Name: Storage Beyond the GPU: The CPU at the Center of Agentic AI Infrastructure
Uploaded: 2026-07-23T18:30:00.000Z
Duration: 53 min 10 s
Description: AMD, Supermicro and Vast Data leaders discuss how CPUs, storage and networking work together to keep GPUs productive, improve AI infrastructure economics and support enterprise agentic AI deployments across on-premises, cloud and hybrid environments.

Clips
More from Beyond the GPU: The Full-Stack Fight for Enterprise AI

John Mao

Global Head of Business Development

VAST Data

Penny Tseng

Product Marketing Development

AMD

Ben Lee

Director, Solution Management

Supermicro

play_circle_outline Supermicro H14 platform and deployable AI system engineering

play_circle_outline When NVMe All-Flash Shifts the Bottleneck: Balancing SSD Throughput, Network I/O, and System Performance

play_circle_outline Harnessing DDR5-8000 16-Channel Memory for CPU-Centered Agentic AI Infrastructure and Orchestration

play_circle_outline GPU vs CPU roles; shifting GPU:CPU ratios toward parity

play_circle_outline VAST Data’s All‑Flash Efficiency: Data Reduction, Erasure Coding, and Software Strategies for NAND Supply Pressures

play_circle_outline PCIe Gen5→Gen6 enabling 800G and future 1.6T networking

play_circle_outline TCO focus: performance-per-dollar and performance-per-watt

play_circle_outline Reliability, data readiness, and organizational impact of AI adoption

play_circle_outline Workloads: unstructured data, streaming, video, biotech use cases

play_circle_outline AMD EPYC Genoa and Turin (SP7/SP8): Powering Enterprise On-Prem, Neocloud, Hybrid and Edge AI Deployments

47:28

Info
Transcript

Storage Beyond the GPU: The CPU at the Center of Agentic AI Infrastructure

John Mao

Global Head of Business Development VAST Data

Penny Tseng

Product Marketing Development AMD

Ben Lee

Director, Solution Management Supermicro

AMD, Supermicro and Vast Data leaders discuss how CPUs, storage and networking work together to keep GPUs productive, improve AI infrastructure economics and support enterprise agentic AI deployments across on-premises, cloud and hybrid environments.

Storage Beyond the GPU: The CPU at the Center of Agentic AI Infrastructure

(INTRO)

Dave Vellante

>> Hi, everybody. Welcome to this special theCUBE Power Panel. Beyond the GPU, we're putting the CPU at the center of agentic infrastructure. Let me set it up and then introduce the panel. AI infrastructure, that narrative around AI infrastructure, of course, has always been focused on the GPU. But accelerators, they only really create value when data reaches them consistently, and of course, at sufficient speed. You got to keep those GPUs busy. Now, Flash, of course, removed the bottleneck from spinning disk and it pushed the bottleneck elsewhere, out to the network, for example. So as Flash removes those traditional media constraints, the bottleneck shifts to the network and the system I/O path. And today we're going to examine how AMD with EPYC, Supermicro's H14 and VAST Data create a balanced and efficient system and platform for AI storage. My name is Dave Vellante. To my right is John Mao, who's the Vice President of Global Business Development at VAST Data. Good to see you again. Welcome back to theCUBE. Many times a Cube alum. Penny Tseng is with Product Marketing and Market Development at AMD. Thank you for hosting us here in your executive briefing center, the beautiful HQ here. And Ben Lee is the solution architect at Supermicro. We're going to talk about how these pieces fit together and what it means for customers. So, John, I want to start with you. VAST is an amazing company. You guys were born for AI. Go back to the beginning of the company. What the founding premise was and what's changed since Renen Hallak and Jeff and the founders started the company.

John Mao

>> Yeah. you know the team and you know our story, but in short, Renen Hallak set out over 10 years ago now to build a new type of company, a data company specifically. He had a long history in building large and fast storage systems. And he always felt in his heart that there was an opportunity to really build a new type of system that could unlock and extract insight from a lot of enterprise data. And so when we first started this 10 years ago, There was a lot of innovation that had to go into rethinking how do you build a storage system that could be fast enough and large enough and scalable enough and economical enough to actually store and house all of the enterprise data. But over the years, we've been starting to tap into Renen Hallak's vision, which is how do we now start to use technologies like AI inferencing to extract insight to make that more intelligent for the enterprise as well. So here we are 10 years later, we're working with a lot of the companies you talked about. And your opener was spot on, right? It's about not only building out large new AI infrastructure using accelerators, but how do we also prepare and really focus the data to feed those accelerators as well?

Dave Vellante

>> Now, Penny and Ben, I want you guys to explain how each of your companies fits into the system, right? You have specialization, obviously, in silicon. You guys at Supermicro develop systems. What role do you play generally in the ecosystem and then specifically in terms of taking AI systems and making them balanced and efficient?

