Snowflake Summit 2026 | Vivek Raghunathan, Snowflake & Matt Luizzi, Whoop

Clips
More from Snowflake Summit 2026

Vivek Raghunathan

SVP of Engineering

Snowflake

Matt Luizzi

VP of Analytics

Whoop

play_circle_outline How Snowflake Powers WHOOP’s Shift from Athletic Wearable to AI-Driven Health Platform

play_circle_outline Raw sensor data types: heart rate, HRV, respiratory rate, blood pressure, ECG

play_circle_outline Snowflake CoWork for Business Users and CoCo for Developers: 3+PB Lake, Iceberg Interop, Polaris Governance

play_circle_outline Observability importance and Snowflake's Observe acquisition

play_circle_outline Using Snowsight for agent runtimes, costs, token consumption monitoring

Info
Transcript

Vivek Raghunathan, Snowflake & Matt Luizzi, Whoop

Vivek Raghunathan

SVP of Engineering Snowflake

Matt Luizzi

VP of Analytics Whoop

In this interview from Snowflake Summit 2026, Vivek Raghunathan, senior vice president of engineering and support at Snowflake, joins Matt Luizzi, vice president of analytics at WHOOP, to talk with theCUBE's Dave Vellante and Rebecca Knight about the path from fragmented data and AI initiatives to governed, production-ready agentic systems. Luizzi details how WHOOP manages over three petabytes of sensor data — growing at 20 terabytes per day — and relies on Snowflake as the foundation for business analytics and accelerating insight delivery. He explains why i... Read more

explore Keep Exploring

How would you describe WHOOP today, and how is the company using Snowflake? add

What sensor data is collected from the wrist, and how is that raw signal data processed into metrics such as heart rate, respiratory rate, heart rate variability, blood pressure, ECG, and irregular heart rate notifications? add

How does Snowflake's architecture help a company like WHOOP operationalize its sensor data — does that data flow through Snowflake CoWork or CoCo, and how is governance managed (for example, using Iceberg and Polaris)? add

What was the rationale behind Snowflake's acquisition of Observe, and how will customers (e.g., a customer like Matt) be able to take advantage of that integration for observability—will Snowflake provide integrated observability built on its analytics platform and what is the intended roadmap? add

What benefits have you observed from implementing Cortex agents on Snowflake (using Snowsight) for your agentic workflows, and what challenges do you anticipate as agents become more autonomous? add

bolt Powered by CUBE AI

Vivek Raghunathan, Snowflake & Matt Luizzi, Whoop

search

Dave Vellante

>> Here we go.

Rebecca Knight

>> Good afternoon everyone. And welcome back to the Moscone Center here in San Francisco. We are theCUBE and we are live at the Snowflake Summit 2026. I'm your host, Rebecca Knight alongside Dave Vellante, co-CEO of theCUBE. I would like to welcome our next guest to the show. We have Matt Luizzi, VP of analytics at WHOOP. Welcome back, Matt.

Matt Luizzi

>> Thanks for having me again.

Rebecca Knight

>> And Vivek Raghunathan, SVP of engineering at Snowflake.

Vivek Raghunathan

>> Thanks for having me.

Rebecca Knight

>> Thanks for coming on again. This is great. So Matt, WHOOP, it's a very cool, very hot company, right, in our hometown of Boston. You sit at the intersection of biometrics and analytics and AI. Talk a little bit about the company now and how you're using Snowflake today.

Matt Luizzi

>> Yeah. I mean, the company started as a wearable device for elite athletes. And we're very much transitioning to a company focused on health. And being able to take that holistic picture and use that to generate proactive insights for our members is something we care deeply about. So as far as how we use Snowflake to get there, we're capturing a ton of sensor data. This is data coming off the risk that we're processing, storing in our data lake. And then we want to use analytics to improve the product experience, understand how to get the product into the hands of more customers. And Snowflake sits at the center of that. We power our entire business analytics function by Snowflake. And more recently with AI, it's really transformed the way that we're able to accelerate our insight delivery and prove value to our members.

Rebecca Knight

>> So talk a little bit about the data that you're picking up from the wrist. What exactly are we talking about here?

Matt Luizzi

>> Yeah. So we're capturing raw sensor data. And then we have a signal intelligence and data science team who's responsible for creating algorithms that translate those raw signals into scores and metrics. So if you think about something as simple as heart rate, that requires an algorithm to translate signal into a heart rate. Now we have heart rate, we have respiratory rate, heart rate variability. We've recently launched blood pressure as well as ECG and irregular heart rate notifications. So the amount of insights we've been able to develop with first party algorithms has been truly great.

