SC25 | Jeff Denworth, VAST Data

Clips
More from SC25

Jeff Denworth

Co-Founder

VAST Data

play_circle_outline VAST Data announced a partnership with Microsoft Azure to enhance data access for customers.

play_circle_outline Software stack integration on Azure supports multi-cloud and hybrid cloud environments for VAST customers.

play_circle_outline Navigating AI Adoption: Enterprises Face Unique Challenges Amidst Evolving Computing Infrastructures and Neo Cloud Dynamics

play_circle_outline Customers prioritize outcomes and AI efficiency over managing their own GPU infrastructures.

play_circle_outline Unlocking Potential: VAST's Federated Edge Computing Platform Enhances Strategic Partnerships and Data Access Across Computing Nodes

Info
Transcript

Jeff Denworth, VAST Data

Jeff Denworth

Co-Founder VAST Data

At SC25, VAST Data co-founder Jeff Denworth joins theCUBE’s John Furrier and Dave Vellante to unpack how AI infrastructure is evolving as exascale compute, high-performance networking and next-gen storage converge. Denworth shares fresh details on VAST’s new strategic partnership with Microsoft Azure, building on momentum with CoreWeave, Google Cloud and other neo clouds to deliver a single VAST AI operating system that spans data centers, public cloud and hybrid environments. He explains how this unified, federated data platform exposes file, object, databas... Read more

explore Keep Exploring

What partnership was announced and what is its objective? add

What is the fundamental principle of the business plan regarding customer computing preferences and software deployment? add

What insights can be drawn about the current landscape and usage of AI cloud services among large enterprises compared to large cloud vendors? add

What are the current revenue sources and future opportunities for businesses utilizing AI services? add

What developments have occurred in the edge network business, and how do they integrate with distributed computing systems? add

bolt Powered by CUBE AI

Jeff Denworth, VAST Data

search

John Furrier

>> Welcome back everyone to theCUBE's live coverage here at Supercompute in 2025. I'm John Furrier with Dave Vellante. It's where all the actions happening on AI infrastructure, and one of the hottest topics is networking and storage. And this next guest is the co-founder of one of the hottest companies in the storage area, pioneering the single name space, pioneering the concept of data platforms versus storage. Jeff Denworth is here, Co-Founder of VAST Data. Great to see you on theCUBE again. I know you're super busy and taking time out of doing deals and coming on, appreciate you taking the time.

Jeff Denworth

>> I always love seeing you guys.

Dave Vellante

>> Deals? What kind of deals are we talking about here? You?

Jeff Denworth

>> You guys got some hard news, just hit the wire.

>> Literally five minutes ago,

John Furrier

>> we were reading it as we came on. You had great news last week. CoreWeave, Google this week. Give us the hard news. Just straight up, take the microphone.

Jeff Denworth

>> Yeah, so today we announced a very strategic partnership that we've created with Microsoft and the Microsoft Azure team. The objective is to basically to bring all of the goodness of what we call the VAST AI operating system to Microsoft customers. We've been collaborating with them for, I'd say the better part, about two years now. And the objective is to basically open up data access to all of the customer's data sets for all the different compute platforms that Microsoft offers. Extend that to customer data centers using boost capabilities from Microsoft so you can build a unified, federated, and hybrid cloud. And we're super excited about it because we get to partner with a bunch of different product teams and create AI integrations in ways that we've done with customers around the world and it's very exciting.

>> Jeff, Dave and I have been talking about this operating

John Furrier

>> system and large scale systems about a year before you guys launched and we been following you guys. I gave you some props on the intro, but the reality is-

Jeff Denworth

>> That's where we got the idea for it, we were watching theCUBE.

Dave Vellante

>> Yeah, sure. Went and copy the, I'm going to build that.

>> That's all we did for the first year.

Dave Vellante

>> Well, hey. - I mean, a lot

John Furrier

>> of people watch theCUBE videos and get ideas.

Jeff Denworth

>> Off of theCUBE. - You guys talk about this OS,

John Furrier

>> but one of the hottest topics now we're in our third year

Dave Vellante

>> where this is mainstream, is the software stack. The news that you just talked about is porting the software stack to Azure so it's a whole other environment. What does it actually mean? Because as these systems get built, we've talked about multi-cloud, but now you have these large scale distributed computing systems. What's it mean for the software stack to be on Azure for VAST and for the customers?