Penny Tseng

>> Yeah, so let's go back to the older AI needed. Initially, I think the traditional, I call traditional people thinking is that AI and everything needs GPU. GPU is dominant. GPU is king. I don't want to say no, but I will say yes or no. GPU plays an important role, but most critical roles played by the CPU. And you see VAST Data working with AMD, the EPYC CPU stuff from the two generations. We will have third generation coming this year. The two generations go from the Genoa. We see the CPU play an important role, not just for the storage, also for the GPU operating or orchestrate for our data center. So that's why we see CPU start to be back to the center of stage for our data center environment. And I think later on we talk about more agentic AI. But a great partner like Supermicro, they are very flexible and variety solution, not just for the VAST and for overall customer, especially for storage partner like VAST Data, they can create the all-flash and different generation from our last gen from the support gen four, for Gen 5 and future with Gen 6 as well. And that's when we see a great partner like Supermicro and VAST, they can work together for the platform, perspective from you and data perspective from you to support our end customer to drive the AI innovation.

Dave Vellante

>> And Ben, you have to turn these things into deployable systems. You can't just smash in silicon and storage, right? You have to make it balanced.

Ben Lee

>> Yeah, basically we put things together. So just like Penny mentioned, we're a platform company, but we are also trying to turn the total solution company. So that's why we have a conversation, very common thing with John and the team, right? With VAST back in 2022, 2023. And we launched the Genoa platform back to 2024. And we're going to launch the, we actually just launched the Turin-based platform in March this year. It has been very successful, right? I think one of the great benefits for Supermicro is we build a lot of AI GPU clusters. So we are not just provide a platform, technologies, we also bring a lot of intelligence back to AMD and also VAST to share what we learned from this AI evaluation. And along with the platform, we are going to launch with AMD for H14, which is our latest and greatest platform that I'm pretty sure VAST will use that. So we're very excited about that.

Dave Vellante

>> Okay, so it's comical to think about this, but back in the day, there was a spinning disc, it was the media, it was the high performance, and they used to have the high performance disk drive meant it spun faster. It was still much slower than the electronic speeds working through the rest of the system, but as flash, and of course VAST is all flash proponent, as that media changed, it shifted the bottleneck. You had to deal with that from a performance standpoint. I'm sure it had impacts on AMD, and it's clearly had impacts on storage. What has that meant in terms of that bottleneck shifting, for example, to the network? How are you feeding that network? And I want to get into the economics of all that.

John Mao

>> Yeah, like you said in the old days with mechanical media, there's literally an actuator seeking on a hard drive.

Dave Vellante

>> Or two of them sometimes.

John Mao

>> Yeah, sometimes, yeah.

Dave Vellante

>> It's just really comical when you think about it.

John Mao

>> That type of latency, and you think about how fast the SSD evolution has taken over, all flash, NVMe specifically, it's put a focus back onto the network, like you said, right? So storage is this unique workload in the world of computer science because it has to balance both directionally from pulling data, reading data off of a collection of SSDs, but also pushing it out over the wire. It's very rare to have workloads where you just process it within the machine itself and not actually have to feed that data somewhere else, right? In today's example, it'd be to GPU clusters, right? To different AI infrastructure, you need to be able to feed that somewhere else. So finding that balance. And I think what the SSDs have done from a system design perspective is it's pushed, like you said, it's put a lot more focus on making sure that the network IO out of those storage nodes, those data nodes are also keeping up, right? It's not becoming the bottleneck, which has really happened. And you see this constant back and forth going between it's squeezing the balloon, right? you squeeze the balloon on the SSD side, then it's the network that becomes swollen and bottlenecked. You do the other side with the network, and then the SSDs become the bottleneck. And so it's a constant balancing act that we see, which makes this all fun, right?

Dave Vellante

>> Penny, what happens when the GPU has to wait on data? What role does the CPU play in terms of feeding that GPU?

Penny Tseng

>> So go back to the GPU architecture. So GPU still need CPU for the orchestrate for the data direction. And we know the GPU pumps so much data, so fast, but how CPU can orchestrate, how the data go for more efficiency, that's I think CPU playing a critical role, right? The second thing I think the CPU evolution gen by gen is based on the memory bandwidth from the DDR5, from our Milan generation to the Genoa. So we see the memory bandwidth improve. So that can give a GPU more extra the balance or margin. And another thing that we see, but we see the I/O die change. That we see the PCIe Gen 4 to PCIe Gen 5. And our next gen, we see the PCIe Gen 5 to Gen 6. That's why when John talked about how we support networking, we see the networking from GPU generation like two or three years ago to about 400 gig. and today we talk about 800G (800G Ethernet) and I think tomorrow next year we talk about 1.6T (1.6 Tb/s Ethernet / 1.6T). So that means you require a lot of PCIe bandwidth and Gen 6 played a very critical role. It gives better bandwidth for 800G (800G Ethernet) and future 1.6T (1.6 Tb/s Ethernet / 1.6T) as well. But even now, in the list, I think this winter around November timeframe AMD is launching in the market, you will see the first PCIe Gen 6 CPU from x86 going to the market. And that can help the GPU accessing the right data. It gives you extra data bandwidth. That's number one. Number two, for storage perspective, we can see a storage vendor like Flash Vendor that can start deliver for the PCIe Gen 6 for storage, for NVMe SSD. Plus, you get in the back, you have PCIe Gen 6 support the bandwidth for 800 gig. So your data will be not stuck in because the traffic is not enough, bandwidth is enough. And that's the most important, that Venice is the first with 16 channel memory. So that's huge memory and cache, help data buffering. And that's CPU play important role for today. You see the AMD keep on talking about CPU is so important for today because CPU give extra, I call it memory bandwidth, and that's move data faster compared to gen to gen. I would say every two years, we see this data moving, it's double, AMD CPU can keep on increasing performance, give a double bandwidth to catch up for GPU needed. That's why I say CPU back to the center stage. When GPU push so much data you cannot handle, CPU can manage a fast data pass through and get faster to move in there and get your GPU more efficient, that save you more cost, and driving AI.

