, Solidigm. Obviously you had some pedigree. You're now out in the open, products selling like hotcakes, form factor, a big differentiator, cooling. This is the key to get shrinking the density down to fitting into the systems. How do you see the business right now and the world evolving?

Avi Shetty

>> John, it's always kind of vectored around four things for us, right? Ever since we've been all... I think we are part of the Intel Storage group, and ever prior to that, we were with the FPG, the Flash Products Group. And for us it's always been low latency, high bandwidth, high density and lower cost, right? You have those four vectors which you have to optimize your solutions for. And now in the AI world, we've been trying to push the industry, trying to push them to adopt SSDs, but now it's actually now where you can't have an AI data center without having optimized high-performing, scalable solutions. So for us, it's just starting and we are excited to have a full portfolio of storage solutions for our AI customers.

>> Hao, you were at the NYSE. We did an interview there. What's changed in your world in the past four months? Because networking is hot, the systems need to be tied together. Low-latency memory is huge for these intelligence and these computation systems that are designed, not as a one general purpose machine. These are large-scale clusters. What's changed in your world? How do you see things?

Hao Zhong

>> Yeah, it's such an exciting time and we keep finding the demand for the AI infra is becoming more and more, and the opportunity we're seeing is we have lots of stuff we need to enable for the robotics and gen AI. Those are things that we need a tremendous data infra for that. However, we need to keep in mind, we have constraint of power. Within the power consumption, we need both capacity for data storage as well as performance for the compute, so how to develop the technology that enable both performance and more capacity with optimization of the power? That's become more and more interesting.

>> And the physical world is connecting in with digital. This is what we're seeing in robotics, certainly. I call it the edge, but it really is the merging of physical AI. And the question is that what's the next evolution in the infrastructure to make that more of a reality? I mean, there's a general conversation in San Francisco right now when the question everyone's asking is when will robots replace the human population? And it's all out there, but 10 years? It's coming.

>> It is coming. But looking at the here and now and around storage, I think we're witnessing right now a major sea change where heretofore flash has been the performance tier and people have used more traditional architectures with hard drives for the capacity tier. But like Hao mentioned, because of power budgets, we're now seeing flash becoming the capacity tier, also because you need more performance from the capacity tier than what traditionally disk drives have offered. So higher performance requirements, lower power requirements means we're finally at, I believe, the tipping point where flash takes over the 90th percentile of just bulk capacity and retention, not just the performance sites. So this is a major shift to see.

>> Let's unpack that, because one of the things I've been squinting through and looking back at the past events over the past six months or so, give or take, the theme has been first token out, okay, we check, and then now with reasoning, it brings up, okay, more tokens, thinking's happening, reinforced learning with human feedback that came out of the Databricks event. That's going to put more demand for tokens, which is going to push the curve for demand for large scale systems. So storage is required to have that kind of thinking beyond the first token out. Okay, speed, great, check that box. But then with memory and with these foundation models, especially computer vision with robotics, a lot more tokens will be needed. What does this do to... Well, first of all, do you agree with that? I'm sure you do, but if you don't, we'll debate it. But if that happens, that changes the storage requirements, and what changes specifically?

Avi Shetty

>> For us, John, it's typically when what we understand from a tokenization of the whole AI data center is the need, whenever you generate a token, your data needs change during the whole data pipeline. I think everyone talks about the AI data pipeline. Fundamentally, it's data, and data needs to be stored, whether it's at the start, whether it's during ingest, whether it's doing your extract transform load, whether you're doing training. You're talking about robotics, which is more of the inference and at the edge. It needs to be scalable. I think David talked about it, to be small so that it has power consumption constraints which it has to take care of, and it has to be scalable. We don't know what density points an edge AI robot might require 10 years from now. We at Solidigm are always looking at pushing the density envelope to give our customers the largest density at the smallest footprint. The more data you have in a single drive, the less space you need. So in an edge use case, you're able to pack in a lot of exabytes in a small footprint.

>> That's where you guys come with the form factor. But let's take the data center, for instance. I was talking to Shad, who runs all the data centers for AWS, and they're doing $20 billion deployments. This is not a real estate deal in getting a data center. This is a design of a computer, basically a super computer. The whole facility is now looked at from a design perspective. How does that change the scope of one, how it's built, and two, how it's utilized? Because it's power-bounded. Okay, energy's going to bound it. Cooling. Maybe you can take a different approach, or rack configuration.