Jeff Denworth

>> The fundamental principle of our business plan is that we want customers to be able to benefit from VAST wherever they want to compute. That starts by sporting all the different platforms that customers may want to compute on. Being able to put our software natively on their early days. We built a reference architecture that you could basically put in about seven different places and that was it. And over the last four years, the company is really a software company. We've opened up the platform support so that we can be much more general in terms of where we deploy now. That's the fundamental principle. And now, customers can get, file an object and database and vector database and event streaming services and agent deployment services next year on Azure. But the really nice part kicks in when we do two things with the types of partnerships that we're trying to create with the Microsoft team is, one, better platform engineering so that we co-design things that express the capabilities of our software as best as they can be. Second is platform integration with the other first- party services that Microsoft has, so customers get a much more holistic data experience as they work across the different Azure products and services.

Dave Vellante

>> Last year at GTC, Jensen laid out the three vectors of opportunity for AI. You said AI in the cloud, AI and enterprise, and AI, physical AI, robots and edge.

Jeff Denworth

>> Yep. - Two years ago, maybe it was even more, Jeff.

Dave Vellante

>> Three years ago you guys were really knocking down the neo clouds. You obviously did very, very well there given you're purpose built for AI. As we go into enterprise, what are you seeing? Because it seems like it's a different world when you talk to the neo clouds and as you well know, they're working those GPUs, they're lit up, they're really efficient in terms of their utilization enterprises, they're still learning. And now, this deal you have with Microsoft all about hybrid you just talked about, so how do you see that enterprise AI adoption going? What's that going to look like? Where are we today? How long is it going to take them to get that muscle memory?

Jeff Denworth

>> The first thing that I would say is that when you see these very large projects that the large cloud vendors or the neo clouds are doing, typically that's an AI lab that has the conviction and the capital to go off and just start running from day one.

Dave Vellante

>> And the skills to the people.

Jeff Denworth

>> But if you do it by count of number of customers, it's very obvious to us at least that enterprises are overwhelmingly the largest user of AI cloud services. And so if I think about Microsoft in particular, I can't go into a customer meeting today where there isn't some sort of endeavor going on with some Microsoft AI product that they're working on. Same is largely true for Google. We announced a partnership with them last week. They're doing great jobs, basically building proving grounds and providing higher level capabilities and bringing the models and everything that you need from a toolkit perspective to basically just start running or at least start experimenting so that you can run. And then I think a lot of people conceptualize this idea that the enterprises are going slow because they themselves aren't building 100, 000 GPU clusters in their own data centers, but that's not a core competence to a lot of these organizations. They're not IT organizations by design, unlike the neo clouds, that's their purpose in life. And so as a result, a lot of them are just consuming different cloud services, and that's one of the reasons that we wanted to partner with Microsoft because there's all these customers that we're working with, they're like, "Here's where we're computing, either you're there or you're not."

Dave Vellante

>> But they're trying to solve a different problem in the enterprise. We were talking about it earlier, they're trying to get outcomes, not necessarily monetize their GPU clusters. And that's hard because you've got to harmonize your data. You're saying, "No, it's not that hard." Or what's your-

Jeff Denworth

>> I think a customer of CoreWeave or Lambda is also trying to get outcomes. They're trying to build a new product and then deploy that as an inference service. So at some point, everybody converges to that same end state where I'm taking a fine-tuned AI model and then I'm using it for some sort of internal or external service.

Dave Vellante

>> I guess what I would say and tell me because you're deep into this, today, the revenue monetization is chat GPT users subscriptions.

>> Yes. - Okay, great. The future revenue opportunity

Dave Vellante

>> that you're talking about, which is the biggest TAM I think

Jeff Denworth

>> who knows about physical AI, that could be the biggest of all TAMs, but it's going to take a while. But those customers, they have to do certain things to get their data ready. They got to harmonize that data. They got a lot higher governance bar, et cetera. Are you feeling like, well, which parts of the enterprise are there today? I mean big banks maybe of course the mainstream enterprises by my research, but you have a different observation space.

Jeff Denworth

>> You reference ChatGPT who's buying that with fifty- hundred-thousand-seat-license contracts?

Dave Vellante

>> Yeah, consumers. And you're right bunch, they've delivered through APIs. I mean, do you see that as the mainstream business model? We see it extremely large. Take some of the more popular AI models, they Gemini-plod, sorry to interrupt, but I think they've said a million million enterprise customers, right?