Dave Vellante

>> So how do you think about that, Vivek, from the standpoint of the architecture of Snowflake helping WHOOP, for example, operationalize, taking that sensor data. Does that go through CoWork? Does it go through CoCo? How does that all flow?

Vivek Raghunathan

>> Yeah. I think of where Snowflake plays the most important role in this ecosystem is helping customers like Matt use their data to make their AI go faster or to use their AI to make their data go faster. And so across the board, we've tried and built Snowflake as something that's interoperable from the get go. So we embrace standards like Iceberg and Polaris, which I believe you guys are on. And our thesis is instead of you fragmenting your AI experience in multiple ways, we will extend the data platform to be able to do lots of the things you would do like on top of the data inside of Snowflake. Snowflake CoWork is aimed at business users looking to extract insights from their data. So this would be someone like me wakes up in the morning, says, "Hey, the amount of code written yesterday was down like 20%. Why was it down?" Snowflake CoCo is aimed at developers. It's aimed at builders. It's aimed at people who are trying to build out the patterns, build out the models, build out the skills, build out the intelligence that the business users in the enterprise will use. And so at every company like WHOOP, they're both personas of customers and they end up using Snowflake CoWork and Snowflake CoCo the way we imagine it.

Dave Vellante

>> And Matt, Vivek said you guys are using Polaris as the governance catalog.

Matt Luizzi

>> We are.

Dave Vellante

>> Talk about that journey to Iceberg, Polaris. You've got, of course, Horizon has additional function, but you're not leaning into that. So how do you sort of manage the governance and how do you think about that?

Matt Luizzi

>> Yeah. Just given the scale of data we have, there was a WHOOP presentation yesterday by some of my data platform colleagues where they shared that we currently have over three petabytes of data in our data lake with 20 terabytes added per day. So just a tremendous scale. And for us, interoperability is extremely important. We need to be able to evolve over time, make sure we're taking advantage of the latest technologies. And being able to partner with Snowflake on that where they've been good partners understanding what's working, what can we do to improve that product experience for us has allowed us to feel like we're meeting the needs of the business, which is really at the end of the day, what we need to solve for on my role.

Dave Vellante

>> So every conference we go to, Rebecca, we're talking about agents and Agentic and we're starting to realize how much more data that's going to create. So with that volume of data sounds enormous, but with agents, it's probably going to grow even more. You remember the days of Hadoop, you probably could have done all the big data stuff in Oracle, but you needed something that was cost-effective to do that. How do you think about the data growth and managing that data growth governing all that data in sort of the near to midterm?

Matt Luizzi

>> It's a great question because part of me wants to say I don't have to think about it because Snowflake just enables that scale. And it's just more around the number of assets and governance. And given Snowflake strong role-based access controls and all of the governance that we get out of the box with Snowflake, a lot of that cascades down to the end users in a way that we don't really have to think about it. And as the amount of data scales, it just allows us to increase resources as needed and use all of the Snowflake observability functionality to let us know when the time has come.

Rebecca Knight

>> So Vivek, one thing we always hear is that enterprises want AI to help their employees to free up their time for more strategic thinking, give them more speed, increase their productivity. What does that look like in practice when you're thinking about a company like WHOOP?

Vivek Raghunathan

>> Yeah. I can start with a company like Snowflake because one of the things we've tried-

Rebecca Knight

>> You are a customer zero.

Vivek Raghunathan

>> One of the things we have tried very hard to do in the last half a decade or longer is be the biggest internal dog fooder of the Snowflake product. Our internal instance of Snowflake, we call it Snowhouse, is probably the biggest Snowflake account in the world. So how I see that transformation look like in practices, there's often a maturity model. It's very easy for us to think binary, like this is a AI native company and a non-AI native company. What I find on every dimension, whether it's processes, products, adoption, people, there's a maturity curve, right? I'll give you some examples. We use Snowflake extensively, extensively to do all of our observability. So all of our observability sits in our ... The first, like what I'd call maturity level 1 would be our engineers writing skills to extend Snowflake CoCo to all of the most important user journeys that someone might want to do when they're trying to do observability of their data. You get there, the second step is, "Okay, I got all these, I got thousands of skills. I know how to do these user journeys. Now how do I do event-driven AI? Some Slack notification comes in, how do I trigger this journey?" Then you get more into like a workflow based orchestration system where you start to automate these things and humans are supervising stuff more. And this is the vision that Christian and Sridhar talked about when they talk about an agentic enterprise. We think we're going to move from us doing work and writing SQL queries to us supervising work that agents do and kind of directing them and coaching them. The notion of the TL of agents is starting to become very popular. The fourth step then, this is like what I call maturity level 4 is a continually learning system, like a multiplayer coding agent that as people in match teams use Snowflake, it gets better for the other folks, of course, with our back and governance and security all built in, right? That vision, maybe 20% of my teams currently internally using Snowflake are on. Step one and 200% are on. And we're like trying to get more people down that maturity curve.

