VMware Explore 2024 | Chris Wolf & Ram Velaga | Broadcom

Clips
News
More from VMware Explore 2024

Chris Wolf

Global Head of AI and Advanced Services, VMware Cloud Foundation Division

Broadcom

Ram Velaga

SVP & GM

Broadcom

Network performance and IT economics help drive VMware’s private cloud strategy

A key focus of VMware by Broadcom at the Explore conference in Las Vegas last week centered around how enterprises can power network performance to drive GPUs and necessary data center infrastructure and deliver artificial intelligence.This need has been highlighted through an ongoing discussion within the tech community around the merits of various networking protocols, such as Ethernet and InfiniBand, to power high-performance computing. Key executives at Broadcom have made it clear that Ethernet is the way to go.Broadcom’s Chris Wolf and Ram Velaga talk with theCUBE about network performance

play_circle_outline Maximizing AI Service Value through Customer Validation and Proof of Concepts for Optimal ROI

play_circle_outline Private AI and networking

play_circle_outline Distributed GPU computing over Ethernet

play_circle_outline Exploring the Evolution of AI: Capacity, Network, and Memory in Private and Generative Models

play_circle_outline Navigating Private AI Challenges: Interoperability, Controls, and Implementation Hurdles

play_circle_outline Networking bandwidth scalability and challenges

play_circle_outline Unlocking Efficiency: Architectural Integration and Orchestration in Private AI Cloud Implementation

play_circle_outline Focus on customer trust and reliability in the industry.

Info
Transcript

Chris Wolf & Ram Velaga | Broadcom

Chris Wolf

Global Head of AI and Advanced Services, VMware Cloud Foundation Division Broadcom

Ram Velaga

SVP & GM Broadcom

The Venetian Conference Center in Las Vegas hosts VMware Explorer 2024, focusing on private AI and network Ethernet, customer challenges, and use cases. Ram Velaga and Chris Wolfe from Broadcom discuss the significance of Ethernet in AI deployment, particularly in connecting GPUs for distributed computing in large cloud deployments. They highlight Ethernet's scalability and performance in handling large GPU clusters as key factors for success. Chris Wolfe mentions cost savings of running AI services on VMware Cloud Foundation compared to public services, with... Read more

explore Keep Exploring

What feedback is being received regarding the cost savings and momentum of running AI services on VCF compared to public AI services? add

What have you been seeing out of the hyperscalers in terms of utilizing Ethernet for frontend and backend networks in relation to private AI applications? add

What network is typically used for the backend of GPU interconnect in large cloud deployments? add

What are the two main classes of customers that Chris, the private AI expert, is currently working with, and how is the evolution likely to occur in terms of their adoption of AI technology? add

What are some of the challenges that customers are facing in regards to AI hardware accessibility and interoperability? add

What is the current status of networking bandwidth scalability and what improvements have been made in recent years in terms of chip capabilities? add

What is the thesis on private AI and the importance of having an integrated solution for private cloud infrastructure? add

What analogy was made by a colleague at the financial analyst meeting earlier this year, regarding the success of teams who have stayed on at Broadcom? add

bolt Powered by CUBE AI

Chris Wolf & Ram Velaga | Broadcom

search

Dave Vellante

>> Hi, everybody. Welcome back to the Venetian Conference Center in Las Vegas. We're here covering VMware Explorer 2024. Dave Vellante with Rob Strechay. This is our 15th year covering VMworld, VMware Explore. We're super excited to be here. Two sets we've gone all week. We're wrapping up day three. We've been talking about private AI. We're going to talk about network Ethernet. We're going to talk about customer challenges and use cases with Ram Velaga and Chris Wolfe. Chris is the Global Head of AI and Advanced Services for VMware Cloud Foundation Division at Broadcom. You're in the right place, Chris. And Ram Velaga is the Senior Vice President and general manager for Core Switching, the group at Broadcom. Gents, welcome to theCUBE. Good to see you again.

Ram Velaga

>> Thank you for having us.

>> Always great.

