SC25 | Andy Pernsteiner, VAST Data

Clips
More from SC25

Andy Pernsteiner

Field CTO

VAST Data

play_circle_outline Overview of theCUBE's live coverage and guest introductions at Supercomputing 2025.

play_circle_outline Evolution of VAST's software stack from HPC focus to enterprise-grade capabilities.

play_circle_outline Significance of resiliency and non-disruptive upgrades in NeoCloud environments.

play_circle_outline VAST's Sync Engine: Revolutionizing Data Migration, Indexing, and Integration of External Sources through Advanced Data Pipelines

Info
Transcript

Andy Pernsteiner, VAST Data

Andy Pernsteiner

Field CTO VAST Data

In this interview from SC25, Andy Pernsteiner, field chief technology officer at VAST Data, joins theCUBE’s John Furrier and theCUBE Research’s Jackie McGuire to discuss the company’s evolution from its high-performance computing (HPC) roots to powering the next generation of NeoClouds and enterprise AI. Pernsteiner details how VAST Data is securing billion-dollar software deals by providing the resilience necessary to support expensive GPU-as-a-Service environments where downtime costs millions of dollars per minute. The discussion highlights how the platfor... Read more

explore Keep Exploring

What has been the experience and atmosphere at the Supercomputing 2025 event for VAST Data and its representatives? add

What advancements or changes have occurred in the company's approach to service offerings in relation to high-performance computing (HPC) and enterprise needs? add

What are the advantages of the platform compared to alternatives that users might consider? add

What is the purpose and functionality of the Sync Engine software product mentioned in the text? add

bolt Powered by CUBE AI

Andy Pernsteiner, VAST Data

search

John Furrier

>> Welcome back everyone to theCUBE's live coverage here at Supercomputing 2025. I'm John Furrier, host of theCUBE with Jackie McGuire with theCUBE research. Andy Pernsteiner is here, field CTO of VAST Data. Welcome back to theCUBE. You were on our original Supercloud 4, I think, or 3, and then at the NYSE. Great to see you again.

Andy Pernsteiner

>> Great to see you as well. Thanks for having me.

John Furrier

>> The show for VAST, your booth is huge. A lot of activity around the company. A billion dollar software deal, very high number. You don't see those very often. I think. I don't really can't remember.

Andy Pernsteiner

>> I don't know of any.

John Furrier

>> Jeff Densmore was saying about that. How's the show going for you? You're out in front of customers, so you're out doing the deals because these aren't like, "Hey, you want to buy a watch." These are serious system deals. Take us through what the show's been like for you at first?

Andy Pernsteiner

>> Also, from a show standpoint, a lot of what I've been focusing on is meeting with customers and kind of reminding them about all of the things that we do. The thing is we meet with a lot of existing customers here at Supercompute as well as prospective ones, and the existing ones, it's always good to hear what their experiences with the product, and also what their ideas are for how to use the newest set of features that we've been releasing. And for me, this is what I get out of it. I go, I try and test out ideas. I see what ideas they have and then we can use them at scale. And so you mentioned the billion-dollar order. Those don't happen overnight-

John Furrier

>> Not for software....

Andy Pernsteiner

>> and the other part of it is that it takes a lot of people on our side to come up with everything that we need to. There's so many logistics and pieces associated with it, but it's really validation that all the work that we've been doing is providing real value to people.

John Furrier

>> We've been following you guys since the beginning, the launch in our studio and watching the success. Renny's been on multiple times. So we've been kind of documenting the rise of the VASTronauts as you're called. But what's interesting is the deals that are announced here with Azure, it's the software stack that's kind of evolving. Take us through where that software stack is now, where it came from and where it is pushing the envelope on. Where are the wins coming from and why?

Andy Pernsteiner

>> Yeah, yeah. I mean if you think about back in time, the HPC community was our focus, primarily the research labs and a lot of the sort of publicly funded research institutions. And our focus initially was providing a cost-effective all-flash storage platform for doing large-scale analytics and data processing. And that's interesting enough, but we wanted to solve bigger problems. We extended the platform to support enterprise-grade encryption, data protection services, all of the things that people would expect from what you would consider to be a more traditional enterprise file system. We brought to large-scale all-flash implementations. And so our ability to sort of morph into a company that could not only service the needs of HPC but also the needs of the enterprise, really sort of kept bringing us forward in terms of the new ideas that we had, the features and functionalities we added. And over the last several years, as you've probably noticed, there's been the dawn of the NeoClouds or the GPU as a service clouds. And we found very similar workloads in those worlds as we did in the HPC worlds. I would say, "The level at which they operate," because GPUs are so much more expensive, they require so much more power, it's much more important to get things right in those environments. I'm not saying it's not important to do it right in the classic HPC world, but-

Jackie McGuire

>> There's a lot of sunk costs, yeah.