Dave Vellante

>> So the observability piece gets really interesting here. Of course, the observe acquisition. As the previous question about scale of data, the volume of data and you bring in agents, you're going to have to be able to observe what they're doing. Again, it sounds like, Matt, you're comfortable. Let Snowflake do the observation now that you've got to observe. Is that the logic behind the observe acquisition? How does a customer like Matt take advantage of that? Are you thinking along those lines? I'd love to tap that a little bit.

Vivek Raghunathan

>> We've always thought of observability as very adjacent to analytics. It is an area where it is often platformed on top of an analytic engine of some form.

Dave Vellante

>> It's a form of analytics.

Vivek Raghunathan

>> And we were close partners with Observe and taking up to the acquisition. So it's a very natural thing for us to acquire them and integrate them because lots of our customers find themselves wanting observability as a first thing after analytics. Directionally, where I see that we're going over a one, two year timeframe is we're going to start unbundling things, right? We're still going to start having headless capabilities that agents can orchestrate over, that Snowflake CoCo or Snowflake CoWork can start orchestrating over. And that direction I think will be the direction of travel for use cases like observability or security or what have you that are built on top of like analytic capabilities. We would love to partner with companies like WHOOP to be able to address some of those use cases in the coming months and quarters.

Dave Vellante

>> How do you think about observability today? I mean, I'm inferring you weren't an Observe customer prior to the acquisition.

Matt Luizzi

>> Correct.

Dave Vellante

>> What's your observability approach? I mean, do you have another tooling that you use?

Matt Luizzi

>> Yeah. For us, the benefit of our implementation of Cortex agents has been we exclusively use Snowflake for our agentic workflows for our business analytics team. And that allows us to take advantage of all of our data being in one place as well as all of our observability being in the same place. So just taking advantage of the native features that come with Snowsight and being able to see agent runtimes, agent costs, token consumption, and start to put more real guardrails around evals and accuracy. Doing all of this directly in Snowsight's been very helpful for us right now. Granted, we're in the early innings. And as we start to see more autonomous agents be deployed, I think it's going to need to mature a little bit more as there's going to be fewer humans in the loop and more autonomous agents taking on work on their own.

Dave Vellante

>> I think so. I think that's why the acquisition was interesting because I mean, I look at companies like Datadog and I say, "Wow, obviously they're doing very well, but I don't think people understand the importance of observability in that world where it's opaque what agents are doing." Kind of off-topic, but it's something that I think we want to be thinking about as we go forward. And you were describing the personas for CoCo and CoWork and you also talked about this learning system through that maturity model. So I'm really curious about that, like how you think about architecting a learning system, what it means to a company like WHOOP that's got all this data. It's kind of a closed loop system that we're thinking about here that begins to learn and infer and enact based upon the reasoning traces of, I guess, humans. How are you guys thinking about that?

Vivek Raghunathan

>> Yeah. I think look, I came from a world of search. I built a search engine that I sold to Snowflake. It was called Neeva. Collaborative filtering has been around for time immemorial. I don't think of these as very complicated problems. I think of these as taking a lot of natural well-known techniques and applying them to new domains. I think the notion of a system that can personalize starting off with you, the end user, and over time extending to your work group. So it becomes that ... We talk about tribal knowledge in the enterprise. Tribal knowledge is all the activities that you and your co-workers do. I think long-term, the coding agents you use like Snowflake CoCo, Snowflake CoWork, are going to capture that tribal knowledge in a governed way, in a secure way, in a way that is with consent, has all the guardrails that you would expect. But in terms of how the mechanics of these things work, I think of this very much the way I think of all systems that did personalization in the past. There are collaborative filtering systems. There are systems that are models that are built on your past session in user history and they make your experience dramatically better.