Dave Vellante

>> Been a good three days, and you guys are still going strong. There's a lot of people here. Are you hearing from customers, Ram? What's the vibe like?

Ram Velaga

>> Oh, so we are from the silicon side of Broadcom, so generally I hear from customers, "What are you doing here?"

Dave Vellante

>> Well, what are you doing here?

Ram Velaga

>> What am I doing here? So I'm like, "Hey, guys. You know what? Let me explain to you. When you start thinking about AI, it's all about distributed computing. And when you're doing distributed computing, it's a lot of GPUs and you have to connect them together with the network," and that's what we do. My team basically does networking, and we work very closely with the very, very large cloud providers. Previously they were deploying CPUs, now they're deploying GPUs, all goes over an Ethernet network. So we look at where AI is going into the enterprise, and we believe it should be built on Ethernet, so we are here.

Dave Vellante

>> Well, we saw you at the financial analyst meeting earlier this year, same week as GTC, which was awesome. John and I were there and course we had Charlie Kowas on at MWC, the whole premise being the world is shifting from a CPU-centric to a connect-centric world, and you guys make all that stuff work together, which is all the hard stuff that nobody wants to do, but if you can figure it out, it's a great business. So Chris, we were talking last night at the bar. A really interesting strategy that you guys are affecting. You saw a piece from a couple of weeks ago. I think you were saying we got much of it right. There were some things that maybe we need to fine tune, but what are you hearing from customers at this event? How is the message resonating?

>> Yeah. If we want to just lead with price, tremendous validation for what we're doing, recognition that running AI services on VCF can be a considerable cost savings as compared to public AI services. We've had a number of organizations look at us as next steps. I was meeting with the CIO of a state last night and our conversation was around what specific GPUs do I need for my use case? They're ready to go. We have a number of POCs. Coming out of the show, at least two dozen POCs we're going to run with customers. Just tremendous momentum for what we're doing and real excitement around this whole suite now of advanced AI services that we're running on top of our platform. Plus, everything that we're doing with our infrastructure scheduling and optimizations have been a real point of excitement for customers too.

Rob Strechay

>> I got to see you guys earlier today and I think, again, one of the things I learned a long time ago, back when I was a networking nerd and things like that, in IT for that matter, was not to bet against Ethernet. I remember ATM to the desktop was a thing at one point in time. It reminds me of InfiniBand and in fact, funny enough, we were talking about InfiniBand and I've actually run an InfiniBand network, so I probably was one of the few people in that room who's actually done that. I think that it makes total sense that it's Ethernet, and I think you had some pretty astounding statistics that you had seen around just how many connected devices across Ethernet. And for me, where it ties in with private AI, and I want to understand this and get your view on this, is that there is the frontend network and then there's the back-end network. That is, the east-west traffic and the north-south traffic, so the applications can take advantage of the AI as well as moving the data around and connecting things together. Go into what you've been seeing out of the hyperscalers because I think, again, you had some pretty amazing numbers there. And even if you can't share the exact numbers, I think when people look at how they're building these networks out, it was pretty amazing.

Ram Velaga

>> Yeah, so when you look at GPU architecture, when you're doing machine learning, as I said before, anything related to AI, generally the workload cannot be run on a single GPU. You need a collection of GPUs together. And generally, when you think about how this works, you probably go to ChatGPT or some place. You type your question. The question that's typed goes to the CPU. The CPU takes that question, breaks it up into a bunch of tokens, words, tokens, and those are digitized, and that that information is then fed to a collection of GPUs. That's how inference works. Now, prior to inference, obviously you are doing training, training of large language models. And when this happens, essentially what's happening is there's a lot of parameters that are being exchanged between those GPUs. That exchange of parameters between those GPUs is usually what you think of as the backend network, the GPUs talking to each other, exchanging the parameters, coming out with these coefficients for these regression equations and so forth. All those matrix multiplications that happens. How you access those GPUs, how you push data into it in the first place, is that what do you think of as the frontend network. So there's a misnomer that is, oh, Ethernet is only used for the frontend network. All this backend, which is significantly higher capacity is based on something other than Ethernet. But the reality is, when you think about the largest cloud deployments today, whether it is Amazon, Meta, Google, Alibaba, ByteDance, Tencent, Oracle, all of the backend of their GPU interconnect is based on Ethernet. Not just the frontend. The backend. And these are not clusters which are small hundreds of clusters of GPUs per cluster. These are 10 to 60,000 plus GPUs in a cluster, and these are all publicly stated numbers. The numbers are actually much bigger than 10 to 60,000. I can only point you to publicly available information, but very large clusters all connected over Ethernet and doing very, very well in terms of performance as well as the economics.