Jeff Denworth

>> Yes, so there's a very vibrant enterprise business underneath some of these commercial models like Cohere and Anthropic and OpenAI that they're selling to enterprises. And it comes back to what I said earlier, not every enterprise wants to become an IT organization. Now, there are certain reasons that people endeavor to build their own models because they can't get something that they've found in the commercial space or the open source space. And then there's reasons why they build the infrastructure for either cost or security reasons that they can't go and use some shared cloud infrastructure, but the number one mandate at the board level within an enterprise is try to figure out all the possible ways to get acceleration and efficiency from AI. And if that's true, SaaS services are the first place you go to start that mission because you can move quickest.

Dave Vellante

>> And the data's pretty good, right? I mean, so generally speaking-

Jeff Denworth

>> The data is the other consideration.

>> How do I bring my data to these models?

Dave Vellante

>> And I think to your earlier point, I mean a lot of the headlines are all the CapEx that we're spending. Having said that, we're probably going to spend about cumulatively $4 trillion in CapEx between now and 2030, and maybe we'll generate half of that in revenue, but that's still $2 trillion, which is about the size of the enterprise IT business in total, it's crazy. And then at some point it's going to cross over, but it'll probably take the better part of a decade to catch up to that initial spend.

Jeff Denworth

>> It's the classic R&D and monetization model just at levels of scale they've never seen before.

Dave Vellante

>> And then presumably when it crosses over, we're going to see levels that we've never seen before. That's the bet that we're all making, right?

>> You start measuring productivity gains in terms

Jeff Denworth

>> of GDP movements. That's the only place that type of money comes from.

John Furrier

>> Jeff, talk about the neo clouds because Dave brought that up. You guys have some CoreWeave expansion going on Google Cloud. You mentioned that.

Jeff Denworth

>> One of the first billion-dollar agreements to ever be announced in our space, as best as we know, we don't see business like that typically done.

Dave Vellante

>> I think maybe one other, Palantir I think had one with the government.

>> A billion-dollar contract?

Jeff Denworth

>> With the government, I think.

Jeff Denworth

>> Just software? I'm not sure.

Dave Vellante

>> I think they did. I think it was for the, well, no,

Dave Vellante

>> no, there were services in there.

>> I'll take your bet. - No, you're right. There were

Dave Vellante

>> services in there.

Jeff Denworth

>> It wasn't pure software.

>> Billion-dollar contracts because.

John Furrier

>> We owe you a beer.

John Furrier

>> See how that plays out, we'll see. We'll be covering that.

Dave Vellante

>> We see they pay all that money.

Dave Vellante

>> You're right. There was services involved there. We'll see.

Jeff Denworth

>> And at times infrastructure. A lot of times that's hosted.

John Furrier

>> Yeah, we'll see how that's recognized. CoreWeave expansion, Google Cloud traction, the neo clouds. I interviewed a company called Argentum there, essentially an asset-light neo cloud aggregating marketplaces where people can just buy compute. This is a real market. I want to get your thoughts on why one, why you have traction in these clouds because, and then two, how do you see that being a service layer for enterprise? I talked to the Lambda folks, they're like, "We're one API click away from getting compute. " I mean, this is basically the model and the business side. Can you talk about the expansion CoreWeave, Google Cloud and this new crop of neo clouds or specialty clouds for AI service AI compute?

Jeff Denworth

>> I think it's interesting. The first thing to consider is that when we started in this space, we didn't start with teams that had very large engineering organizations that needed to go out and make partnerships with third- party software providers. And the reason that I would argue we're the gold standard in this space now is because we built something that can scale massively that just works. I can't overstate how important that is for organizations that are paying SLAs back to customers when their clusters fall down. You think about a 10,000, 20,000, 100, 000 GPU cluster when these things go bump in the night, the cash machine starts ringing, but in the opposite direction of what these cloud vendors hope for because they have to pay SLA penalties. And so it's actually funny where it's Supercomputing, I met a neo cloud that we had been chasing for years just in the line for picking up my badge. And he looked at me and said, "We're tired of what we're doing right now. We were wrong in dismissing you guys. We just need something that works," and something that just works that can withstand the compute power of 100 or 300,000 GPUs is not something that comes easily in this space.

John Furrier

>> And that's nuanced, I want to just unpack this. I think it's a huge point. The scale penalty is SLA wrong direction. The upside potential is better performance.

Jeff Denworth

>> You guys are going to interview somebody today, I can't tell you who, who their data platform vendor has told them that they might be able to get access to all of their data sometime in late December. And so these types of dynamics don't even play in a space where an AI lab just got a billion dollars and they have to go turn that into an intelligent model that generates revenue. This is why people first and foremost use us. The second reason that we're a standard in this cloud is it's more than just storage. It's got integrated database services, event streaming services, vector database services-

>> Operating system.