Dave Vellante

>> Well, I think three or four years ago, I don't think Snowflake was thinking about going after that sort of tribal knowledge opportunity. And I think the realization is that's where the opportunity is today because we're going to sort of automate a lot of the coordination tasks and we're seeing that in real time. Then it goes on to judgment. Well, how do you automate judgment? You need tribal knowledge, access to process, knowledge, the underlying application, semantics.

Rebecca Knight

>> And also the continuous learning, because the institutional knowledge is not static. I mean, things are constantly changing as they should be.

Dave Vellante

>> Okay. So how do you think about that internally at WHOOP? I mean, you've got a cleaner slate than, say, we had Thomson Reuters on earlier, right? You could probably move a lot faster than they can. When you sort of AI-ify, you don't have to necessarily come at it from the edge, you can kind of go at the core. Is that how you think about it?

Matt Luizzi

>> Absolutely. We've been putting a lot of effort into generating that clean semantic ontology over the past couple of years. And that's really enabled us to take products like CoCo Desktop and immediately dive in and see value. And what we're seeing now is a shift in where humans are able to add value and where they're needed to add value. So if you think about just getting this tribal knowledge into a skill, that's something that the team spent a few weeks on upfront when we first got access to these products, started to see a lot of success. And now what happens is it just comes time for the human to add judgment and make those decisions at the end of the day. But if you want to talk about where humans want to add value, it's going to be on the strategic decision making, the high leveraged tasks. And we're starting to see some early signs that we're really automating a lot of those more trivial workflows that just require tribal knowledge and manual work and start to repurpose these same humans on these strategic value add tasks for the business. And what we're not seeing is the lack of need for humans. We're seeing all this more work kind of propagate and we're able to do a lot more things as a business, move faster, and ultimately deliver a lot more value to our customers.

Dave Vellante

>> I mean, I think we all see that personally. AI is just making us busier because we have more fun stuff to do. Our backlog is growing. But the leading organizational thinkers now, they'll talk like McKinsey and Bain and BCG. They talk about flattening the organization. But more than that, it seems like the org chart, which is today built around hierarchies of humans, will increasingly be built around intelligence and agents. Do you buy that? What does that mean for you? Is that too futuristic?

Matt Luizzi

>> It's funny you say that. Back when we first started using some of these more autonomous coding agents like CoCo CLI, CoCo Desktop, I was chatting with the Snowflake product and engineering teams and saying, "This first agent you interact with is almost the CEO. It's just going to understand the business and not know too much about the details. And it's going to have to delegate to some other more specialized agents like your chief product officer who's then going to go to their subagents and get tasks done." I do think that we're starting to naturally see that develop because humans are taking their knowledge, putting those into skills, enabling agents to have access to those skills. And it's almost naturally forming an org chart just based on where the tribal knowledge is persisting as markdown files that are available to agents. So I think it's an interesting take on it. And I think it's naturally happening that way.

Dave Vellante

>> Yeah. I mean, it seems like that necessitates. Obviously, you want good data, sure, but then you're going to have a knowledge substrate that, I'll say it this way, that avoids silos of automation and sort of flattens that knowledge base across the organization. That seems to be the future of where all this productivity is going to come from because otherwise we're just going to create all a bunch of silos of really cool AI workflows. And that's not the answer, is it?

Matt Luizzi

>> Yeah. And when you say flat in the organization, the way that I've been positioning it internally at WHOOP is almost compressing it horizontally, right? The lines between what a product manager is, what an analyst is, what a designer engineer is, the lines are blurring. Traditionally, those were very distinct roles because they required highly specialized expertise in those specific tasks. But I'm not an engineer by trade, now I can write code. You're not a designer by trade, but you can create a prototype just by going in and working with agents. And I think that's where we're going to start to see a lot more change in the coming years.

Dave Vellante

>> Do you think the highly specialized roles that we saw come out of the big data era, data scientist, data engineer, business analyst, data quality engineer, will they also collapse horizontally or will they just sort of get better at what they do?

Matt Luizzi

>> I'll defer to Vivek on this one, just given what Christian was talking about on the keynote this morning with these pre-bundled skills and all these things coming out of the box. So maybe you can take that one.

Vivek Raghunathan

>> Yeah. I'll actually take a spicier take on your .

Dave Vellante

>> Yeah, let's hear it. Great. Hit us.