Dave Vellante

>> Well, I've spitballed out there saying that the big internet companies, they want to get to a million GPU clusters. That's where they're headed. And it was interesting listening to Charlie talk about the ROI of the consumer AI, very clear. You build bigger clusters, you're going to be able to target better and you're going to be able to print money In the enterprise, I say the ROI is chatty. It's little singles that people are hitting right now. Tech summarization, maybe do some marketing copy. Chatbots, obviously. Code generation. Nice things, but the really interesting stuff is yet to come. theCUBE Research put out the gen AI power law talking about really where private AI, is where I'm getting to, as things that are going to be done on-prem. They might be small language models, they might actually be larger language models, but it's going to be private proprietary data sets that actually solve bigger problems and drive bigger ROI. So Chris, private AI, big, big thrust around that. Help us understand what customers want to do and how that evolution is likely to occur.

Dave Vellante

>> Yeah. I think there's two classes here. So you have the customer that has been running what they would call ML. So if we rewind, we used to call this ML-

Dave Vellante

>> Legacy ML....

Dave Vellante

>> not too long ago. Right. They've been doing this for years. And what they've seen us come along with our platform and how we can do distributed resource scheduling at scale for them, we're solving problems that they have not been good at with their own DIY approaches on bare metal. So that's one class of customers, so we're already converting them now to our platform and we're seeing stronger than expected traction there. Then the second set, this is to your question more on the generative AI side, they're looking for their first successful use case. And we're helping to guide them into what that is, because what they want to do is pick something that's not just in the abstract. Give me something where I can measure the actual business value. And the easiest thing and the most repeatable thing that we see across every customer segment is really your call center, the contact center use case, because I know in a given week this is the number of tickets that my agent closes. When I bring in generative AI, I might see a five to 10% efficiency gain in terms of how many more tickets they're closing. That's a measurable number that adds value to the business that's really cut and dry, and that's what we're seeing is the first place that they start. And this is something where they can start relatively small with their private data, so it might be one or two clusters to begin with, and then as they continue to onboard employees, then they can just continue to horizontally scale that. That's been a very common pattern that we're seeing, and we're going to continue to repeat that, I would think, for the next couple of years. Other use cases are catching up, but that's where we're guiding people because we want to have a successful result and we want them to be able to prove business value.

Dave Vellante

>> What are the headwinds that you face? We write a lot about privacy and the legal compliance. Those are well-known. I presume you're seeing that. What else are you seeing that customers are telling you some of the challenges that they're facing?

Dave Vellante

>> Yeah, the challenge is still the accessibility of hardware. I still need to be able to get enough GPUs to run some of these services, so that's key. There's still a fragmented ecosystem. Customers are asking for interoperability, but it's easier said than done. You don't have a truly interoperable software stack today. It doesn't exist. There's parts that exist, but you don't have that true end-to-end solution, but it's not stopping the AI applications, and that's why we've led with Nvidia because there's a lot of trust for Nvidia hardware. We're having good success getting customers reaching successful outcomes on their GPUs. And the supply chain is what it is, but for some of the lower-hanging fruit use cases, we can get them started on the L40S, and then we can graduate them from there. At the far end, we would have customers that are taking 8x H100 and putting those per node, and definitely in the financial services space, that's typically the design pattern that we're going for. It's like, let's just build this thing the first time and then scale it. One other observation I would say that really matters is we're starting to see this bifurcation where the data science teams might have gone out and bought some DGXs and are using that for some R&D work, but when it comes time to actually operate the service in production and run these inferencing use cases, that's when they're coming to us because we have all the operational controls, we can support security audit, all of those things, and give them that efficiency of being able to share that AI capacity. And I'm saying AI capacity because people focus on the GPU, but that's just one part of it. It's the network. It's memory. There's lots of other parts that are going to comprise that application. So that's where we're seeing things going now, and I think our roadmap is going to get even more exciting where we're looking at more advanced ways to do scheduling and workload placement that's beyond what you're seeing from anybody else today.