Andy Pernsteiner

>> I mean, when you're spending billions of dollars just on GPUs, you need to make sure that when they're running, they don't have a hiccup. And so a lot of the largest customer successes that we've had, including the most recent deal with CoreWeave, is based on the fact that our platform can scale, but then also as highly resilient so that they can get the most out of all of their investments, not just out of storage.

John Furrier

>> Well, they can quantify. They can quantify a hiccup. Define hiccup.

Andy Pernsteiner

>> Millions of dollars per minute.

John Furrier

>> Define hiccup? What's a flub, what's a hiccup?

Andy Pernsteiner

>> Yeah.

John Furrier

>> Take me through what a hiccup is?

Andy Pernsteiner

>> Well, a hiccup obviously could mean-

John Furrier

>> Again, where things break down? You got do it yourself. You guys got the platform. When I say, "Hiccup," things that could go wrong that-

Jackie McGuire

>> So I know what can go wrong. I am a data scientist. I come from a data science background, but I went to a boot camp, so I learned how to write Python over six months. I never really learned how to do things like data at scale. I learned you open a Jupyter Notebook, you import everything you could possibly think

Andy Pernsteiner

>> You could just bring your laptop.

Jackie McGuire

>> You hard code all your API keys. So when I was reading the blog you wrote on the VAST Data engine that one of the things that can go wrong is your data scientists don't necessarily understand infrastructure. And one of the things that you highlighted in your blog was kind of the ability to use simple Python to do things you need to do. And I was like, "Yes."

Andy Pernsteiner

>> Basic things.

Jackie McGuire

>> Because I did so many-

Andy Pernsteiner

>> It doesn't have to be crazy....

Jackie McGuire

>> bad things that I didn't mean to do because I didn't have a background that would've enabled me to understand the risks I was taking like running up-

Andy Pernsteiner

>> And to be honest-...

Jackie McGuire

>> five or six figure cloud computing bills because I didn't limit the number-

John Furrier

>> You left that S3 bucket open.

Jackie McGuire

>> I didn't limit the number of things I was pulling from Kibana, put a query in, went to lunch, came back-

John Furrier

>> So what would be a NeoCloud example of, because they're on a tight rope, they're spending dough, they've got to earn cash. What are some of the things that you guys solve for them versus alternatives that they would use?

Andy Pernsteiner

>> I mean to start with, the platform is resilient. We started in HPC because we knew they'd beat the crap out of us, and we made-

John Furrier

>> We did....

Andy Pernsteiner

>> sure to be resilient in the face of that. We also came in from the world of having enterprise backgrounds where downtime wasn't an option. Upgrades had to be non-disruptive. So let's just start with the basics right there. But in terms of the NeoClouds, we're increasingly seeing them want to offer more services to their end users. I mean renting GPUs by the hour is one thing, but providing service layers on top of it. I mean if you think about let's say, "CoreWeave," do they want to have to build their own block store, their own database table format? Do they want to have to create their own mechanisms for doing all these things, or would they rather take something that exists and is on the same platform, what they're using for everything else, and expose it. And so we can increase the ability for them to, or reduce the time to market, you could say, for them to offer these higher level services to their users. It just means that they can get the most out of their investment.

John Furrier

>> And I think what I like about I said earlier, following you guys, the software stack. Took your lumps in HPC, that's kind of like spring training. You get through that, and you can win anything, right?

Andy Pernsteiner

>> I think so.

John Furrier

>> And so you go to the NeoClouds, they need to have a software stack, so they plug you in. Okay, next question is enterprise. I won't say, "The laggers," but they are kind of the laggers because the pressure on the NeoClouds, because of the demand curve. We were expecting to see the enterprises rolling, so, okay, what's the solution there? It's not as clean to swap out fabric.

Andy Pernsteiner

>> One of the challenges in the enterprise besides they have a lot more process to follow. They have policies. There's regulation, especially in the financial services world that prevents them from just moving too quickly. We find that oftentimes, and this is something common I see in the large banks, that they can start a pilot to create an AI factory for example, but they can't bring it to production at scale because they don't have a platform that gives them all of the resiliency and the security and the traceability and data governance that's required. And so it stays in a sandbox forever. And so our goal is to help people understand how to leverage things like the data engine, including all of the policy controls that we have, to make it so that you can take something from your laptop, deploy it at scale, but in a way that aligns to corporate policy or aligns to data governance policies that's-

Jackie McGuire

>> Provides resiliency....

