theCUBE + NYSE Wired: Physical AI & Robotics Leaders | Renen Hallak, VAST Data

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

Renen Hallak

Founder & CEO

VAST Data

play_circle_outline Introduction of Renen Hallak, CEO of Vast Data, discussing AI and data infrastructure.

play_circle_outline Shift in focus from training to inference and now to agentic workloads in AI.

play_circle_outline Transforming Data Architecture: Adapting to Unstructured Data and AI Workloads in Vast Data's Evolving Landscape

play_circle_outline Navigating Governance and Compliance: Simplifying Data Lineage and Management for AI Applications in Enterprises

play_circle_outline Discussions with financial institutions transitioning from sandbox to production AI environments.

play_circle_outline Scale and performance needs of AI workloads exceeding capabilities of older infrastructures.

Info
Transcript

Renen Hallak, VAST Data

Renen Hallak

Founder & CEO VAST Data

Renen Hallak, founder and chief executive officer of VAST Data, joins theCUBE’s John Furrier and Dave Vellante at theCUBE + NYSE Wired: Robotics & AI Infrastructure Leaders 2025 event. The discussion focuses on architectural innovation in AI infrastructure and the enterprise demand for scalable, high-performance data platforms.

Hallak shares how VAST Data is evolving to meet the full AI lifecycle, from training to inference, while enabling real-time performance at scale. The conversation also explores partnerships with companies like NVIDIA and the g... Read more

explore Keep Exploring

What is the current status and recent developments in the storage industry, particularly concerning Vast Data and its data operating system, over the past year? add

What changes are occurring in the field of AI and how are organizations adapting to these shifts? add

What changes in data analysis and infrastructure are necessary for handling unstructured data as compared to traditional big data techniques? add

What are the challenges and considerations for governance in the context of data security and compliance as AI systems evolve? add

What changes are occurring in the approach of big banks regarding AI applications and infrastructure? add

What are the key constraints and challenges related to scale and performance in modern data workloads compared to older systems? add

bolt Powered by CUBE AI

Renen Hallak, VAST Data

search

>> Hello, and welcome to theCUBE here in our Palo Alto Studios. I'm John Furrier, Dave Vellante. We are here for theCUBE's series on robotics and AI leaders. This is part of our NYSC Wired community presentation. A great event last night. We've got a great guest here. Renen Hallak is the co-founder and CEO of Vast Data. Renen, great to see you again. Thanks for coming on this special AI leader series. Appreciate it.

>> It's great to be with you. - So infrastructure leads, it's our second year doing this and you've been on multiple times, but the world's changed. So give us a quick update. You got CHIP software, geography of the big issues in the industry. Storage is now everywhere, but the data operating system, which you guys are doing is key to that. Talk about what's changed for you in the past, let's say 12 months.

Renen Hallak

>> Well, everything. The world shifted from training to inference. We're now moving from inference to agentic workloads. The organizations that are doing AI are changing from experts, the model builders, the big AI clouds to enterprises, all of that means that the weight is shifting from compute to data. And these organizations need to be spoon-fed, need something that's simple, they need it to be secure. They need to basically be told how to adapt these new technologies.

>> Two years ago you were on the stage, you launched the company's product. It's been worked on for a while. We had that conversation. A lot's changed since then, but the original North Star's kind of playing out. Architecturally you're starting to see the enterprise, certainly the hyperscalers and large- scale systems providers are all moving quickly, now the enterprise is heating up. Is the architectural change happening the way you originally saw it? Has it changed a bit? Is the market changed at all in terms of the orientation? A lot's happened on the chip side, but AI is certainly more hyped up than ever before.

Renen Hallak

>> Yeah, I think more and more the different parts of the architecture are starting to shine. When we were talking large-scale training, you needed fast access to a lot of data, but that was basically it. You didn't need resilience. It wasn't a critical workload. Latency didn't matter, distributed systems didn't matter. Now as we make this shift to inference, all of those things come into play, and especially as we add agents into the mix, that's a whole new dimension. Before we had people and data, now we have people, agents and data, and that introduces all kinds of questions around what data can these agents see? What data can they communicate to us, to other agents? How do we have observability into what they're communicating between themselves? How do we have reproducibility of how they gave that answer two years ago? All of these things are data related, and I think more than ever, we're seeing that data is also helping accelerate the compute aspects. Things like KV caches are moving the onus onto the data platform in order to save NGP resources.