Dave Vellante

>> So of course, the spinning disk used to be the bottleneck in the system, and then the end prices came down and the volumes went up and changed all that. Is networking now the new bottleneck and the big challenge from an infrastructure standpoint?

Ram Velaga

>> I would say the good thing about networking is the networking bandwidth has been scaling quite a bit. If you just look at our history of chips that we've been coming out with, in the last few years, we went from 12.8-terabit chips to 25-terabit chips to 50-terabit chips, and most people can predict when we will have a 100-terabit device coming out. So the network bandwidth is keeping up. You obviously have to architect your network with these high-bandwidth devices. As we scale bandwidth, I think interconnecting between these switches, what today is copper cables and some of this would have to be optical cables, and that is where I think more work needs to be done. But networking is there. It's ready to take this challenge on.

Dave Vellante

>> You'd like to avoid optical, I presume, wherever you can because of the cost, and you must be pushing copper pretty hard these days.

Ram Velaga

>> Yep, absolutely. We will push copper as hard as we can. For example, our 100-gig SerDes today have a reach of four-meter distances. Soon we will have products which are based on 200-gig SerDes. Just the nature of the fact that it is twice the data's rate means the reach will be lower, but still, you're going to try to pack as much as you can within the reach of copper. But at some point beyond that, when you're building these very large clusters, you will have a lot of fiber running around.

Dave Vellante

>> SerDes, your secret sauce.

Ram Velaga

>> Exactly.

Rob Strechay

>> You brought this up earlier today with when you were connecting multiple data centers together as well. And I would assume, and I pushed you on this a little bit earlier today, about how to connect the full stack. Because we're talking about VCF, we're talking about private AI running on top of that, and now you have NSX in there. And you talked about this, I thought, eloquently when I pushed you on it a little bit about, "Hey, there's that control plane." The different control planes that you have to have. Elaborate on that for people who are looking at it going, "Okay, you're talking silicon, I'm talking NSX and different things." How to bring that thread together for people.

Ram Velaga

>> Sure. When we talk about silicon, we're talking about silicon that goes into these pizza boxes or chassis that makes a physical switch. Switch/router. These are very high-bandwidth devices. Anything from 25 terabits to 50 terabits to multiple of hundreds of terabits together put into a physical box. Historically, there wasn't a real clean separation between how do you actually do things like overlay networks and how do you do a very good physical network underneath? But when you look at large clouds, large cloud providers, they've done a very good job at it. They said, "Hey, look, let me deploy a very high-performance network and that's based on the fastest speeds which are available," whether 25 gig in the past or 100 gig today, and so on and so forth, and build a very large topology. "Now, on top of that, we want to be able to have, let's say, multiple different customers and we want to keep all these customers separated," so you have some kind of a segmentation between these customers. You don't have to do that segmentation in the network. When you try to do it inside the physical network, you're trying to make sure every time you have to deal with segmentation, you have to change the underlying physical network. Instead, they're saying, "Let the network do what it does very well, connecting point A and point B at very high speeds, very reliably, with the right QoS attributes." But when I want to do things like segmentation, security, isolation, I can do that in the upper layer. That is where they would use their own internal software or this is where customers could use something like NSX, which will provide you the segmentation, and you can let the decisions happen independently. The control plane and what you do with segmentation and security could be done completely differently from, "Oh, I need to get faster speeds right now. I need to go to the next bandwidth point. I need much more burst absorption capability." So you can separate those two out and both of them can operate as ships in the dark.