Andy Pernsteiner

>> the hard thing.

Jackie McGuire

>> Provides-

Andy Pernsteiner

>> All of that....

Jackie McGuire

>> multi-region.

Andy Pernsteiner

>> Definitely.

John Furrier

>> What's that blocker? How do you break through that blocker? It's like we need a TSA CLEAR pre-version of these projects because there's this huge sprawl of sandbox backlog that's kind of stuck there. Some call it POC purgatory, but is it skill set, bandwidth, lack of standards? Why is-

Jackie McGuire

>> Headcount, right?...

John Furrier

>> the blockage, maturity?

Andy Pernsteiner

>> Where I see things getting broken down is as you mentioned earlier as a developer, you learned how to write code, and how to analyze data and process data. You didn't know about infrastructure or scale. A lot of the infrastructure people don't understand the problems you're trying to solve. So our goal is to bridge the gap between those two groups of people and be a layer where we can offer services for developers to do things at scale, but we do so in a way that allows the infrastructure people to manage it very easily, because that ends up being a breakdown where people might sandbox something, but then they have to go and deploy it somewhere. And there is no infrastructure team that knows how to deploy these things at scale.

Jackie McGuire

>> Yeah. And you can't just hire three people because it's going to take months for them to get up to speed on what you've done, what you need to do. You can't just hire somebody, and on Monday they're in your cloud doing deployment at scale.

Andy Pernsteiner

>> It's very challenging. The other part of it is that I look at a lot of the sandbox projects and they're going and iterating on which vector database do I use, which models do I choose? And we're not necessarily going to be opinionated on everything, but we've built most of the functionality into the platform so they don't have to go and POC every random vector store. They don't have to POC every single Kubernetes control plane. If we offer these, and they don't have to do anything besides just enable a feature on the platform, it speeds up their time to development and deployment and it just makes it easier for them to manage. Because nobody really wants to go manage a separate Kafka cluster, to manage Knative services on their own. They don't want to manage all these pieces. It'd be much easier if it was integrated into a platform, and that's where we're really focusing energy.

Jackie McGuire

>> We're going to go through the RFP process and manage five different SOWs. And that's sometimes-

Andy Pernsteiner

>> And that takes time....

Jackie McGuire

>> even just the-

John Furrier

>> All right, so the silos, or you guys are breaking that down, the software stack moves around. What I'm intrigued is that you're now moving to different environments, super cloud, multi cloud, whatever you want to call it, but that's not just cloud. That's like you're going to have behind the meter on some edge cases. You're going to have now, the software stacks will be distributed. So what part of the value proposition in AI is horizontal scale and having data access across environments with models doing their job wherever they need to go, that requires a connected intelligence. How do you look at that? Are you seeing any use cases around that kind of deployment?

Andy Pernsteiner

>> We do. I mean we can even focus, even taking AI out of the frame of reference for a moment. Even in general purpose, existing workloads, having the ability to compute where the data is and having the data able to be accessed wherever the compute needs it to be, is a challenge that everybody has to deal with even outside of AI. AI obviously amplifies that. And so our goal is to again, make a simple layer that makes it easy for data to be where it needs to be. And I'm not going to pretend like we defeat speed of light, right? We can hide it a little bit, but we can also create mechanisms for people to easily move the data where it needs to go without having to make extra copies, without having to manage large dependencies. We can make things much simpler. And it's like anything else. If you don't make it simple, people aren't going to do it.

Jackie McGuire

>> Yeah, with the data engine, I was at Cribble for a while, and so what you find is that generally I'm the security guru here and the whole processing the data where it is, where it needs to be, people pay based on ingest for a lot of platforms. So the more you can do where the data is to shape it into what you need to do, to ingest it in the right format, to process it in the right way and then send only what needs to go to a platform only that data there, that actually must be a huge efficiency gain, but also financial gain.

Andy Pernsteiner

>> Ultimately everything ends up coming down to cost. People talk about the word performance. Performance is another word for cost, right? And you mentioned the ingestion. There are a lot of platforms out there that charge based on ingestion rates, that charge based on CPU hours used. And if we can reduce those down, either by advances in our database or advances in using our data engine to do pre-processing in situ, it does reduce the amount they have to push somewhere else. It reduces their egress fees into and out of the cloud. There's all kinds of side benefits that you get from it. Yeah.