Jeff Denworth

>> It's in a runtime and that's an operating system for AI.

John Furrier

>> And if I think about a data monetization opportunity, nobody other than VAST gives you the best chance to go and take a set of data products and turn them into services that can really provide a comprehensive experience for either an AI lab or an enterprise customer.

Dave Vellante

>> Well, it's so important because you're appreciating your CapEx pretty rapidly

Jeff Denworth

>> For training, that is true.

Dave Vellante

>> Okay. Finish that thought I want to ask you is today's training infrastructure, tomorrow's inference infrastructure?

Jeff Denworth

>> I would say yes, but with the exception of a certain number of models that need to be fit into the bigger and bigger GPUs. But the second thing is you said inference. There's a data processing element that fits around this. We work with one of the largest AI coders and they have a reinforcement learning system that they've built that requires mass scale event processing. They had to go build their own thing. They couldn't use Kafka and they didn't even know that we existed. And those are the types of stories that we're now starting to pull through these neo clouds saying let's build a holistic AI service delivery platform that can scale to whatever proportions that you want. Super efficient, and that's the big differentiator with VAST.

Dave Vellante

>> There's one other thing for VAST that I want to bring up and test with you. Did you hear Satya when he was talking to Gerstner? He said, "Yeah, in the early days of software, when we shipped it on CD-ROM, " do you remember he said this or maybe you didn't see it?

Jeff Denworth

>> I didn't see it. - Well, you ship a software on CD-ROM,

Dave Vellante

>> your marginal economics go to virtually zero.

Jeff Denworth

>> Yes. - But he said with cloud, we had cloud cogs.

Dave Vellante

>> Okay, we all have seen that, the SaaS, the trillion dollar paradox, we heard that and he said, "Now with software in this AI era, it's different. We have big expenses, big CapEx, big OpEx as well. " So the gross margin is actually from gross margins. Everybody not nearly as attractive as it used to be. It's not the case for VAST, right? You've got pure software, pure software, you don't have to incur-

Jeff Denworth

>> Well over 90% gross markets.

>> So if you've got 90 over 90% gross margins on a billion

Dave Vellante

>> dollar software deal, you do the math people. And so that's a different business model than what Satya was talking about, and it's actually unique juxtaposed to some of the software firms that are associated with AI.

Jeff Denworth

>> It's true. We've built an extremely unique business model, cash flow positive, growing multi-hundred percent per year, and gross margins that would be enviable of any company in the world.

>> It's super capital efficient relative to-

John Furrier

>> So your customers, the neo clouds coming back to them.

Dave Vellante

>> You mentioned about the economics of bad decisions, event streaming. That's a great example of something at scale that if you go the wrong direction. Only takes one piece of the puzzle to crash and everything will go sideways. That brings up the whole bubble conversation. There's a lot of.

Jeff Denworth

>> That's a pretty liberal jump from one to the other.

>> Well, I mean the neo clouds are going

John Furrier

>> to turn all this CapEx and not get the software, right? I mean you could just say mean the economics, but the over-investment and the combination of operational efficiency, they do go hand-in-hand your take on the bubble concerns because obviously we feel it's a over-inflated market, but certainly not a bubble like the dot-com, which by the way, that all happened anyway. What's your take?

Jeff Denworth

>> Any indication of lower demand? I mean, well, you think at the expense of offending a few of our partners, they'll say there are a few businesses that were never built to succeed in this space because they were not built for long-term, taking short- term contracts, focused on a specific GPU generation. And as NVIDIA and other processor companies bring out new capability, they're stuck trying to monetize that thing that they never locked in a customer for long term. That's a really, really dangerous business model. On the other hand, I've been talking to what you, I don't know if you know the semi-analysis guys, but they create a ranking of the different clouds of the bigger clouds, or I would say the better ranked clouds in the space. I've been talking to all of them this week and every single one of them says, we don't have enough to sell right now. So power is the limiting factor in terms of how they can continue to grow their business. Everybody's out scouring for power all the way around the world. Now you're going to space. I don't want to talk about that, but there's serious demand right now for the clouds that are serious. Somebody said something to me the other day in Australia I thought was interesting. It's a lot easier to sell 2000 GPUs than it is to sell 20 GPUs in the marketplace today. And I see that all over the world. Why is that? Because of the factory configuration versus no, because the sales process of working with an enterprise customer is just starting to dabble with, or maybe a research team just starting to dabble-

Dave Vellante

>> Back of the line.