>> I think we are on the cusp because flash is now being designed in as the capacity tier. It's going to make some major differences. We'll see over the next five years adoption of direct contact water cooling for storage. It's never been done before. Solid-state is going to make that possible. Probably see the last vestiges of mechanical hard drives. The hot swappability and faceplate space, we'll probably see that abandoned in the near future to get higher density using the depth of the rack. So I think there's going to be some major transformations to form factor function around how it's cooled, how it's integrated to leave behind, finally, the inheritance that we got of hard drives and the premium on faceplate space to be able to slot it. The form factors that we inherited, there's going to be major transformation around that that's also going to facilitate liquid cooling and other things, because these data centers are going to high temperature. You wouldn't even be able to put in hard drives if you wanted to because it's running at too high a temperature. It has to be silicon to be there and it has to be cooled with not just liquid cooled, but high temperature liquid, direct contact.

>> What does that do for the practitioner? Oh, you want to say... Go ahead.

Avi Shetty

>> No, absolutely. I think I'm aligned with David, but before we get there, right, today's systems, you have typically liquid cooling only for the GPUs and the CPUs. Storage requirements are going up in terms of voltages when you go from a Gen 4 PCIe to a Gen 5 and soon Gen 6. So today's infrastructure does require some innovation on the cooling front. And that's where I think, John, you were there at NVIDIA GTC, where we introduced the world's first liquid-cooled SSD, where you have the coolant coming and touching the cold plate, the cold plate cooling. I have actually a product here for you.

>> It's so small. It's smaller than an iPhone.

Avi Shetty

>> It's the same form factor where you have innovations happening at the form factor level.

>> What's the capacity of this?

Avi Shetty

>> We can go up to 16 terabytes on those. Those are E1.S, 9.5mm SKUs which are directly connected -

>> Does this work?

Avi Shetty

>> Yeah, those .

>> Thank you.

Avi Shetty

>> So I think you talked about a solution where eventually there'll be liquid directly connecting to the-

>> You're already doing that.

Avi Shetty

>> But we're still cold plate, but it's still an industry first.

>> There we go. Again, shows you storage is not just storage anymore, data platform on one side, essentially really critical infrastructure on the other. Hao, what are you working on now? Because last time we chatted, the systems were being networked together. Networking is a big part of things. You got to connect these systems. They've got to get the data moving from point A to point B across multiple coordinated systems. This is distributed computing 101 meets large scale AI.

Hao Zhong

>> Yeah, correct. So that's why we really focus on optimizing along the data pipeline. So as David mentioned on the storage side, we need more capacity. More capacity also means more power saving. Imagine that if we can put hundreds of terabytes in one small factor SSDs, that's equivalent to a whole server in the past with HDDs, right? And one server of this was to the whole rack in the traditional architecture, and David is developing amazing software to make that really working efficiently to optimize the power and footprint for the data center. And Solidigm, Avi, is I think the front-runner on the large capacity SSDs to really help on that, while at ScaleFlux, we are developing more and more efficient chips to enable those capacity in the system. So this is on the storage end, and then move to more compute side, the more performance will be needed because the more performance also means more efficiency of using the compute power. Also means power saving, right? You probably recently heard that NVIDIA is pioneering the concept of the Storage Next. How about we boost the performance by 10 times, 20 times? And that's going to make the GPU computing even more efficient to reduce the data movement. So that's why we need both. It's a data pipeline. We need to optimize the whole pipeline to make the most efficient infra.

>> That's great. And you brought up GTC. I did see that slide on stage from Jensen and the scuttlebutt around the storage aficionados around me said, "Which one of those are AI native?" And I'm like, "That's a good question." So I'll ask you guys, the experts, what does AI native storage mean? Because I think this is where we're starting to see the transition from large hyperscalers who have been doing stuff for a while. As practitioners, they got the design side down at large scale. Now they got the enterprise just sitting there kind of feeling like they're modernizing. That's the buzzword, meaning they got to get their you know what together, so-

Hao Zhong

>> Well, if we look at where it's been going, it's been about disintermediating the touch points along the way. The whole GPU direct was all about getting the host root complex out of the data path when it arrives at the client and gets into the GPU without having to get copied through memory. That's great. I think we're going to see the same thing on the storage side to get to ever higher densities and ever lower power architectures like Hammerspace allow you to disintermediate the need for storage servers altogether and have flash that's more directly connected to the network, so completely bypass it. And it's funny how history repeats. The whole early days with Fusion-io and the flash becoming the performance tier, a good part of that was stepping over the legacy buses and controllers and storage networking and now what we're seeing is the next step to get to ever higher density and more direct connectivity is to disintermediate the storage server and the fat software stacks that are needed to turn that into usable data. So when you talk about AI native, I think, okay, all the shit that can be bypassed so that things can be closer, because at the end of the day it's about reducing that distance and time.