John Furrier

>> Andy, talk about the realities on multivendor. Not every company is going to want to roll out VAST 100%. You guys are coming in with a great platform for all the benefits you talked about. How do you guys look at other data sets? Talk about how your strategy was to kind of, I won't say "Encapsulate," bad word, but how will you-

Andy Pernsteiner

>> I mean at first we just told everybody to move everything to us, just do it. They didn't just do it.

John Furrier

>> Well, you got a billion dollar order.

Andy Pernsteiner

>> Come on.

John Furrier

>> It's working at many levels, but you do have a reality where you have some legacy, enterprise specifically sometimes brittle, but no one wants to play with it. Do you guys look at heart, extracting that into the platform? What's the rollout look like?

Andy Pernsteiner

>> There's multiple facets here, but the one that we've been focusing on lately is we have clients that are going to have data in multiple data centers on multiple storage mediums, and sometimes it's not going to be on VAST. And so we developed another software product that we are actually giving away for free, at least right now, to all of our customers that allows them to not only migrate data from place to place, but also to index everything where it is. And so they can build a holistic picture of all of their data, identify, okay, some of this data looks like it could move over to VAST for the most efficient way of processing it. And some of it maybe it needs to sit in Deep Glacier or wherever it happens to be. We're not picky. So it's called Sync Engine, and it has the ability to index and migrate data both from POSIX related file systems as well as object stores. We've also built connectors for Confluence, Google Drive, and we're working on OneDrive now, because we also find that as people build data engine pipelines to do document classification or RAG or any sort of derivatives, they need a way to bring the data in. You can't just magically do it. It has to go through a pipeline, and to bring that data in at scale reliably is challenging. And so there's another blog post I'm about to launch which focuses on that aspect of it. Moving large quantities of data from different disparate sources and bringing them through the platform and through a data engine pipeline.

John Furrier

>> So it's a nice way to give some inventory views of where you've got your data. Let the customer decide. Let that sleeping dog lie right there, we'll come to it later. Or I don't really need to have real time or whatever AI onto it.

Andy Pernsteiner

>> Definitely. And we have customers who are moving massive amounts of data using this, 50 petabytes at a time. And there's customers using it to move small amounts of data. We're not so picky. We try to design everything with scale in mind, and so it can start small and get very large, and that's just a principle that we've always had.

John Furrier

>> All right. What's the best meeting you've had here at the show that you can share? You don't to name the name, but categorically lay out that-

Jackie McGuire

>> Pick favorites.

Andy Pernsteiner

>> So there was one, and you're going to figure out who it is in a minute, but they're working on building out a sort of publicly funded AI factory. And the idea is that they can offer GPU, compute and storage to service for very scientific related purposes. At the end of the meeting, we all clapped because, well, it means that we all get to go to Italy to help build this with them.

John Furrier

>> I think we interviewed that company on theCUBE here, but that's the scope of the scale. I guess my final question would be how has the scope of the scale changed just from three years ago?

Andy Pernsteiner

>> Oh, my gosh.

John Furrier

>> I mean you guys just launched.

Andy Pernsteiner

>> It's so hard to even remember three years ago.

Jackie McGuire

>> Yeah, I know. I'm like, "What year is it?"

Andy Pernsteiner

>> From a storage standpoint, you can fit five petabytes in two rack units without any kind of compression or anything. We see customers who can fit 10 petabytes in two rack units. When I worked, I had previous jobs at other companies, 10 petabytes wasn't even a thought for a gigantic system. And so three years ago, I feel like the scale of our systems, we were definitely on the trajectory where we were headed in this direction, but now for a system that's 100 or 200 or 400 petabytes, it's kind of just noise. It used to be that that was all that-

John Furrier

>> Huge deal....

Andy Pernsteiner

>> was focused on. Now it's just they happen almost every week.

John Furrier

>> There's more data coming, everyone's, more data's coming in.

Andy Pernsteiner

>> Everybody wants more of it, and they want to process all of it. So it's very interesting.

John Furrier

>> Andy, thank you so much for spending the time. I know you're super busy at the show. Thanks for coming on.

Andy Pernsteiner

>> It's very much a pleasure. Thank you.

John Furrier

>> Thanks for coming on. I'm John Furrier with Jackie McGuire breaking it down. The data's coming. The systems are getting bigger and in all ways more dense and this power in this AI generation. Thanks for watching.