Dave Vellante

>> Thank you. And Ben, you have to architect for all this. We heard Lisa Su, I think it was in the last earnings call or maybe it was some other conference, she said that the ratio of GPU to CPU was 8 to 1, and then it dropped to 4 to 1. She predicted we could even see it go to 2 to 1 or even 1 to 1. And I think that's generally the consensus now. So what is that relationship between the performance per GPU, capacity per GPU, the CPU ratio? Where does that all fit? Can you explain that dynamic from a systems perspective?

Ben Lee

>> Yeah, I think this is a very dynamic question. But first I want to just echo what the evaluation about SSD. So when we started to engage with VAST, we feel like what Renen talked about is all-flash, it's back to like 10 years ago, it's actually a very, very bold move. And it's happened I think you might have that mentality that we're always trying to be the first mover when we believe something. So we actually have E1L platform, from all the way back to maybe 2014, 2015, right? But it did not work out in the end, but still we continue to develop our petabyte-scale solution. So that's, and with more agentic AI happening now it's proved that essentially SSD is much needed for this data ingestion, data tier platforms, right? But then because you have so fast SSD, right? So that's why you definitely need a high-speed CPU. Because the CPU can really not just be the bridge for all these PCIe lanes for the SSD, but also for NIC card, and also provide sufficient powers, computation powers to deal with this workload. And on top of that, you can still see the memory bandwidth increase from the 4800, 6000, and now it's going to be 8000, right? So we try to work very closely with AMD and also VAST to tackle this problem and always try to be the first mover for the products. And back to your topic, in AI inference field, both capacity and performance are very important because the main capacity driver is the KV cache. You know, we all know this is the big problem in the world. Just look at how crazy the memory and the SSD prices are. But in addition to that, you also want to really keep a very good speed, high performance. You can feed the data into the GPU, the AI factories, right? So you want to keep your AI factory busy. So it's very dynamic, but I think the best approach to tackle this challenge is really have this great engineering conversation. All three parties, we really have constant conversation to talk about how to address that. And if you talk about a ratio, I actually feel the CPU ratio would be a bit more than the GPU. Because the GPU is more powerful and has more memory within the GPU. But the CPU still plays a very important role for these orchestrations. GenAI tries to fit all the data into the GPU.

Dave Vellante

>> So what exactly is the CPU doing? I want to dig into that. But before I do, I want to take a little side trip because Ben mentioned the supply chain challenges. The NAND fabs are all giving way to HBM and DRAM because it's so much more profitable. So that puts a real challenge on you guys to really drive efficiency. What are you doing specifically to make Flash more efficient? we're seeing Seagate going through the roof and there's only a couple of spinning disk vendors left. but they're at a disadvantage relative to flash. So if you can be more efficient with the flash capacity, how are you doing that?How are you dealing with that?

John Mao

>> you're spot on. there is definitely a supply chain issue right now when it comes to NAND for all the reasons you mentioned. We were just on an internal call yesterday. We have an example of a story for a customer recently from VAST where literally we told the customer, we did the math on what we do, and I'll explain what we do in a second, but our solution ended up being what we call now cheaper than free, right? Because of the increased price of NAND, a lot of the innovation that VAST brings is within our software. And so when you think about things like how we handle data protection with our unique erasure coding algorithms, how we do our own patented similarity data reduction, which is a unique form, a very, very aggressive form of doing global compression. The net benefit of that is you just need less raw NAND to hit whatever type of logical capacities needed for the workloads for the customers. And so usually in a non -supply chain tight situation in the industry, that's a nice to have. it's a, yes, we can give you benefits, but if NAND is cheap, then okay, the marginal savings is marginal. But now when it's exacerbated and you see this huge hockey stick price increase in NAND flash, every single if we're talking about being able to save 50%, 100 % in terms of using software to do data reduction, the gains, right, on the hardware cost savings alone for SSDs is massive. It pays for the software. It's why we say it's cheaper than free. So yeah, those are some of the innovations we've been building on. This goes back 10 years ago when Renen started the company, was how do we build an all -flash exabyte scale system that was economical enough? Because back then, people thought that's crazy. Nobody would do that. And now you fast forward 10 years, and you see the rise of AI and some of these high -performance workloads, and it's a necessity. And so we're in a very good position. And we see now, we just published another press release the other day with another service provider. but we're doing true exabyte deployments on the regular, and these are all flash configurations.

Dave Vellante

>> And Ben, you mentioned that you actually see the ratio of CPU, GPU favoring CPU over time. So what's specifically going on? There's obviously data reduction. I think John mentioned that. You've got other copy services. You've got orchestrate. From a system perspective, what is that CPU doing? Why is it so important?