Jeff Denworth

>> Right, versus XYZ organization just has a new idea for a new, let's say, world model or physical model or something like that. They get a billion dollars, they want to go right now.

John Furrier

>> It's really prioritizing where the sell-through is.

Jeff Denworth

>> It's still very in research and our observation is the world right now is about from an infrastructure perspective, half inference, half research, and then you have a lot of organizations that are trying to figure out how to balance the same infrastructure for both workloads.

Dave Vellante

>> Sorry, that's today's balance and then in the future 80/20?

Jeff Denworth

>> It has to get to 80/20 at some point just to pay for the investments.

Dave Vellante

>> I have to ask one quick question on the last one. China's got they're not power constrained, they're GPU constrained.

Jeff Denworth

>> Yes. Right. So not just with NVIDIA, it turns out that their own internal initiatives are also constrained in terms of what they can build and how quickly they can build.

>> Any thoughts on how that plays out?

Jeff Denworth

>> I don't play geopolitics.

Dave Vellante

>> He's going to take a pass. All right, so final question.

John Furrier

>> . - Running out of time, we got

John Furrier

>> to get the aging question and obviously performance

Jeff Denworth

>> and driving the agentic infrastructure, that is where people are talking about the value will be extracted. How are you guys seeing that? What's your role in the agentic piece? Because as more enablement comes in many you're saying that's going to be the action, including us. What's your take on the agent workloads?

Jeff Denworth

>> Earlier this year, we announced that you'll be able to deploy agents from our system early next year. We want to build an agent builder that allows you to put agents directly on top of your data. I think the whole market has changed over the last 12 months or so. Now that reasoning models are here because with the generative AI area, you basically pump a lot of data into some sort of training run. You get a bunch of weights out and then you forget about data during the compute time and you definitely forget about it once you've actually deployed. Your service agents need data products in order to make decisions. So now, data systems become an integral part of my compute framework, and if all of that data is pre- existing within an organization is just waiting to be unlocked by agents, then our belief is that the world is about to embark on one of the largest technology refresh events in history now that people realize that they need to up- level their data infrastructure to feed these new agentic systems.

Dave Vellante

>> So you guys are growing like crazy. You got a bunch of money in the bank, you're profitable. Are you IPO ready?

Jeff Denworth

>> In some respects, yes. We've got all of the controls in place. We hired Amy Shapiro who used to be CFO of Shopify. We announced that last year. She's running a very tight ship.

John Furrier

>> Getting ready. - The business composition looks very much

Jeff Denworth

>> like the first few quarters of some of the most successful software IPOs in history. So in that regard, we could go public today. We have everything in place to go and be a public company In terms of all the teams in place and things like that. Are we reporting on a quarterly basis? Internally the board, we are. To the company, we are. I think we're probably still a few quarters away from being really ready.

>> We got to wrap. But final question, I have

John Furrier

>> to throw it out there, the DGX box is getting smaller and smaller. The AI factory is beholden in the hand. Jensen's like, "Here's the factory in my hand. " The edge, your vision on the edge real quick.

>> It's a very, very vibrant part of our business.

Jeff Denworth

>> Well, it's becoming more vibrant, I would say, John. We've made an announcement with Akamai earlier this year. Similar in vein to some of the stuff that we're doing with Microsoft, Google, and CoreWeave, where they want to build out a full- edge network with VAST. We have a few other telecommunications companies we're also going to be doing this with, so I see it now starting to percolate in a big way.

John Furrier

>> And an operations system has to have distributed computing, nodes, edge, core, cloud?

Jeff Denworth

>> Well, that's why we built federation into the platform about two years ago. So now you can take all the places that you want to compute, your edge, your core, your cloud, your neo cloud, and they can be all one canonical VAST system where anybody that gets onto any part of the network can see all of their data, they can process all their data. We built this, think about a compute router that basically goes around the world, finds the most available resources with the most available data and runs the compute right where the data is.

John Furrier

>> Well, we'll put a pin in that. We'll have to come back and do a deep dive on the edge. Jeff, great to see you. Thanks for coming on. Congratulations on the momentum and future IPO. We'll see you before then for sure.

Jeff Denworth

>> You will. - I'm John Furrier with Dave Vellante.

John Furrier

>> We are here Supercomputing 2025, our fourth year in a row, unpacking now the AI Supercomputing show. We'll be right back with more after this short break.