theCUBE + NYSE Wired present the East Coast AI Leaders Executive Series | Andy Byron & Pete DeJoy, Astronomer

Clips
More from theCUBE + NYSE Wired present the East Coast AI Leaders Executive Series

Andy Byron

Co-Founder & CEO

Astronomer

Pete DeJoy

SVP Products, Co-Founder

Astronomer

play_circle_outline Balancing Agentic Systems and Productivity in Enterprise AI: the Evolution of AI Workflows

play_circle_outline Building a Strong Data Ops Foundation for AI: Collaboration and Co-location in Generative Workflows

play_circle_outline Maximizing Data Security and Cost-Efficiency for AI Models with Northern Trust and Successful Customer Use Cases

play_circle_outline AI unlocking new value in data, shifting data engineering paradigm

play_circle_outline Maximizing Data Decisions: Shift Left/Right Workflow and Unit Economics for Differentiated Outcomes

Info
Transcript

Andy Byron & Pete DeJoy, Astronomer

Andy Byron

Co-Founder & CEO Astronomer

Pete DeJoy

SVP Products, Co-Founder Astronomer

During theCUBE + NYSE Wired: AI + Cloud Leaders Media Week event, theCUBE Research’s top industry analyst John Furrier sits down with Andy Byron, CEO of Astronomer, and Pete DeJoy, the company’s SVP of Products. They discuss how AI’s rapid evolution is reshaping business models, tech infrastructure and the broader data ecosystem.

Astronomer empowers organizations by orchestrating data pipelines essential for production-scale AI, Byron and DeJoy shared. As AI becomes more democratized through pre-trained models, they emphasize that clean, well-managed... Read more

explore Keep Exploring

What are the key factors for driving strategic outcomes using AI in businesses? add

What is the importance of a strong data ops foundation in moving forward with real production AI use cases? add

What is the importance of a strong data engineering foundation when working with AI engineering for domain-specific outcomes? add

What problem are we particularly excited about solving in the future? add

What is the importance of understanding the unit economics of data infrastructure in relation to the cost of deploying data platforms? add

bolt Powered by CUBE AI

Andy Byron & Pete DeJoy, Astronomer

search

>> Hello, welcome to theCUBE's here, the NYSE Studio, partnering with the NYSE Wired committee and Brian Baumann building out a great network with theCUBE. This is our East Coast presence. We are here with Andy and Pete here with Astronomer, we're going to break down all the action of CUBE alumnis. Last that we up, we talked with guys. Great. Thanks for coming on. Appreciate it.

Andy Byron

>> It's great to be here, John. Appreciate you having us.

>> So give us the update. You guys were just recently on at re:Invent, Amazon's annual conference just last year. One's coming up again, we haven't really touched base. What's the update? Obviously gen AI's hot, but give us the update real quick then we can get into some of the AI questions.

Andy Byron

>> Yeah, John, I like to use the analogy, I think 10 years ago Andreessen Horowitz came out saying, "Software's eating the world and I think data's feeding the world." We're right in the middle of that right now. So what we do is central casting and really I'd say the central nervous system for delivering data to hydrate these large AI models. And we're seeing the explosion of that happen in our business right now.

>> Yeah, I mean your customer base has been doing quite well. There's been a little hoarding of the GPUs, but still, I mean you see the sector, it's a secular trend, obviously. I love the supercycle word. Dave and I love that. Some people don't like it. I had a previous guest on Arcus and they're all school networking guys. We were talking about the 1986 to '96 years. I was actually living in the middle of that super cycle, literally the 10 years in the landscape was completely changed. Overnight. Proprietary network operating systems running on the big mainframe and many computers to open systems. Now we're seeing an open systems revolution again combined with large scale supercomputing power being democratized. So you're seeing every theater of technologies exploding with innovation, infrastructure, middleware, just generically saying, middleware is just software. And then obviously the applications are being infused with data assistance, agentic systems coming around the corner. I mean, this is a seismic event in the industry. It's a supercycle.