Vivek Raghunathan

>> I think the focus on the shape of the org chart is itself maybe a little premature, right? I think a lot of what we are seeing is here to start from first principles and ask what are the jobs to be done. What does this function in the organization do? What are those end-to-end activities? Some of those activities will be activities where the input is well-defined and structured and the output is well-defined and structured. And if that's the case, then an agent will do that activity, right? Some of those cases, the input will be well-defined, output needs some judgment, and there you'll have what we call human in the loop. And there'll be a third class of activities which are the input is fuzzy, the output is fuzzy, what I'd call fundamentally human activities. And I guess some level we have to start from first principles, define what are the activities each of these functions do and say, what are the things that can be completely autonomous and what are the things that need you to exercise judgment and what are the things that need you to exercise creativity? So I think we'll end up reorienting the work that me and Matt do and others in our teams around what I think of as the residual work, the work that is not fully identified. In that world, I think you will have both the collapsing horizontally and the collapsing vertically, but you might have very different roles. You may have roles that map to a job to be done versus a functional thing that somebody's very good at. And that I think is something we're all going through. I think Andy Grove would say this, "You need to let chaos reign before you reign in the chaos, right?" And I think we're right now very much in the let chaos reign phase of the transformation. And so as leaders, I tell my leaders, it's very important for you to not reign that in too quickly. You need exploration to happen, you need to see... And Sridhar will say this, in any transformation like this, the heroes will self-select. There'll be a bunch of curious people that will go and explore. And you need to find those explorers and amplify them and create almost a pedestal for them to stand on and become famous in your organizations.

Dave Vellante

>> So in your scenarios one and two, the first one seems like low hanging fruit and maybe even can be dealt with with a lot of today's deterministic software, maybe it can be improved. The second piece requires, as we were talking about before, a learning loop as Rebecca was saying. And the third is it's out there and it's complex, but it seems to be the direction that you're headed and where organizations are headed. That judgment piece that's fuzzy still going to require humans in the loop, but more and more of that we keep saying tribal knowledge will be, well, I guess be executed by bringing the deterministic and the probabilistic together. You are both of those ends. And then in the middle is this intelligence layer that's emerging that kind of nobody has nailed. Palantir ontology as an example, but that's a heavy lift with forward deployed engineers and-

Vivek Raghunathan

>> And I would like to think of Cortex Sense as our first steps in that direction.

Dave Vellante

>> Yes.

Vivek Raghunathan

>> Fun story. Cortex Sense was a December hackathon project by the guy who was the lead for both Snowflake CoWork and Snowflake CoCo. He first built Snowflake CoWork and then he was off building Snowflake CoCo. And on a lark he felt like it was complicated to curate these skills and create these semantic views and these ontologies. And he said, "Let me try automate some of this stuff. Let me have work with a coding agent to actually figure out how to do this." And that exploration, one thing led another, now we have Cortex Sense, which is twice as good as like a vanilla coding out agent out of the box on hard data tasks. I think that is the direction of travel that that intelligence layer would take. I think it'll be a combination of what I call, I don't want to use the word context, everyone uses it. Enterprise intelligence, the word you use. And I would say Cortex Sense is that. It is how do you take all of the stuff happening in enterprise and just package it up in a way that coding agents can use in a structured way. The second thing I would say is what I'd call multiplayer personalization in the enterprise. So what you called a continuous learning loop, like something that improves but across roles and people in the enterprise. And the third thing I'd say is bringing all your data to your coding agent in a governed way. And that's why we acquired Natoma because we feel like I talk to enough CIOs and CISOs who see MCP and see all this value of connecting all your data to a coding agent. There's part of them that also goes, "Wait, how do we govern this? There's like five Jira MCPs in my enterprise and seven ways to connect to Salesforce." So how do we do that in a governed way that gives enterprises the confidence to bring all this tribal knowledge into one place and just help amplify the organization?

Rebecca Knight

>> Exactly. We could talk about this all day.

Dave Vellante

>> Oh, are we done? Oh, we are.

Rebecca Knight

>> that.

Dave Vellante

>> Because I was late.

Rebecca Knight

>> Thank you guys so much for coming on the show. I appreciate it.

Matt Luizzi

>> Thank you.

Dave Vellante

>> Thank you guys.

Rebecca Knight

>> I'm Rebecca Knight for Da Vellante. Stay tuned for more of theCUBE's live coverage at the Snowflake Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.