Ben Lee

>> I think it's just because GPUs mainly do the math, right? all these AI, inferencing. But for the CPU, you have to do a lot of things, just like what John mentioned, the data reduction, even the very typical erasure coding, things like that. And we, I think from our perspective, we are, we expect we're going to see more workload running on the CPU, again, just because of the agentic AI. Right? And to John's point about the SSD price go through the roof. And our observation is guess what? but people still buying that. I think it's just because all the investment in AI, people has been very optimistic about the AI future. So even the SSD price is going like that crazy, they still like to buy it because they need to build this complete AI cluster for sure because without the data tier platforms, there's no way you can keep the GPU busy. On top of that, I think the best way for us is like, We need to try to optimize the platform, how to optimize solution. So that's required three parties have a partnership, right? We try to find the best sweet spot to do optimization, right? And I think even on power efficiency, also can save some great dollars for the whole data center. So these are everything we can work together, but we always want to work in this partnership.

Dave Vellante

>> So Penny, the relationship, let's get into the partnership.it goes back well into the last decade, but let's start in this decade. Let's go back to Genoa, which I think is probably 2022 timeframe maybe, but Genoa, Turin, now Venice, and of course H14. What has changed across these generations that's tangible that customers can see and take us through that roadmap?

Penny Tseng

>> Yeah, so if we see the AMD roadmap, we're doing a little bit tick-tock. So Genoa was our first Gen 5 tech and that's when we did the first DDR5 solution. And that's what gave us PCIe Gen 5 at the same time. And then Turin, we improved the memory bandwidth because we know like VAST Data, they say, hey, we need more memory bandwidth because I need more cache inside a box, the EBox and CBox. So when we start to design every silicon, we're not thinking only about how we improve the core performance. When you think about our ecosystem partner, right? for the first one, we think about Supermicro. When they develop the platform, how can we consistently build in a 19 inch server and still manage certain power and certain power efficiency and the performance balance at the same time. And at the same time, want to give us best performance for the CPU to manage for data orchestrate at the same time and give it enough PCIe bandwidth, give it extra memory bandwidth, the margin. So that's why we designed Silicon, not just thinking about Silicon itself. And from the Genoa to Turin, we see gen-to-gen change, and we benefit VAST Data because they're driving so much AI storage, at Supermicro developed a great platform for support for KV cache or database storage. And we see that the adherence by the different CPU generation, and going to the next-gen, we see the performance improve for next-gen. for future GPU like AMD Helios come in. We need more KV cache storage from VAST Data for sure. And we need to build more H14 for us for the Helios. But not just this way. When we talk about CPU importance, KV cache is very important for the GPU. But think about when the GPU calculates everything, finish the inference, you go to the agent, you need tools. Tools need the CPU running. Again, we go back to Supermicro and VAST Data. They still offer database storage. Don't just think about everything. I know AI drive you guys crazy and us crazy, but database storage, I call it fundamental for our infrastructure. We also did a good job like 10 years ago at Supermicro building a platform. So that's why you see so much CPU adapt and not just KV cache. People think about KV cache because our competition only mentioned about a portion, but we not mention about KV cache only. We think about the database storage. And believe it or not, VAST still did a strong job for that one. And that's what we see when we design Silicon for our partner together, we think about not just KV cache and the whole entire ecosystem for the AI workload. I think that the CPU plays a critical role. And that's why I see the one-to-one ratio coming out.

Dave Vellante

>> Interesting, so when you think about H14, Ben, as Penny just said, the CPU is orchestrating access to all these tools. You need the CPU to manage that. Workloads are shifting very dramatically. you're close to the customer, you have to think about, okay, as the customer needs change and the workloads need change and the profiles change, you obviously have to design a system for that. And you also have early qualification. It's clear you guys do some serious engineering.

Ben Lee

>> Yes.

Dave Vellante

>> I'm sure a lot of back and forth. But how does that early qualification map to where the customers are? And what does that mean in terms of time to market?

Ben Lee

>> Yeah, I think that at least between Genoa to Turin, we see a 40 % at least performance bump, right? So that's a great indication that keeps our engineers and also their engineers very excited, right? So people, especially from VAST team, they start to really chase for early sample validation. So we work very closely with AMD. We try to first start with some review on the paper, try to understand what's the benefit. And even if there's some architecture change, what's the implication for the software. And then we are, because we are always the first mover, and thanks to VAST, now they are chasing me to be the first mover. So we definitely be ready to have this platform launch, while AMD is launching that. But the partnership, I spoke to the engineering almost weekly, and we have a weekly sync up call, and we try to understand from different perspectives what does that mean? Because I think that software also needs to be ready. Is there any new architecture they can leverage from the Venice? So these things that we can help. And on top of that, we also help them to really make assessment about the future key parts. For example, the DDR5 memory 8000, which is pretty critical to the future SSD the Gen 6 800 gig NIC, things like that. So these are all the things that Supermicro as a platform total solution provider, we can provide to the VAST. And they have been a great partnership, right? They bring a lot of good business with us. So we definitely want to be, we want to treat them as a VIP for sure.