Rob Strechay

>> And we were talking just yesterday as well around the private AI stack, and how organizations are looking to have more of that control. And you were talking about your internal use case where you've switched models some crazy amount of times to go and see that. How does that work as well? Because the network comes into effect because the data... Everybody wishes it was all in one place, but the data is not all in one place. And how you're bringing that together as well.

>> And let's back up a little bit, because our thesis on private AI is if you take your same terrible IaaS that you keep trying to homegrown and build yourself and think you're going to automate a complex AI application across it and be successful, you're fooling yourself. I think that's first and foremost. The approach that enterprises have tried to take to private clouds for the last 20 years, piecemealing the thing, bringing in all of these different pieces and parts and building an orchestration layer above it, thinking you're going to have the right level of reliability and the lowest cost, zero chance. So the first thing we had to do is get really opinionated to say, "If you really want to do private cloud, you need to have an integrated solution. That's without question. You're not going to be successful without it. Your costs are going to be higher, you're going to have higher challenges with reliability," and so forth. So it's get your IaaS house in order first. Then we can start to do all of this incredible orchestration that we do above it. And that's been what's really key, and that's where our products are really standing out. There is nobody that has anything to the sophistication of our distributed resource scheduler. It's not even close. And now how we're able to look at GPU profiles, understand different memory blocks that we want to pre-allocate and reserve. These are things that you're just not finding by anybody else in the industry, and this is where you're going to see far better intelligent orchestration and placement for AI workloads, and then load balancing where we need to. So this is how we're seeing architecturally a lot of this come together, and then what we're doing is just we're working now... I'd say if you're thinking about what's coming next on this, we're building capabilities that are looking at the size of the model, the context window, understanding how many inference endpoints are we expecting and what is the frequency of that, so we can start to pre-size out expected capacity for these workloads and then optimize as we need to as well with the monitoring tools that we have.

Dave Vellante

>> So obviously, your cost comparisons, Hock's keynote really played up your stack versus in the cost relative to the public cloud and the cost advantages. He told the story in the analyst meeting about what your CIO did in terms of doing RFP. The cloud guys came in 2.5 to 4x more expensive. You guys spend about only 1% of revenue on IT. Very, very efficient. So compare what you were just describing to my options in the cloud, and we can talk about the cloud sometime. It's all the same thing. AWS, you have these fragmented services, but really strong services. Microsoft's got abstractions and abstractions. Google's in between. Great data chops. How do customers think about those alternatives and what's your message to them?

Rob Strechay

>> To me, this is why repatriation is such a huge trend right now. Steven Elliott was even citing earlier this week that it's in the 80 percentile. We've seen other research as well, I believe it's from Barclays, that has seen similar types of trends. So there's a strong desire to start to move back. And move back doesn't mean you have to build a data center. We have plenty of service provider partners that would give you the capacity so you can still maintain that control at a far lower cost than what you're paying on a hyperscaler today. And all of the things that we do through virtualization to provide infrastructure optimizations, that goes to the customer's margin. They're putting that back in their pocket and that's what it's also helping to drive down the cost. So we know and our customers know they're paying the public cloud premium for the velocity. That's what they're paying for. And there is valid use cases for that. Don't get me wrong. It's not all or nothing. We don't expect it to be. But that said, everybody knows if you can have a well-run private cloud and if you can simplify it, your costs on premises are dramatically lower. And even our on-premises competitors, there are some competitors today that their prices are 2x ours. They're double. Look at their list prices. You can find it right from their price books. So we're beating our competitors on price on prem. We're superior to the public cloud as well when it comes to total cost of ownership. And these are just facts. It's basic math, so it is what it is.

Rob Strechay

>> And I think one point I'd love that you made this week, and I'll throw you a softball for this one, is-

Dave Vellante

>> I can follow up with the hardball.

Rob Strechay

>> I would expect nothing less. Cop, bad cop. And you brought up somebody like an AWS with Bedrock and how you're going to be feature parity and things like that. How have you found in the last nine months really leaning into that innovation? Because to me, having been at AWS, I understand how they work internally and how they're really focused on that innovation. How are you seeing that happen now that you're part of Broadcom in the last nine months?