Pete DeJoy

>> It certainly is-

>> Where are we in this right now? Where's the progress bar? Where are we on this progress?

Pete DeJoy

>> Yeah, no, it is a really great question, John. Look, we're certainly feeling the tailwinds of this movement in our business and we're very grateful for that and also very optimistic about our future. I think one of our key observations is that folks have really been doing machine learning and AI for a very long time, but until the admin of these foundational models, you had to sink a pretty big CapEx investment into actually making AI happen. You had to go hire a bunch of very specialized resources. You had to spend a lot of money on GPUs to go train these models that your specialized AI researchers are going to write, and you also have to find a way to deliver them trustworthy clean data. Now with the advent of large language models, the barrier to entry to doing AI has decreased considerably. Now you don't need to go hire specialized resources and you don't need to spend money on GPUs to train these models because you can just access them via an API and pull them off the shelf, these pre-written pre-trained models. However, we are still seeing a lot of challenges in actually going to production even with generative AI. And even Gartner is predicting that 30% of these Gen AI projects are going to fail by the end of next year. So one of the key questions on our mind is why is that? Why are these projects failing? And our perspective is that the challenge of actually delivering clean, trustworthy data to fine-tune these models is the gap between prototype and production. It's now so easy to prototype of large language models. I've even seen interns build really revolutionary projects that make my jaw drop, but actually making sure that those models are fine-tuned with domain-specific data that is trustworthy and reliable is the thing that's preventing folks from actually going and driving highly strategic business outcomes with AI. And our perspective is that that's just a data engineering problem. It's the same data engineering problem folks have been working to solve for a very long time. Delivering reliable, clean, trustworthy data requires a bunch of facilities that are just challenging to implement and that's where we play at the end of the day. We want to make it easier.

>> Yeah, Pete, you bring up a good point. I mean, I do like this idea of prototyping because it's easy to build a rag system when you have a small data set.

Pete DeJoy

>> That's right.

>> Okay. Once you start connecting either larger-scale data or trying to integrate in intelligently with other data, you start to get into a little bit of a messy situation. And you guys again have experience with the large-scale data sets. This becomes the crossover from the scale question to real where AI needs to be successful. I would also say from our observations on theCUBE is that the boardrooms are saying, "We need an AI strategy." And then the bottom's up works like, "Wait a minute, our platform's not really ready. We're just getting Kubernetes automated." So you're seeing a lot of that in the plumbing and also there's a lot of engineering going on at the processor level. So you have a lot of infrastructure stuff that will flip soon to be completely stable. But now the app side is like, I got to play with the data. So we're in this experimentation mode. So take me through how you guys are seeing your customers now because training's obviously hot because that's one thing, but inference will be the killer app because that's when you graduate school. When you're trained, you're in school basically. And then inference is like, okay, I'm using data. Now in connecting to intentional relationships or delegating authority to another data set, you're in a different service level than just say microservice, just say, hey, I'm going to do, Andy, a data exchange with you. But it's not an API call. You're connected to me.

Andy Byron

>> That's right.

>> The data has to connect and be accurate, vetted.

Andy Byron

>> That's right.

>> That's not easy.

Andy Byron

>> So I think when you go through these big use cases, John, and I think actually to bring it even back to what you mentioned before, the transformation that's happening now is happening in the course of months, not years. And what we're seeing is that the constant foundation of that is people need data. And what we do is just deliver data at scale to a modern stack. So that connection for us is super important. Our ability to make that connection, which is what we do, is really important. But then also the ability to say, okay, once that connection's made, we can provide you observability into that connection. We can provide you quality of service into that connection. So the data that's being delivered between you and I as an example, we know that that quality of service is going to be really high. We can assure that the security around that data, which is on everybody's mind today, is very high as well. And then the third thing is around cost management. We hear that all the time, which is when that connection's made between you and I and then frankly hundreds of thousands of other connections, the cost associated with that, which is on everybody's mind right now has to be low. And we provide that ability to keep the cost low as well.