Dave Vellante

>> So we always treat them as VIPs. Well, so obviously Venice is important for storage because it gets you PCIe Gen 6 support. But what else? How does it enable faster networking? And what happens to the balance of the system? How do you prevent the bottleneck? One of you said the balloon squeeze. I think it was you, John. Maybe let's have that conversation.

John Mao

>> Gen 6 is obviously necessary to enable 800 gigabit NICs. And so that's about getting IO out of the box, right? And so if you're running and there's always a lag because you also have a transition period going from PCIe Gen 5 SSDs to PCIe Gen 6 SSDs. And that doesn't happen overnight because you have a very mature, complicated supply chain. So this squeezing of the balloon with Venice pushes a lot of that back onto the SSD part of the componentry to ride that wave and wait for that curve to tip over from Gen 5 to Gen 6 for SSDs. Gen 6 drives come at a premium today, right, for example. But today adding 800 gig helps tremendously because it allows us to squeeze out even more performance from those Gen 5 SSDs at an aggregate level, at a system level, at a cluster level. Right. But I think beyond networking, the important thing, like Ben said, we see from our side, from the software side with the move from Genoa to Turin and now to Venice, we're seeing somewhere between a 30 to 40 % bump in performance, which is huge because what that means for our customers at the end of the day is a much, much better TCO when it comes to our systems in their data center. And the thing we haven't really talked about is that TCO is not just about cost. Obviously, there's price implication. But today, the bigger challenge we see is really around power, the limiting factor within most data centers. How many rack units do you need to take up? How many racks do you need to take up on the floor space? How much power is required to drive a certain amount of capacity or performance from storage systems? Those are all equally important. And then to Penny's point, many people may not know this, but VAST does a lot more than just high performance storage. We support database workloads. We have our own database, our own data warehouse that we built. We have our own streaming engine, right? All of these things are very, very compute oriented. So by moving to the latest and greatest and by riding the curve of the innovation that AMD at Supermicro bring to the table, we also get a lot of those benefits that we can pass through for those capabilities in our software to our customers as well.

Dave Vellante

>> So what does this mean for customers? how should decision makers in IT think about the economics of AI generally and storage specifically? You mentioned a fixed power envelope, so that's obviously important. But what should they worry about? Is it throughput? Is it performance? Is it GPU utilization? Obviously, TCO is important. What are customers telling you that matters to them from an economic standpoint? Maybe we can go around the horn.

Ben Lee

>> Yeah, I think it's more from the TCO perspective, right? Because all the customers even not just in NeoCloud they've been doing the AI cluster for years. But the enterprise, they really want to embrace the AI, but they're still in the learning step, right? So they want to understand from a TCO perspective, how they can leverage all the solutions, The VAST Data platform not just on high -performance storage, they have different workload and using Supermicro platform and maybe leverage the needs as a complete like AI clusters for the enterprise sectors. So I think one key factor is like you have to, the CIO needs to make a fair assessment about what's their data structures, understand the data. and I think that's where VAST can help a lot and try to see how they're going to embrace the AIfrom different perspectives and then I will always say that don't just look at buying more GPUin the past two years ago I said, oh AI means buy more GPU. I think that's changed now because all the agentic AI people realize that the data is actually the core for AI. You know, if you need to run AI you have to understand your data structures, your data platforms. So start with that and come to talk to Supermicro, talk to VAST, and even AMD. We can help you to architect the right solution, right platform that you can start with. Anything you'd add?

Penny Tseng

>> So you see, before joining, I did build a data center. And back in the old days, people, the IT decision makers always think about how much per node, how much per rack I need to pay, and how much power I need to pay. So people always think about hardware cost, but I think John's point is when you go into 2026, in today's world, yeah, that cost still going crazy over the roof, but that's how AMD talk about performance per dollar. You pay this dollar, but you get a certain performance. I think most important, your cost, your pure hardware cost for sure you'll be improved, but how much performance you gain. So performance per dollar or performance per watt is how we always emphasize. So every gen EPYC CPU we want to talk about our competition and talk about our leadership. We truly believe the cost you spend, you need to get certain performance. That's what we believe. So today in AI world, people think about GPU punch and we need more GPU running for more data. But no one thinks about how we store your data, how we structure data. What's the right platform to structure data? John said, they have their own data warehouse. But how to help your customer to structure your data and make it more efficient. When your GPU keeps on giving so much data, which I don't want to say useless, but sometimes you cannot use because it's too quick and you don't structure well. Today's CPU vendors like AMD, we're working with Supermicro and VAST, we tell the customer how we orchestrate your data, how to manage your data, make more efficient. That's how we see the gen-to-gen transition, right? From Turin to Venice. We're not just giving the bandwidth, we give a better performance. John said, we see the 40 % improved for the bandwidth, but think about how much the CPU core performance. So we don't give you so many CPUs today. You can see the power efficiency that's how AMD always every generation we will try to improve cpu from our last gen Turin we use a TSMC 3 nm process (TSMC N3 / N3E) and 4 nanometer based on the Zen 4 and Zen 5 and Zen 5c but go to the next gen we start moving from TSMC 3 nanometer to 2 nanometer for architecture so every generation we give the customer the right direction say we give you true technology. We're not giving you the PPT technology. That's really PowerPoint technology. We call it a joke. But every generation we see improvement is true. 4 .3 is for last -gen for 2 .0. For 1 .0, it will be 3 .0, 2 .0, all from TSMC. That's why we see Gen2Gen improve. Again, platform vendor and the solution provider, I love calling it the solution provider and platform working together with silicon vendor. I think that's the best combination for today.