>> It's been fantastic. The efficiencies that Broadcom has brought to our software division have allowed us to really work across a group of senior engineers that are execution machines that know how to get things done. We've taken out a lot of layers of bureaucracy that have allowed the company to be able to be decisive and innovate faster. There has certainly been zero slowdown. We announced private AI a year ago and yes, we expect in VCF9 we're going to have feature parity with the Vertex AI, AWS Bedrock, all the core services you're going to have, but there's a key difference, or a couple. Lower cost, full control of your data, and also our ISV ecosystem. It is very difficult if you're an AI ISV to approach a customer with a hyperscaler, because that hyperscaler is selling their own AI services, so they're directly competing against the ISV that's trying to partner with them. We're not a direct competitor. It's very easy for the ISVs to work with us because we're just going to approach customers together and get to a good outcome. So it's really those differences I think is what's helping us to stand aside from where the hyperscalers are going.

Dave Vellante

>> It's interesting. Snowflake AWS is the poster child for that type of competition and they seem to have a bromance going because AWS knows the Redshift guys might not like Snowflake, but the EC2 guys love them, so it all comes out and seems to work out. But I want to come back to the whole repatriation thing. We had a conversation with Steven Elliott and Mike Cannon, so it doesn't really show up in the numbers this way. And there's a weird thing going on. The macro is growing at let's say three point a half percent IT spend. The big three clouds combined would probably be 200 billion this year. You guys sell to them, so you know how hot that market is, and they're growing at 20%. The on-prem market maybe is growing in the mid single digits. Maybe. Some are shrinking. Cisco, Dell, IBM, HPE, Lenovo, Oracle. Soft. So I think what's happening, I'd love your take on this, is the old legacy stuff is declining. Maybe the repatriation is helping prop up the new stuff. Maybe there are some new workloads, the AI workloads. And also, that market's bigger. That install base is much larger. But that's why I say it doesn't show up in the numbers because the hyperscalers are still growing so fast, so obviously people are adding dollars to the hyperscalers. How do you square that circle?

Rob Strechay

>> I could start. I'd love Ram's perspective here too. So first, we've already covered the fact that the hyperscalers are charging a premium, so they already have a higher margin on the customer. Your modern workloads are running Kubernetes predominantly, containers, and we can really densely pack these things. We've done a ton of optimization in our own virtualization software that's giving you greater density per core than you've had before. So our software is licensed for core, so customers in turn are getting greater value from their investments. So I think that's the gray area that may not be appearing in the numbers here, but it's certainly important to the conversation. When customers have a modern app and they see the steady-state cost, on-prem is lower cost for them than a hyperscaler. Again, it's just basic math at that point. So customers that have the aptitude to do it, that's what they're going to do.

Dave Vellante

>> It's renting is more expensive than owning. We know that. And you're right, AWS's operating margin is 35%. It's not quite Broadcom operating margins, but different business. But still, it's an basically a hardware company, AWS, with 35% operating margin, which is like Oracle's and they're a software company. So I understand your point. There's an umbrella there. But given that, and given what you just said, that should be a real tailwind and we should see continued momentum in the on-prem business. And I would expect that, but I also see continued hyperscale growth. It's just astounding.

Ram Velaga

>> So maybe another way to think about this, there is the hyperscale growth and you actually double-click on this hyperscale growth. There's companies that are focused on the digital native companies. There's a growth in the digital natives. Now, you could separate out the so-called digital natives from a traditional enterprise. So there's cloud companies that serve this digital natives, and there's growth there. Then there is these cloud companies that have their existing software stack that they move into their cloud, and then they show it as their cloud growth, so there's another one. But then when you start talking about Dell, Cisco and the others, those are generally hardware companies. So in general, what you're seeing is there's more efficiency coming out of the hardware just because there's more and more transistors being fit in the silicons, and generally, there is an eroding price curve on the hardware and so forth. But when you look at it in the overall context of what you might see from the hardware companies, that revenues not keeping up, what you don't take into account is the fact that still, these enterprises are still spending a lot on the software, the application software stack, the database software stack. So you have to add all of that up including the services cost of it and then say, "Okay, what does the growth there look like compared to somebody like the clouds?" So maybe you'll have a double-click-

Dave Vellante

>> Fair enough.... needs to be done here.