>> And density, too, is a factor, too.

>> And density, which is all gated by the power consumption. So at some point, we have to disintermediate the whole notion of a wall between the flash and the client that is a proprietary storage system and have it so it's a more open platform with flash more directly connected on the network accessed in parallel with direct zero copy.

>> All right. Avi, react to that, because what he's saying is it's an architectural shift, because you can't have the old way, whether it's an accelerant or offload or old architecture. What is the new architecture from your perspective? First, do you believe what he said to be true? And then how would you describe the new infrastructure?

Avi Shetty

>> Absolutely, John. I think any innovations which lead towards lower latency architectures are the right step forward. When it comes to the way we see storage evolving at this point especially, there's one hungry GPU who always wants data, and you need fast low latency SSDs very, very close to it, direct attach. And there is this second set of data, I think network attached, which could be on an ethernet in future. Right now it's still PCIe-based, where the focus is all about density. The goal there for AI data center operators is dollar per density while the goal for the other one, the compute node is IOPS per dollar where you need millions of IOPS reading the data back and forth in between the GPU. But this architecture does require a scale up architecture which requires the densities to grow at the same footprint and allow incremental TCO for the AI data centers to build on.

>> Hao, what's your reaction? Obviously density is big, close to the target, packets moving around. Physics play into this.

Hao Zhong

>> Certainly. I think it's a radical shift from all the stack from the chip component, sub-system all the way to the rack level. If you can imagine with one rack in the future, we're going to store extra bytes of data and that's in the typical what's whole data centers of data, and once that happening, we're definitely... Our journey is very different.

>> If you were designing a data center, say you and I were college roommates, "Hey, let's design a data center to power of the region of the Middle East." We're in Riyadh, or we could say North Carolina. Amazon just did a big 10, 20 billion dollar deal down there. How would we design it, Hao? And you guys could weigh in too. How would we think about it from clean sheet of paper? Knowing what you know, power requirements, rack configurations, how would we look at the design thinking... As practitioners, what would be our approach?

Avi Shetty

>> I think number one, there's normal hard drives. That's clearly the picture.

>> We can all agree on that.

Avi Shetty

>> We all agree on that.

>> Spinning rust, as they say, is gone.

Avi Shetty

>> And number two, every AI data center, greenfield data centers right now are prioritizing their capital investments on the GPU. So they want the rest of the components to maximize the investment portfolio. So what they're looking for is you have your TCL, you have your density needs, you have a fixed power budget, and within that power budget, what's the maximum density I can provide in that data center for the compute nodes? And obviously once you've moved on to the SSD, you are saving power. In an AI data center, there's nothing called saving power per se.

>> You use all the power.

Avi Shetty

>> So you reroute the power back to the GPU or you add more GPU components for the powers you saved by doing efficient storage. I think using ScaleFlux optimized controllers, Solidigm SSDs and then Hammerspace solutions, you're looking at efficient storage or in actuality where you are giving the data center the best TCL to maximize its CapEx investment on its GPUs.

Hao Zhong

>> I think the efficiency is the key. And let's take Jensen's words, "Scale up first and then scale out." So if we can do most efficient way on both scale up and scale out, that's where we want to be in our future.

>> And the silicon makes a big difference, to your point.

>> So I think specific from the architectural point of view, that means the performance tier of storage actually moves into the compute tier. Local NVMe inside of the GPU servers themselves becomes the shared storage tier. You're not going to get better connectivity than between the GPUs, and this is especially important for cloud architecture because the networking in the cloud is a very difficult problem because it's very spread out, very wide. So if you can have the performance tier live inside of the compute nodes that are on the same interconnect or fabric, now your performance tier is as high as it's going to get because it's in close proximity. So the performance tier is going to move forward. We talk about that as tier zero, and the capacity tier is going to shrink, disintermediate the servers, and be something that's more directly connected on the network and at a somewhat higher semantic level speaking file as opposed to raw block, because block would require that you put some layer out front, an intersection that has to reroute the data and put a file system on it. So we're going to see the file system evolve back into the SSD and the SSD get put more directly on the network, much higher density. So the way you get the efficiency is having the storage dissolve into the compute tier or have the storage server dissolve away and have flash directly on the network.