>> So I want to dig into this. I wrote a piece last month, did a deep dive on VAS, called it Unpacking VAST Data's ambition to become the operating system for the thinking machine. And what I wrote was VAST is attempting one of the most audacious pivots we've seen from an infrastructure specialist in the last decade, talked about the $2 billion in software, et cetera, et cetera. The new AIOS layers distributed scheduling, event streaming, vector search, database indexing, and an agent execution framework on top of this. Disaggregated shared everything you call it tays.

>> Days. - Days, sorry. So I wrote why this matters. We think the AI error is going to be defined less by monolithic applications, more by distributed agentic, kind of what you were just saying, workflows that continuously embed, enrich and act on data might, when I first heard this, again audacious, I said, "Wow, operating system for AI. " I thought that was Jensen. So okay, so help us understand where you fit in that whole value chain.

Renen Hallak

>> So we work very, very closely with NVIDIA. In fact, all of our latest announcements were combined with NVIDIA. If you saw at COMPUTEX, Jensen gave us more than a minute of stage time in his keynotes. We are working very closely with his enterprise team. We're working very closely with the software teams. And so when you build that operating system, I like that analogy because it's right in the middle between the hardware and the application. And NVIDIA obviously built all of the hardware pieces, the GPUs, the DPUs, the networking for these new AI workloads, and they're also in that enterprise application layer. They're building NIMs and NeMos and blueprints, and we are there as that middle of the sandwich piece in order to enable all of this and bring it together. Why is that important? Again, as we move from a world of experts to a world of laymen, we need that operating system to make it simple, to make it widespread. I like to give the analogy of the personal computer. In the '70s, you needed a PhD in computer science to operate a computer. In the '80s and '90s everybody could do it because Windows made it simple for us to do it. That's the stage that we're at from an AI perspective, and that is the piece that the world, I believe needs in order to make this next leap. Now, why is a data platform a good place to start when building this operating system? Every one of these technological revolutions has a defining characteristics. For the PC, it was compute, for the internet, it was networking, for AI it's data. And so we're right in the best position to build out this stack and to make it easy for everybody to adapt.

>> So help us understand the data platform because it feels like you're not directly trying to take on a Databricks and a lakehouse or a Snowflake. And we heard certainly last week from Databricks in a lakehouse, we're going up stack. Snowflake is kind of moving some of the function out into the open source world. So when you say data platform, are you rethinking what that is?

>> Yes. - And would you explain that, please?

Renen Hallak

>> So if you look at older companies like Databricks or Snowflake, they were built in the era of big data, analyzing numbers, mainly columns of a database. And that has a certain scale to it, it has a certain ability to it. It's mainly CPUs doing that analysis. As we shift to deep learning or AI or especially these agentic workloads. We're talking about not numbers, but pictures and video and sound and genomes and a lot of unstructured data. We're talking about tens of thousands, hundreds of thousands of GPUs that are so much hungrier and require that much more parallel access to all of this information. We're talking about needing access to all of your historical data, not just the last two days of information. And so for all of those reasons, you need a new architecture and you need a new data platform to do unstructured data to give it structure and meaning, to drive these workloads in a data engine. All of these pieces need to be different for this new world than what those older companies built. 15-

Dave Vellante

>> So another way of thinking that is you're bringing in a new dimension of context, I would say. Whereas those historical platforms are historical systems of analytics, what happened and maybe why it happened, but not really good at what's going to happen and what's the next best thing to do. Is that the type of context that you feel that you can support? Not that you're actually, you're not a SaaS vendor. Maybe that's what up the stack, they're going to do that, but you're going to be the platform to enable that. Is that the right way to think about it?

Renen Hallak

>> Yeah. When you say up the stack, I would say you need to start from the bottom in order to do this, right? The fact that we started at the bits and bytes and understanding exactly where each piece of information sits on hardware, that's what gives us an edge because well, that's what makes us the operating system. If we were to start from the top and leverage cloud services underneath us, that would be very different.

Dave Vellante

>> So, John, what's interesting about this is you see this from Snowflake and Databricks, they are moving up the stack and they are moving into the domain of Salesforce and ServiceNow. I'm not hearing that from you, Renen. You're not trying to build an ecosystem around the governance layer and try to be the source of metrics and dimensions and do that. You're just trying to be the best infrastructure and software to enable that to occur.

Dave Vellante

>> That's exactly right. - So you're not trying to take those up the stack folks on.