Dave Vellante

>> I want to ask you, please add a comment, but I wanted to come back to you. No, okay.

John Mao

>> I was going to answer your question.

Dave Vellante

>> Yeah, please do.

John Mao

>> I think there's a lot of, there's hard dollars for IT decision makers to think about. And I'm a performance junkie like everybody else when it comes to technology. But I think there's also soft dollars that IT decision makers need. So some of it, Penny and Ben talked about, which is around data readiness. So there's a lot of other things that have to go into success, especially when it comes to AI projects that aren't talked about enough. And they mentioned some of it's around data readiness. It's about how do you structure the right data? How do you provide the data? There's obviously performance angles to that. And then the thing that is the running joke, at least inside of our own company, right, is reliability is such an unsexy, but so critical aspect of everything that underpins what we do, right? We joke that if you could have the highest performance storage system on the planet, but if it's offline, it goes to zero, right? So those kinds of things are not talked about, and you don't see it per se on a data sheet or on a spec sheet for a chip or a system or even a piece of software. But those things are something that people need to think about, especially because these AI systems are no longer sandboxes they're becoming mission critical.

Dave Vellante

>> And I think VAST has established that it's more than a storage platform. You were involved beyond that. You mentioned your own data warehouse, data lakehouse, whatever you call it. You've got streaming capabilities. And you're at the heart of AI. And to your point, yes, there is the cost of moving to a next generation PCIe. Okay, there's a hard dollar capex there, but the faster you can get on the AI curve, the more AI projects you're going to be able to enable, you're going to create a flywheel and that's going to drive a 10X delivery of productivity or else we're all dreaming. And I don't think we're dreaming here. We are entering a new era. I think we all believe that we can see that. But so what are the workloads that you're seeing today that are taking advantage of some of these capabilities that VAST has beyond storage?

John Mao

>> I think that a lot of what we see from our perspective is a lot of this bridge between streaming real-time data feeding into some of these agentic systems. And that's a very complicated pipeline because historically you have to do a lot of data processing to make that data come from a real-time, think about events or sensors or whatever it may be, into a more structured format that then computers can actually process, right? We see a lot also happening in the space of unstructured data. So unstructured data, think about documents or PowerPoints, things that are very common to the enterprise, but more specifically things around imaging or video, right? And you can just think about all the different types of industries that that pertains to. I think I read a statistic that 90 % of the enterprise data is still unstructured. So how do you then apply AI to actually make that data ready for agentic applications as well? Those are the workloads that we're focusing on beyond your traditional IT workloads like virtualization and databases and so forth in the world of AI that's becoming top of mind.

Dave Vellante

>> And there's so much focus on jobs and I think not enough focus on work, the type of work that enterprises are doing. And much of the work that we do in enterprises is, to your point, It's unstructured data. It's tacit knowledge. And the whole software architecture is completely changing. if you think about it today, generally the analytics are in their own stovepipe. And the operational systems, even within the operational systems, the HCM and the ERP and the CRM, they're all in different departments. Those are all coming together. There's a new user surface being created. Obviously, we're talking to LLMs, but if you're coding, you're coding in whether it's Claude Code or Codex or Cursor, other agent, that system is coming together to create a real-time representation, to your point about real-time, of the enterprise.

John Mao

>> It's funny because those examples you just gave are very common. You'll find them across every single enterprise. It's how they built data strategies for the good part of the last 20 years. But it was always built for humans to consume. And so what you don't see between those different silos is the latency that sits between them. You have real-time systems, transactional systems, analytical systems. They're all built for a human consumer at the end of the day. And now with the rise of AI, that has to merge into one giant high -performance system because AI never sleeps. They don't go to the bathroom. They don't eat. They can process way more data than any human could ever do. And so rethinking data strategy is what gets me super excited because most of these organizations, to really see the value of AI, they're going to have to rethink and transform a lot of their data strategy from years past.

Dave Vellante

>> We've all been in meetings where somebody presents some data and then we spend 15, 20 minutes, a half hour arguing about the data. Somebody goes off and asks, what's the true version? Don't need to do that anymore. The future is that will happen virtually in real time.

John Mao

>> That's right.

Dave Vellante

>> I'm interested in your thoughts on the enterprise consumption specifically. We know there's a lot going on in the AI labs. There's a lot going on in the Neo clouds. We have a panel around that. What about the enterprise? Most data centers are not geared toward liquid cooling. There's a lot of air-cooled systems that Supermicro ships, and it's unlikely that the data center is going to overnight transition to full liquid cooling. So there's a balance there. There's a hybrid architecture. There's public clouds. There's neo clouds. There's on-prem. There's edge How are you thinking about that? What are you seeing in terms of demand from the enterprise again? I'd love to go around the horn and what it means for each of your companies specifically, but then also generally how you work together

Ben Lee

>> Yeah, I think we did we have been doing great in the NeoCloud space, but I think the enterprise is the next, our next focus I think just like what John mentioned, when enterprise seeing the power of AI and when they see they put structured data and unstructured data together into high -performance data platforms, they're going to see a lot of workload needed. So I think from a market perspective, we always be the first mover and we always, for example, the H14 platform going to launch, it not just be one platform, we have the phase one and the phase two. So this approach we're going to have is like we're going to have different platform that's meet different workload for enterprise need. Not just on the GPU server and it's air-cooled, liquid-cooled, right? And also the storage data platform or very basic computational 1U, 2U servers. So we're going to provide this complete product portfolios to the enterprise sectors, right? And on top of that, we work with the solution partners like VAST and also other partners, try to tackle this workload. For example, the video. Video might need some special tool or solutions to address the video workloads. So this is the thing. We should, for my goal, try to become the total solution platform company, and we try to serve the enterprise better.