>> You brought up cost. That's a good point. I want to get your both thoughts on this next section of conversation around productivity. Obviously AI is showing it's not an either or in inspection. Go big or go home. There's a nice middle ground of productivity, low-hanging fruit you can knock down on an enterprise when you get through the prototype, you go, "Okay, let's pick up a use case." Productivity gains at the expense of high CPU costs. So this is where you start to get into that, the lines cross. You got to have your innovation value proposition yielding enough value that the cost envelope is within that reason. A lot of people I talked to are like, "Look, I can throw GPUs at a problem and generate some productivity." But we didn't win back that piece.

Pete DeJoy

>> Yeah. I think what we're finding, John, is that productivity and dollars aren't actually the limiting factor in driving strategic outcomes. You mentioned earlier a lot of the discussion in boardrooms is finding a way to infuse AI into all of these core businesses. I think every single company in the world, and we actually see this in the field from supply chain to Ford to some of the most advanced e-commerce businesses in the world. All of them have ways to leverage AI in a way that's strategic, but at the end of the day, the outcomes that you deliver are actually not going to be constrained by how fast people can prototype because prototyping has become a commodity at the end of the day. It's very easy for someone fresh out of college with a CS degree to go build something fairly robust in a little rag application. Now, the truly differentiated outcomes are going to come when people are solving problems that are high context, very important and strategic to the business outcome itself, and are actually domain specific. So you have some level of domain specific data that's differentiated that allows you to build in IP mode around what you're actually doing at the end of the day. And we see this in our customer base right now, it's pretty remarkable what some folks are doing. I mentioned Ford. Ford is using our product and orchestration generally to train a lot of their advanced driver assistance systems models. So we're talking about terabytes of data every single day coming from radar and onsite cameras on these vehicles, coming back to the data platform team to then go retrain these autonomous driving models so that they can actually go be competitive with some of the more modern players. In this case Ford has obviously been around for a very long time.

>> What's your profile customer size wise, what's the engagement look like? Take us through a day in the life of Astronomer or when you guys went to customers. Is it big data sets? Is it workflows? Is it which asset class? I mean workflows are assets now, not just data. Data and workflows are assets. Scope where you guys are winning, where most of your engagements are.

Andy Byron

>> Yeah, I think first of all, any company that has a large proliferation of data we can have a great conversation with. And that data can be used for really a couple different things. And we're starting to see the evolution of the data being used for what we are starting to term as a data product, which is a byproduct in some cases of the AI stack. And I'll give you an example. So in one of the largest retailers in the world as an example, they're using their data to make better decisions around their support cases. They're also using their data to make better decisions on how to route their supply chain to get people packages in a much more faster, consistent way. In another case, there's a large technology company that uses data to make more decisions around their product, who to sell to, and frankly to have a better experience around their products. So for us, that's how we think of data and every company in the world's trying to do that today. So we're not tethered to one vertical or one market. I think anybody that wants to get more value out of data, that's a great company for us to speak to.

>> Enterprise AI is such a untapped market because the workflows are domain specific. You mentioned that earlier. You can't just throw AI at the enterprise because there's a lot of knobs and buttons that are pushed all over the place.

Andy Byron

>> That's right.

>> Using that as a metaphor, but identity access. Okay, I'm going to have an edge use case that has to tie into a core cloud. I got to harmonize that data set and have compute on the device. I mean, an architectural shift is happening at the same time as everyone's rolling in these data value oriented workflows.

Andy Byron

>> That's right.

>> How do you guys engage with customers there? Is it a mix of service and platform? Because if you look at agentic systems, that's going to change the professional services business and move that into software economics.

Andy Byron

>> Yeah.

>> So that's a good scenario for platform companies.

Andy Byron

>> So maybe Pete, I'll hand it over to you in a minute, but-

Pete DeJoy

>> Sure, yeah.