>> Yeah, I mean this comes up a lot. I want to get your thoughts on this because what Dave's mentioning is that the old world, old world, a couple of years ago, two years ago, data analytics drove everything. You mentioned some of the things around how they evaluated the data, and then you mentioned the windows. So you mentioned also KV Cache. So at the layer of you're at, there's a lot of networking too. So software is now controlling where things go. You mentioned scheduler events. It sounds like an operating system to me and first principles. So the question is what has to change in the enterprise or people who think in a data analytics world because you have disruption and enablement going on at the same time. So you're seeing that in software development and you're seeing it in sales and marketing kind of obviously low use cases where AI shining today and agents are kind of picking that up too. But the analytics in the data world have the dogma and the old school thinking around, we built these systems to analyze, to build dashboards. This is not like a cottage industry. This is pretty big and sizable, huge investments. What do enterprises need to do and what's the connective tissue that you're dealing with at your layer that's going to enable that transformation from analytics and intersect and bring in gen AI? Because gen AI is coming in like a freight train Into the analytics world, and those people not necessarily aren't platform engineers. They're thinking business logic, UI, application level agents. They're not really thinking about the plumbing. So as an OS, what do you think about that and what has to happen for the enterprise who are hugging onto their analytics systems like a tree hugger?

Renen Hallak

>> Yes. So there will be connectors from the old world into the new world. Some people call these MCPs, Salesforce calls it A2A and Google does. It's a way to give agents access to older information and to data sources that are already existing. But I think if we think of this new era as data analytics, we're doing it a misservice or a disservice. These new agents are the new applications. They're half new applications, they're half artificial people. And so it's a new concept that is emerging and they will do tasks for us. They will go out and study something. They will learn new things. They will come up with eventually new ideas as they operate on our behalf. And so they'll need access to data, but it's not an analytics workload. It's all of the workloads when we're talking about legal work or accounting work for the white collar type of agents, if we're talking about robotics, it will be everything. Construction, workers, gardeners, these agents will literally have a life of their own and we need a way for all of them to communicate, for all of them, to share thoughts for all of them, to organize their internal and external information. And that's where this-

>> So data brings up the analytics and the governance. I mean the governance piece, which sits to enable that. How are you guys enabling that? Is it origination of where the data's stored? Because you're getting closer to the NVIDIA low-level piece, so you have access to where the data comes out versus post, it's out there and I'm going to analyze, it's in a corpus or whatever the format is for analytics. As you get closer to where it's originated, you can see things. There's metadata, now you've got S3 tables and AWS. You've got Databricks with Iceberg. So now you have a storage paradigm of how to store it, which is open.

Renen Hallak

>> Governance is super important. I think one of the big reasons why the enterprise isn't moving faster is because they have compliance and regulation and other things that they need to adhere to. For example, access control lists, who is allowed to see what information? If you're talking an object store and then you start to vectorize the information, you need the same security model across your unstructured data and the rows of the database. If you're talking agents, you need to make sure that this agent is not seeing something that they're not allowed to and that they're not communicating something that they are allowed to see. Especially as we move into this world where we have, today it's all static. It's one AI, it's a version, whatever it is that OpenAI came out with, and all the agents are running on the same model. Over time, each of these agents will start to fine- tune their own model and we will have different AIs interacting with each other. But that introduces another challenge for governance because all of the data that they use to fine- tune is potentially going to leak through interaction with them. And so we need to remember what that data is. We need to know who is allowed to see it. We need to be able to observe what happened and where all of these models came from in order to-

>> So knowing the storage, knowing the origination plays into the knowledge of the lineage.

Renen Hallak

>> All of this goes back to metadata and an understanding of where information-

Dave Vellante

>> So to put a finer point on, again, where VAST sits, if I'm in the VAST system, you take care of that. But now above you, there are other governance, there's Unity, there's Polaris, there's Calibra, there's Alation. So you will surface, right? But you'll surface. They have access to your metadata, you will provide that to them so they can then go do their job. They'll hand that off.

Renen Hallak

>> That's exactly right. And we make sure that it's secure and that it's simple and that it's real-time and they can focus on building their applications and building their agents.

Dave Vellante

>> So your TAM expansion strategy is not to go up, it's to go-

Renen Hallak

>> Wide.

>> Wide. Yeah. - Talk about the conversations you're having with customers. You're talking off camera. I won't say the name of the company, but big financial services, you guys are talking to everyone. Certainly you had great run with the hyperscales needed to do business there, but then you have the next layer of, I call, I won't say tier two, but tier one B, large-scale enterprise, and then you have the kind of the mid-range. What are some of the conversations you're having with them? Can you share any stories or anecdotes around the thinking? Are you whiteboarding stage? Are they under pressure? Obviously with agents exit, the complexity increases, but now functionality and business value increases. So I'm sure you're being brought into not just technical conversations, but business conversations.

>> Yes. - Share some of those conversations.

Renen Hallak

>> We speak more and more to the AI people versus the infrastructure people because they are driving the agenda and they have been playing around for the last year and a half and now they're realizing the real value and they need to go into production.