Ram Velaga

>> That's why I think AWS is actually a good example, because they essentially are a hardware company with some paths. But by the way, our research suggests that their EC2, their compute, is now under 50% of their revenue, which I think probably occurred sometime during the pandemic. So there's other services, data services-

Rob Strechay

>> But I think, having been there and partnered with all of the hyperscalers, what I look at from a cost curve perspective for them is the layer stack. B Because they are building off those core services, which is the "hardware layer," and then they're building that up. And some have taken a different approach where they have single-stack products. And I think that's why when you say Google and being able to be competitive and complete to some of the stuff that they're doing in the AI space, I look at them as that they went there first. And I would say Bedrock was a little bit behind on that from that. Are you seeing that people are not only comparing against, but it's workload by workload that they're comparing it? Or is it, hey, that's what Hock was talking about of being broken, and CIOs and CEOs have to get aligned on, you've got to have a full stack so you get these efficiencies and cost reductions? Because to me, that's the difference. When you're selling private AI, you got the whole stack versus it being piecemealed.

>> Yeah, and that's the thing. And that's a key thing that Hock's trying to make, I would say, a point is if you're saying, "Well, it is what it is and I accept my internal org structure as this sunk cost that I'm just never going to be able to break that," then you know what? You're not going to get the same results. We're saying that you have to take this harder pivot and you have to rethink your approach, just like in the cloud. I can't tell AWS how to build their storage stack and I'm fine consuming a service from them and I'm fine paying a premium. So what we're saying is you can get that same level of service, that same level of automation, at a lower cost with us. To us, it's a no-brainer. Are we going to replace all your SaaS in the cloud with us? Of course not. It makes no sense. But for the applications that you're trying to build using containers and maintain control of the app stack and your intellectual property, and not buying them to some type of proprietary storage interface, we're the best option for you, period.

Dave Vellante

>> And that's your home court advantage. No question about it. We've never advised people rip out their core VMware infrastructure and move it into the cloud. That would be bad business. But you go right after their Achilles heel, which is cost. Of course, the cloud strengths, they get great developer affinity, and they have really strong data stacks. Sometimes overly complex, and they have a marketplace and an ecosystem that's very strong. And those are the things that I know you're very much aware of. But so you got to defend the base and then work on those other pieces. Tanzu attach rates obviously critical for the developer piece, and you sort out the rest as time goes on. The data piece, I think is some work to be done there. Really interested to see how you guys evolve that, especially for AI.

Dave Vellante

>> I think our partnership for object storage with MinIO is a key part of that. That's obviously a critical component for AI. And then what we're seeing is the customer demand for AI and really alignment to our value proposition is now creating new pull for modern applications back onto our platform. So that's really, I think, changed the dynamic in ways that it hasn't. You'll see us continue to grow the breadth of the data services that we support, and we're getting more ecosystem partners looking to us. And we even have the hyperscalers talking to us as well about extending their services to the data center or Edge on our platform as well. So there's just even recognition, I would say, on their side that private AI is a valid use case even for them and something that they want to address in the market as well.

Dave Vellante

>> And I think Broadcom's generally a misunderstood company. People think, oh, Broadcom's like a P.E. They just suck the life out of the companies and leave the carcass, and that's clearly not the strategy. You guys are an engineering firm in a way. VMware's got a very strong engineering team.

Ram Velaga

>> Oh, we do. You don't need to point to them.

Dave Vellante

>> Right. You identify durable businesses and you invest crazy and you dominate. That's the strategy here. And we're clearly seeing that playbook, if I can use that term, unfold with VMware. You guys, you live it every day. You and your 26 business unit partners. You're the 27th, I guess.