Andy Byron

>> I think the other piece that's happening too, John, is we're seeing more companies that are building out a modern stack to go and implement AI, and then there's in many cases a legacy stack too, and they're trying to figure out what to do with that stack. Do they bring that data over to the modern stack? Do they keep that data where it is? Do they keep the modern stack? And a lot of it is around productivity and cost efficiency there. So we help on both sides. Our business value is we can help companies deliver on a modern stack and then we can help companies over time move more of these legacy data workloads onto the modern stack to make their company more productive.

>> And modern stack, you don't mean rip and replace.

Andy Byron

>> No.

>> Add a new stack into the new elements, into the mix and integrate.

Andy Byron

>> That's right.

>> And then if it goes away, I think it goes away. If the old goes away, then you don't need to kill the old to bring in the new. Yeah. Yeah.

Andy Byron

>> And specifically in that legacy stack, there's a lot, I mean, that's a highly valuable stack for many, many companies.

>> AI is an abstraction opportunity to pave over legacy, but the future isn't paving the cow path.

Andy Byron

>> No.

>> Okay. It's to build the new thing.

Andy Byron

>> That's right.

Pete DeJoy

>> I think the caveat there, which I think is you make a really good point, John, is a lot of these modern AI workflows are just extensions of the data work that people have been doing in the modern data stack for the last decade or decade plus in the advent of the cloud. A lot of the AI problems are really just data engineering problems at the end of the day. You mentioned earlier some of the maybe enterprise red tape or guardrails that are common when you go actually attach to one of these AI initiatives that's being rolled out, you have to have good governance. There are a lot of knots and swishes, good access controls, good security. All of the things that come into actually adopting software and production context in the enterprise are very true for this AI category of workload. Now, those things share a lot of commonality with any sort of data initiative. So the foundation that we're building and what we're focused on with respect to both migrating away from legacy systems and helping with services is building upon that strong foundation so that folks are actually given the governance abstractions that they need, the observability abstractions they need and the security abstractions they need to go actually drive these projects forward and not be limited by their internal revenue.

>> That's a great point. In fact, when I had the interview with you guys at re:Invent last year with theCUBE in the press area, I was energized by the conversation because when you see the wow, this thing, these guys got it right, the data ops looks a lot like the old security shift left conversation.

Andy Byron

>> That's right.

>> You have two things going on right now. You have the operating player saying, look, we have to get a stable environment to your point and get it set up, and then have the practitioners in the company who are the real domain experts become the value brokers because the value will come from the people who know the systems. And so I call that The New IT, because they have to set the table for the trust and the delegation of the agent connection because automating things for the sake of automating doesn't really do anything. I mean, you can automate bad process, but if you don't delegate trust and have that workflow nailed on the domain specificity, you're going to be hosed basically.

Pete DeJoy

>> That's right. And that's what you're describing is exactly why the platform is so important because if you deliver or if you push value down to those brokers, at the end of the day, they need to be enabled with the technology and infrastructure needed to actually drive outcomes. There are a lot of people that can go work with these large language models, but can you work with them in context inside the four walls of your organization? And from our perspective, that ties to your point about data ops, right? The key to actually going and moving the ball forward with real production, real world AI use cases is a really strong data ops foundation.

>> It's interesting, the data science and the analytics team have been doing business analytics for over a decade. I remember when we started theCUBE 15 years ago in New York, we'd have huge Hadoop events. We all know that was hard to implement. Spark came in, then data lake emerged. Okay. But so, progression. These data scientists doesn't mean they're mutually exclusive against the data engineering that's being done right now. So platform engineering and cloud native to me is a precursor to what will be data engineering, which is happening right now, and they work with that next layer, which is the data science business people, I call the business people, and that's where I think see that coming together because the agentic has to have the guardrails because you can't have horizontally scalable data if you don't govern it because low latency data access is the key to generative AI because it's not a pre-programmed thing. You don't know the prompt. Things are going to happen in real time and iterate and self-learn, the reasoning turns into reinforced learning, self-healing, when that hits, you can't have a centralized repo.