Penny Tseng

>> I love this topic. My team is called Enterprise Business Group for AI. And that's why I really love to talk to the enterprise. My personal manage for the OEM platform, a Supermicro. So in November we're going to launch called the Venice, the codename Venice called SP7 based on the TSMC 3nm, 2nm for Zen 5 and Zen 5c. But in the next year, it's coming out here in first quarter, end of first quarter, early second quarter, we have another Venice platform called Venice. same architecture, but small silicon and more affordable, less power consumption, and still give you PCIe Gen 6 by the way, so enterprise still can enjoy PCIe Gen 6. You still can get DDR5-8000, we call it the Venice SP8. We're going to launch next year. So from silicon perspective, we have an HPC, high performance GPU hosting for platform SP7, going to launch in November. That's what we're going to talk about the embargo for July. Another one is the Venice SP8 coming next year, smaller silicon, less TDP, still give tons of PCIe bandwidth Gen 6. So Supermicro and VAST still can build an enterprise storage box to support them. And because we believe that enterprise, not everything go to a super HPC or supercomputing performance. You need okay performance, right? But we don't want to sacrifice the old generation of silicon. So we don't forget enterprise. We give the same quality silicon, just smaller. And this core comes from silicon to support enterprise. That's from a hardware perspective. For silicon, the second portion we think about enterprise is toward on-prem. Today, they cannot build in a data center, a neocloud. They cannot go to neocloud, say, hey, I want to build my agentic AI environment, can you build for me? They cannot afford it, to be honest, for overall enterprise. And maybe Fortune 100, they can, but the top of Fortune 2000, they cannot. So when you talk about this Fortune 2000 customer, how we help them to get to similar AI infrastructure. So the building, the Supermicro building, air-cooled solution, they still go to co-location. You don't have gigawatt data centers for those co-location. You only have like one megawatt, two megawatt data center. Do they need AI? Absolutely, right? So you still need a compute platform. There's still a storage front of us. They still need Supermicro building that platform for them together. So on-prem AI, I think before we all talk about enterprise, on-prem data, but I think the next year we talk about more enterprise try to adapt AI. They want their own on-prem AI to manage the tools, manage the database. They still can lease the compute power from whatever neocloud or public cloud, but they can drive their AI, protect the data on their own. I believe in the next year, I'm very excited to see the VAST and so on, I can create more and more variety of enterprise platforms for them to run on-prem AI. That I think needs a new term, but on-prem AI for enterprise sounds very cool for them.

Dave Vellante

>> Well, John, Penny mentioned supercomputer class infrastructure and scaling that down actually makes it more cost effective for enterprises. I was at Supercomputing in St. Louis last year. I attended a VAST breakfast. And of course, the audience there was very supercomputer heavy. So you guys are doing great in that world, great in NeoCloud world. But also, enterprises are building their own on-prem AI stack, to Penny's point. And many, if not most of them, are evaluating VAST because you're built for AI. So what are you seeing in that space?

John Mao

>> Yeah, I think for most of the enterprise market, it's still pretty early days but I've been working with a few customers um I mean there's some trends that I see at a very macro level for example a couple of um well one example of vertical that's already leading the pack is in biopharma right so they are already able to buy like they're they found the ROI they found the you know the workloads that fit uh makes sense with their business and they're leading the pack and you know deploying thousands of GPUs uh which is no small feat, right, for an enterprise that this is a very, very big change from general computing, right, in the past. For other enterprises, whether it's financial services, right, retail, there's a long list of AI applications. Most of it, though, the value that you're starting to see in all the studies about how some AI projects have stalled, most of that, in my opinion, in conversations is always attributed to where does the data in which the AI has to act upon reside, right? And so you have this choice. You have to think about either bringing AI and compute to the data or you've got to bring the data to the compute. There's only two ways to skin that cat, right? And one way is more efficient and more cost -effective. Yeah, and so I think there are a lot. What is interesting, I just learned about this the other day, right? The whole SP7, SP8, two different packages of Venice CPUs. I think this really shows a broader market movement, which is we're going back to a time where you have more specialized compute for different workloads versus we probably in the last 15 years we went through an era of just general purpose computing where it was the era of VMware and virtualization and trying to shove everything onto something that looked the exact same because you're going to virtualize it anyways and we're now this constant yin yang of going back to a market and a mode where, well, you actually need specialized compute for all the reasons you mentioned. You can't fit liquid-cooled systems into traditional legacy enterprise data centers. Not going to happen. You can't retrofit those very easily. So then the choices left are you have to either workload optimize, or we do see also enterprises now thinking about partnering with neoclouds, right? But there you have to start to dissect the application. There are some things that have to do data processing on-prem using maybe smaller CPUs or GPUs and then use the big hitters, the big GPUs for other more agentic computing workflows. And so hybrid, I think, is going to be a real part of that strategy moving forward for most enterprises. But there will absolutely still be AI infrastructure deployed on-prem. It just won't look like what I think probably we all think about when we think supercomputing class AI. It'll be a different AI computing on-prem.