>> So they're locking in on visibility of use cases and architecture.

Renen Hallak

>> Yes, these big banks have lists of hundreds of applications that they need from an AI perspective, and they are in the process of shifting from playing around in a sandbox to moving things into production. And at that point, you need access to all of the data for this to work, and you need all of the governance and compliance that we discussed a minute ago. And so it makes it a real challenge for them to do that with the old infrastructure. And that's why with a renewed sense of urgency, they're coming to us and asking us to help.

>> What are some of these with the old infrastructure? Is it latency? Is it the number of hops? Is it speed architecture? What are the key constraints that they're dealing with?

Renen Hallak

>> It's scale. When you look at the older workloads, you had a thousand people in the organization and you needed a million row database in order to do that. When you think of agents, you're going to have a thousand agents per person. That is much larger scale. When you think of the type of information that now can be analyzed. It was stored before, it was documents, it was pictures, it was videos, but nobody analyzed it. Nobody knew what to do with it. Now they can. And so that level of scale and performance that is required is several orders of magnitude more than what the old stuff was built for. And that's why-

>> They kind of knew what they wanted to do it but they couldn't do it, right?

Dave Vellante

>> Coming back to your comment about the sandbox, that sandbox is oftentimes in the cloud and then they bring it on-prem. They're actually doing some sandboxing on-prem. Oftentimes they don't have, whether it's the center of excellence or the skills. Now the big banks have the resources to do that. How do you see that playing out, Renen? Because you've got the big banks, even they are constrained on power, on liquid cooling. And so how will that evolve? I mean, financial services tends to be a harbinger of what's coming cloud and then the financial services and then the rest of the enterprises. There's a big debate, not even a debate, but there's a dissonance in you don't really need GPUs to do this. Others say, "Hey, inference is a lot of matrix math. You actually need GPUs."

Renen Hallak

>> You do.

>> What's the fact from a technical perspective?

Renen Hallak

>> You need GPUs, especially for the first half of inference, what's known as pre-fill. And now we have disaggregated inference. So we're disaggregating pre-fill from decode, which is basically understanding the question versus giving the answer. You need a lot of GPUs, especially as we move into these reasoning models where inference is not a one-shot. I'm just going to give you an answer, but it's a multistep process. It's a lot of trying things out and seeing what works. And so inference is going to be the majority of the GPU consumption going forward. Training, I would say is something that you need to do once, you get to a certain baseline and then inference takes over. I think the big banks are, as you say, a harbinger. I see three phases. The first phase is playing around in the hyperscalers and the big clouds. Once they get to a certain level of scale, they realize the big cloud infrastructure wasn't built for AI. And then they start talking to some of these AI clouds like CoreWeave and Lambda and Crusoe, and most of them will sign contracts with those for this intermediate stage until they're ready to bring it and bring it in-house.

Dave Vellante

>> And what about the fourth wave, which is like everybody else? Are they going to be able to afford that? When does that go mainstream?

Renen Hallak

>> I think everybody else, as in smaller organizations, will consume AI applications more than build their own agents. And so my guess is very large technology companies like an Adobe or a ServiceNow will be the ones that will help them make this transition.

Dave Vellante

>> And that will drive demand for cloud. A lot of those guys might decide to do some of their own data centers as well as they already are.

>> They already are. - You're enabling the ecosystem above you. What's your view of that? Because for you to be successful, you have to enable which you're talking to, what happens for them? They have to have agents that talk to each other. You mentioned that. We've been saying that on theCUBE for a couple of years now, Dave, and so that's happening now. What does that ecosystem look like and what are some of the requirements for success from your opinion around what that looks like? Partnerships are changing. You mentioned MCP that's kind of creating kind of this API vibe around, Hey, we can now talk to each other. What has to happen? What do you need to do to enable that? And what does the ecosystem above you need to look like?

Renen Hallak

>> Yeah, so if we are going to be the operating system, we need to partner with application developers on top of us or agent builders on top of us as well as the infrastructure people underneath. And so we talked about partnering with NVIDIA, partnering with these clouds, partnering with companies like enterprise in order to make our way into the enterprise. But we really want to enable agent builders. And so what we're now starting to do is write our own agents. To just give examples, they're all open source and we can show how easy it is to do that.

Renen Hallak

>> - We are.

>> So I have to ask you, this comes up a lot. We certainly debate on the queue part every Friday pretty much at this point. Is agents going to replace SaaS or is that an evolution in your mind? Because there's a lot of SaaS companies out there.

Dave Vellante

>> Agents talking to CRUD databases, or is that clippy?