>> So either way, the tiering, these bottlenecks will go away. I mean, just network hops-

>> They have to....

>> on the scale up.

>> They have to.

>> You can go down by two hops now in the last-

>> AI native is that those things get factored out.

>> I mean, right now one hop savings just can give you massive performance. Scope that for the audience that doesn't understand some of the scale up challenges, because as you tier things, storage and networking, the hops matter. Talk about the importance of the hops and why that-

>> Well, if you counted today from a traditional NAS storage system, the data has to go through 10 retransmissions between chip to chip or over a network cable 10 times because it gets sourced from NVMe, goes through the root complex out the network adapter, back in on another controller node, through it. By the time you get to the GPU, that data has been transmitted over a bus or a network 10 times.

>> All right, so scope the consequences of that because that means what, that you have potential data loss, and-

>> Every single one of those things is power burning to signal-

>> Idle GPU times, maybe?

>> The GPU sitting idle. All of that is latency and power burning along the way. GPU direct got rid of one of those 10, right? Because now it didn't have to go through the host root complex, went straight into GPU. NVMe over fabrics gets rid of one of the remaining nine or down to eight. The way you get rid of the vast majority of them is using a parallel file system architecture. Parallel NFS lets you get rid of five of them because you don't need a controller node anymore. And then you start putting the flash device directly-

>> All right, so I lost count. Where are we? Three?

>> So that's down to five.

>> Okay. All right.

>> But if you put your compute tier, if you use local NVMe, you can reduce it to just one hop, right?

>> So that's where we're getting to.

>> It's in the compute node, or two if it has to go RDMA to appear. So the performance tier becomes one hop or two hop instead of 10 away, and then the capacity tier is even only three away if the SSDs are natively connected.

>> That's a constraint we're designing around if you're a practitioner.

Hao Zhong

>> That's why I call it the radical change from not just the component level, but at the whole system level architectural design, as David mentioned. So we need to look at this thing holistically, right? There's so many hops in the past and it's not efficient.

>> That's a huge deal. All right, so let's pretend... I love talking to you guys because it's like OpenAI, NVIDIA, you guys see the future because you're in the systems design, who you sell to or making roadmap decisions out. Without giving away trade secrets to your customers that you work with, what are the key architectural things that we're going to see unfold over the next 18 months? Because these kinds of innovations are on the table. I'm sure there's some other stuff. You talk about silicon, you're probably in design cycles. You guys are in design cycles now at least, and they've shortened down to what, eight months ?

Avi Shetty

>> I think first things, since the AI wave, you've seen a substantially large amount, almost all of it, the attention being on the GPU and the HBM. Now over the next few years, what you are eventually going to see is people realizing that's not enough. We need efficient storage, scalable storage, and as a result, the things in our pipeline, obviously. We are in the Gen 5 world right now. We'll move to Gen 6. That focuses on the bandwidth portion. We'll keep on the wave of high density, I think. I have here, just to show, another show and tell, John-

>> Does this work too?

Avi Shetty

>> This works too.

>> Thank you.

Avi Shetty

>> Yeah. This is-

>> Cube tax....

Avi Shetty

>> 122 terabytes on a single form factor. You have a robotics team here in the event here today. Imagine eight of those on a backpack behind a robot and you have one petabyte. That's today. And we're not going to stop. We're going to keep pushing the advancements on the .

>> Yeah, and with the design, with some of the examples we went in the weeds on around hops and areas, efficiency is going to be there, more power efficiency.

Avi Shetty

>> Yep.

>> All right, so the question then to you guys to wrap up is when in San Francisco are we going to see the population of robots surpass humans? What year? 10 years, 20? Pick a year. Pick a guess. Who wants to jump at that first? David, you're looking like you're ready to go.

>> I think we'll see serious in about five years. I think we'll see-

>> We'll see humans be less than robots?

>> I think we'll see some remarkable things happening. I mean, with LLMs and language models, a couple of years ago it sort of started passing my touring test of being equivalent in intelligence, and now it's to the point where using those tools can literally make you feel superhuman, accomplishing things that would've taken years or been nigh to impossible before, like using the-

>> You can get a PhD by going to OpenAI and paying 30 grand and get a PhD.