Dave Vellante

>> And you mentioned some of these projects have stalled. I look at that as progress because you're going to hit a data wall. Anybody who's ever built this stuff, they're all building agents. And it's just not as trivial as the podcasts make it sound. You have to learn and take your hits. But that, to me, is all progress. The other thing you mentioned is AI is a game changer for biotech. Biotech's been in the tank. I'm from Boston area, and it's like the heart of biotech. And people are so excited now about the potential because it just accelerates discovery. And I think you're going to see a resurgence in that industry. Okay, last question. We'll go around the horn. The one action item for IT decision makers that you want the audience to, we want to leave the audience with. We'll start with you, Ben.

Ben Lee

>> Yeah, again, I think it's obviously, everybody is embracing AI, right? So I think the companies, enterprises, different sectors, they have to try to understand how they're going to embrace AI. So, I would say just come talk to Supermicro, just give us a conversation. And we are here to help all the customers to bring the AI to the realities because we have so many platforms, we have so many products, we always launch the product as the first mover. And on top of that, we are a pure engineering company. We build everything in -house. We try to help customers to integrate everything together. And John mentioned about the abilities. From hardware perspective, we are also very important. And we have a proven record to build the largest cluster in the world. And so we believe we can leverage this experience and our experience in the AI space to share to everybody. So, great to have more conversation with our audience.

Penny Tseng

>> So I think for the IT decision maker, you're thinking about two years ago, ChatGPT come in, get crazy, people talk about GPU. I think that's great driving AI, but you see that today, 2026 next year, future couple of years, we found the generative AI to the agentic AI. And the agentic AI is not just a GPU running the performance for power, crunch so much data. The agentic AI helps you execute your tools and make AI really work. Today you see the generative AI is more like Q and A style, but tomorrow we want the data really work for us. But when those agent people, agent tools, these agents work for us, for humans. So I believe the AMD, the EPYC CPU, whatever our SP7 or SP8 will give a variety selection. That means in all your agentic AI infrastructure, whatever you are super compute gigawatt data centers or megawatt data center, we have a different selection for customer. I think that's the best part of working with partner. Now we don't really normally say, hey, this is my solution, just follow it. Working with Supermicro and VAST, they can create a different value to fit our different type of customer environment, not just gigawatt data centers. When I talk about enterprise, they can build a solution for you. And we have silicon ready to support. So back to the agentic AI, I think AI is not, I'm not called as a billion people to own ownership. We want more coverage for everyday people can use it. And I think AGI can be changing the world. And EPYC CPU gives you more options working with VAST at Supermicro. And you can run your agentic AI. You don't need to think about the billion dollars. You can run in the stack a couple hundred thousand to a million dollars running your own data center. I think that's what the EPYC CPU can offer the best value, variety selection and strong partnership with VAST at Supermicro.

Dave Vellante

>> You can get started John.

John Mao

>> I'll be biased and just say again that I think I would challenge the IT decision makers to put in, I won't even say the same amount of thought, but a good portion of thought into understanding what their data strategy is. A lot of our partnership stems from helping make a lot of the infrastructure decisions more seamless and more transparent and hopefully we've done a good job in serving our customers so they no offense but they don't have to know who AMD is but we use the best technologies under the hood so that they can just solve a business problem but i think the reality is that data for most of these enterprise customers it is the one biggest thing and it is shifting i am seeing more and more customers now talking about it compared to a year and a half ago two years ago, but I would implore them to continue to think about that because that is going to be their measurement of success when it comes to AI.

Dave Vellante

>> Excellent.Well, Ben, Penny, and John, thanks so much for spending some time with us here. I appreciate it. And I'll leave you with this thought. One of the things we talked about in the panel was the value. Yes, there's an expense, there's a CapEx, but when we went from on-prem software to SaaS in the cloud, everything changed. The operating model, the technical model, the business model changed, and it's happening again. And so my advice to IT decision makers is you've got to get on that AI curve. You've got to learn, as John said, think about your data strategy because that's going to be a fundamental part of getting on that AI curve. But unlike previous transitions, like I mentioned into SaaS, that affected the IT department. It didn't affect the entire company. This transition is going to impact your entire organization, every department, the way you're organized. And so you need to think about that. That's where the value is. Of course, it's all about the applications that are running on top of this infrastructure that we were talking about today. And this is the enabler. Now these power panels that we're running are all a part of supporting Advancing AI, which is AMD's big event in July. theCUBE will be there. So check out all the coverage on theCUBE.net and social media and SiliconANGLE.com and their respective websites of companies including VAST and AMD and of course Supermicro thank you for watching and keep it right there for more coverage from theCUBE