Ram Velaga

>> Yeah, so I actually probably could give you a little bit of historical context to the story here, which was in 2015, my then boss texted me saying, "Hey, buddy. I want you to go visit this customer in Europe and I can't make it last minute." I said, "Hey, you skipped on them last time. What excuse do I give this time?" He's like, "No, no, just go do it. I want you to go do it." So I went to see this customer and one of the other corporate execs was there with me in the meeting. He was constantly looking at his phone. I'm like, "Hey, buddy, Mike, here's a customer. Why are you looking at your phone?" And then right after the market closes, a new set comes out that says, "Avago acquiring Broadcom." Next thing you know, my phone's going off the hook. All kinds of very colorful words. "Who is this company? Why are they acquiring Broadcom?" All kinds of stuff. "Who's Avago?" Then I had very big customers reach out to me saying, "Hey, look, we really enjoyed working with Broadcom. Good luck, man. I hope you'll be able to keep your roadmap going forward. All the best." And, "You're being acquired by a private equity company. That's how they run." X, Y, Z. I was like, okay, maybe I won't have a job in a year from now, but at least I've got a year more before this acquisition closes, so I'll just keep doing my work. But you look back now, 2016 when the acquisition closed to 2024, eight plus years. In my team of over close to 1400 people, my attrition rate is about 1%. 1% attrition rate in a tech company. And if there's any talent out there we want to get, we are able to get. Any talent we want to keep, we don't lose. What happens to us is Hock is very, very good at this. He says, "Look, you're very good at doing X. Please do X, and go as fast as you can and do X. Don't sit here and think about doing Y, Z, or something else." And what we believe in is every company out there, there is a core business they start with, core technology, and eventually they reach the peak of their core technology. And when they're in a public market, they're looking for alternate businesses to grow because that's what the market's expecting out of them, and so they start to dilute themselves and they're not investing in the core business. And that's what actually happened to Broadcom in our core switching group. Broadcom is trying to get into the mobile side of the business. We were trying to compete against Qualcomm on the processor and so on, so forth, and they were really, really starving the core networking business. Hock came in and said, "Look, you guys are going to get all the investment you need. Whether it's a technology investment or a people investment, go do it." We have consistently been the first in coming out with our products to the market at least by 18 to 24 months ahead of anybody else. And these are very other large companies that came and said, "Hey, look, we'll build our own silicon. We'll be better than Broadcom." And it's proven out. One year out, two years out, three years out, we've just continuously out-executed. Why? We have a process, we have an absolute focus, and we get all the resources that we need. And we are able to keep the best talent, and we are very, very low attrition. Why will people stay if they're not doing exciting stuff? And our team stays. We're doing super exciting stuff. I think as VMware goes through this transition, there's always going to be a nine-month, 12-month period as they adjust themselves to how Hock thinks, and he expects leadership and absolute execution, and the customers will benefit from it. So I believe they will do extremely well, just like all our teams have done who have stayed on at Broadcom.

Dave Vellante

>> One of your colleagues at the financial analyst meeting earlier this year, I think it was in April, gave the analogy of the waves. I think he was a surfer. Yeah, anybody can ride the easy waves. And then it starts to get the bigger, the bigger waves. And then the really gnarly waves, that's what you guys ride, and VMware the same. You guys always had that focus on making stuff work, recovering from failure in a way that makes customers trust you. Guys, great segment. Thanks for helping us wrap up. Nvidia has announced stocks down 2% in after hours. We haven't dug into it, but we will. Thanks for watching, everybody. Guys, really appreciate your time and the insights.

>> It was a pleasure.

Dave Vellante

>> Thanks for having us here. Great show this year.

Ram Velaga

>> Thank you for hosting us. Appreciate it.

Dave Vellante

>> You bet. All right, this is a wrap. Dave Vellante, Rob Strechay for John Furrier and the entire CUBE team. Guys, nice job. Awesome two sets again. 15 years at VMware Explore, VMworld. Thanks for watching. Go to siliconangle.com for all the news. Thecube.net is where all these videos are up on-demand, and theCUBE Research is where all the deep dive research is. And check out theCUBE AI. It's our new rag. Ask it questions, get answers, get clips. We'll see you next time on theCUBE.