>> Well, I mean there's an architectural shift, but I mean, the cloud era proved that SaaS could work and got app stores, and we all know the history there don't need to revisit it. But now agents are going to hit the scene with huge numbers. Is that an extension in evolution of SaaS or is that kind of a wholesale replacement in your mind?

Renen Hallak

>> I think it depends on the company, whoever they are. If they make this transition and start to take advantage of these new abilities, then they'll be fine. And if they stay where they are very quickly, they will be over-

Dave Vellante

>> And they have some real advantages in that they have the data, the metadata, the business logic, the process logic. And so if they lean in, you would think they'll be able to at least maintain their franchise. You'll see how far they can go. When we get a technical visionary like yourself, I want to ask you about, you mentioned MCP and A2A. My understanding is that MCP is stateless A2A is state. A lot of people think it's like competing standards, maybe it is and eventually will become that. How do you think about those two emerging standards, because both seem to have momentum.

Renen Hallak

>> Yeah, it's early days. I never make a bet on one standard versus another. I look at the concepts behind it and the ideas behind it. I think the idea of both of these standards is to put some type of structure formalization behind how agents communicate. I think it's super important not to confuse agents with web services, which I've seen a few people do as a result of these types of standards. I think agents are long-lived things. They have long-term memory. They have experiences, especially as we move into these physical agents and robots. They have sensors that take in pictures and sound, and maybe one day somebody will develop a good smell sensor and they develop thoughts. And so it's unstructured data coming in. It's understanding what's in it, it's giving its structure, and it's comparing that to old thoughts and old memories as we generate new ideas.

Dave Vellante

>> As opposed to hard-coded sort of microservices that are static. This is an organic system.

Renen Hallak

>> It should live. And in some sense, the vast operating system should be the habitat for these new organs.

>> Yeah. Jensen made a controversial comment, Dave, we obviously covered it at CES when he said, " It will be the HR department for agents tongue in cheek," but he's actually not wrong. So what came out of the Databricks event was the focus on evaluation where you have reinforced learning with human feedback, but the human feedback is being statistically calculated as you're managing your kid. How'd you do in school today? How are your grades? So this supervision of agents going into the pre- agent evaluation before they're deployed. So that's almost like an HR department, like, Hey, how'd you do?

Renen Hallak

>> I think he's right.

>> What's your thoughts on this? Because this is where we're seeing the math and the computer science really shine on agent deployment. Evaluate first, understand the task, agents supporting agents, supervisor. And we even had Augment on and some code companies talk about the fact that you have human supervision where the developer becomes the tech lead. So the code's being generated by machines. So this idea of supervised, kind of an HR department.

Renen Hallak

>> It is, and I wouldn't call it HR, it would be AR. But yes, I think-

>> Agent relations.

>> Resources. - Resource agent resources.

Renen Hallak

>> I think humans have a really important role in teaching agents what we know. And then once this reinforced learning happens, I think it takes a life of its own. It's not just reinforced learning by human feedback. It's reinforced learning by agent to agent feedback. It's reinforced learning by interacting with the natural world through these sensors. And if you really can over time, fine tune your individual model, each of these agents will have a different model of the universe. And that's how ideas-

>> Yeah, you're enabling this. And I wanted to bring it up because one, it was fun to reference the Jensen thing, but in the computer science world, this is real science going on. This isn't just random evaluation. So folks who are afraid agents, agents will go off the rails and turn into Terminator, do some weird things, ship code into production that's not supposed to be there. So this is where, can you share your thoughts on why the computer science is important behind this one point?

Renen Hallak

>> Well, I think that's where the security aspects come in, and we have to be in control as much as possible. But more than that, we have to know what's going on. We have to be able to observe communication. We have to be able to look back in time and see what happened. All of these things come back to a persistent layer that captures all of this information and for us to be able to read that information in an easy way. And so we need to make sure that it doesn't go off the rails, but there's a lot of good things that can happen by us giving them a little bit of freedom.

>> Thank you so much for coming on. Again, you're a leader in the AI infrastructure and obviously that impacts robotics, which is an edge. We didn't talk much about that, but continue to nail that data platform, that operating system, getting close to the hardware, orchestrating scheduling, managing events, originating the data, all part of the big wave. We're on. Thanks for coming in.

>> Thank you. - Appreciate it. Okay. The co-founder of Vast Data is in the house here really sharing his thoughts on the enablement required to bring in the agent era that we're seeing, which interests more value, but complexity. And if you get the data in control, the infrastructure will do great. This is theCUBE bringing all the action. I'm John Furrier with Dave Vellante. Thanks for watching.