Hao Zhong

>> That's right. I think the realization of that in the mechanical with robots is only a few years away and will be a game changer.

>> And by the way, speaking of robotics, we even talk about computer vision. I mean, they're going to require very strong storage, strong compute. Thoughts on robotics, computer vision? I mean, they're going to need to see.

Avi Shetty

>> I don't have a viewpoint on how when robots overtake.

>> No pun intended, viewpoint.

>> We got distracted. Pick a number .

Avi Shetty

>> But I'm a big fan of what Jensen said, right, where by the end of this year, every workforce will have the opportunity to add tenfold AI agents for each of its employee and workforce base, just 10X or 50X the productivity. I think that is the key one where instead of doing 10 projects now, we can do 50 with the help of AI. I think that's where Solidigm is kind of focused on is looking at opportunities where we take advantage, whether it's in design, manufacturing, supply chain, on bringing certain AI solutions into our corporate environment to increase the productivity of our workforce.

Hao Zhong

>> Yeah, I believe in a few years you will see a lot of those.

>> Five, less than five?

Hao Zhong

>> I would say less than five.

>> Once home penetration happens, I think that's when the numbers skyrocket.

Hao Zhong

>> So one thing that everyone should pay attention is the robot taxi. So I think Waymo and others are ready to spread more, not just in San Francisco. Once that happens, it's already-

>> You already see in San Francisco, Waymo's already eating into Uber's numbers and people who are happier with no driver. They play music. Your name's on the marquee. They can go into neighborhoods. You don't have worry about getting canceled on.

Hao Zhong

>> See, that one has elements of the robots, right? The computer vision and the processing, and then it's just different format of robots. And that's why I'm seeing that's coming soon, very soon.

>> Well, I'm grateful for all of you for one, being CUBE alumnis, and also being part of this great event as we kind of mid-year. Put a plug in. Each of you go around and talk about the cool things you're working on, breakthroughs, round of funding, something you're proud of, a story. Hao, we'll start with you. Talk about a breakthrough and put a plug in for what you're working on.

Hao Zhong

>> We're big on the future storage, especially for the Storage Next. We believe that's something that in the data storage domain is going to further boost the efficiency of the future AI computer infra, and also more densities. So those are the things ScaleFlux has been very focused on to be essential building block and a partner for AI ecosystem.

>> Yeah, reliability is going to be key-

Hao Zhong

>> Of course....

>> as the silicon goes in. Avi, what are you working on? Tell a story. What big breakthrough? I got a few props here to show a little bit of the breakthrough a little bit.

Avi Shetty

>> Yeah. I'm honored and proud to represent our entire Solidigm team. We have one of industry's broadest, fullest AI portfolio, the fastest SSD, the densest SSD, the first liquid cool SSD. And that's I think a true testament of our team putting a solution mindset versus just focusing on storage component. Yeah, we believe we will be an integral part of the AI ecosystem and working with the right partners, and obviously-

>> And the big breakthrough for Solidigm, people don't understand, what's the big breakthrough?

Avi Shetty

>> Oh, it's a floating-air technology, right? We've been the pioneers of bringing quad-level cell into the market. The first one back in 2018, and ever since. This is the one which I showed you is our sixth iteration of QLC technology, which is what is going to bring the world's highest density SSDs into the market in the future. So yeah, I think-

>> Keep bringing those SSDs. We love solid state. We love flash. David, go ahead. We'll wrap us up.

Hao Zhong

>> So our thing is you shouldn't have to choose between having a high performance file system, super computer class performance and something that's plug and play, easy to use, easy to deploy. Right now you have to choose which of those you're going to sacrifice. With Hammerspace, you have parallel NFS-based, native to Linux. You get these super computer class performance with something with the convenience that everybody knows how to run NFS. That's unexpected. Most people think NFS is not fast enough, but this puts it as world-class. That's number one. But because it's standard space and native to Linux, it allows you to move the data into the compute tier, so you're going to be hearing a lot about our tier zero technology, the ability to orchestrate data into the compute where you can get this extreme performance.

>> I love that little clinic of 10 to one ops tiers. Thank you for sharing. Great conversation, gentlemen. As always, great pleasure. Love getting into the low level because this is where the canary in the coal mine is. This is where the action's happening. This is theCUBE. The infrastructure and AI leaders are here with robotics. I'm John Furrier, your host. Thanks for watching.