Andy Byron

>> No.

>> Now you can centralize things, but you can co-locate it horizontally so it can be domain-specific. Those, the two things have to be going on at the same time.

Andy Byron

>> They do. And that's actually, John, how Pete and I on the way in here, we were talking about this specific problem, ironically. Look the way we engage customers, we're going through this whole stack with the whole transformation happening and everybody's turning to us in particular around just data movement, data orchestration on what do we do. And I think our engagement model as an example with Northern Trust, which is a great customer of ours, they're standing up a modern technology stack and like we mentioned before, moving some of these legacy products onto their modern technology stack. It's not a lift and shift, but it is over time. And we are the ones that are set up really well right now to go help companies with these conversations. We have the knowledge and expertise around how to hydrate these AI models securely and at scale and cost-effectively.

>> Explain what hydrating the AI models means. A lot of people might not know what that means.

Andy Byron

>> What that means specifically, it's a really simple concept, John. It's supplying data through a pipeline to an AI workload. It's making sure that that AI workload has the right data to tune LLMs and do it securely in that scale. So we're the ones pumping that data into these workloads.

>> You got a good connection. The water's fresh.

Pete DeJoy

>> That's right. You're right.

Andy Byron

>> Well, it goes back to what we were talking. It's secure and that fresh water's going to make sure that that AI model grows and grows and grows. So we have that on one part of a company and then on another part of the company, it's like how do we take the existing data that we have and potentially leverage that into these modern models? And so in the case of Northern Trust as an example, we went and prescribed exactly how to do that and we had a really successful migration for that company in particular. The other thing we look at as well around just customer success and making sure that our customers are getting value out of what we do is just the adoption rate of our product. It's consistently over 90%, which is best in class and in software.

>> Give some customer use cases. And I know you guys got a lot of big customers, but you really can't say their names, but for the ones you can say their names, say their names. I mean people want to see use cases because it's very much a hands-on all around the companies we talk to, everyone's got their hands in. I mean, how could you not get excited about generative AI? It's definitely game changing, no doubt about it. But what is the use cases that you guys have been successful at? Can you share a few?

Pete DeJoy

>> Yeah, I can maybe start. Andy and I spent a ton of time with our biggest strategic customers on how they're implementing our product and really try to glean as much as we can about what they're doing. At the end of the day, what we find is people start with data engineering when they work with us and work with orchestration. But data engineering is just, or excuse me, AI engineering is just data engineering at the end of the day. And without that strong data engineering foundation, hydrating these models and actually delivering them clean, trustworthy data that you can tune, and actually use, to drive domain specific outcomes is impossible. So what often happens is somebody will adopt us for a data engineering use case, and very quickly we will be in the weeds of AI discussions. The largest travel website in the world is a customer of ours or one of the largest travel websites in the world. And we started working with them on a lot of their internal reporting pipelines at the end of the day. But very quickly the conversation changed to going and fine-tuning an off the shelf, open source large language model to serve an agentic workflow inside the application. So now if you're going and booking a trip on their application, you can just talk to the bot and say, hey, I want to go to Sicily. Tell me what the best time to go is, the best place to stay. And it'll just take care of it for you. Another really exciting one that we're energized by is a lesser known company, but one that's doing really important work called Vibrant Planet. What they're doing is they're actually ingesting a massive amount of weather and climate data on a daily basis and training their own traditional machine learning models to predict where wildfires are likely to happen for local fire departments across the western continental United States. So what they're doing at the end of the day is using machine learning and using this next generation technology to actually do something that's very good for the world.

>> What's the name of that company?

Pete DeJoy

>> Vibrant Planet.

>> They should have been on when we were here for climate week.

Pete DeJoy

>> Yeah, no-

>> Next time for sure.

Pete DeJoy

>> It's incredible stuff.

>> Well that's a game changer.

Pete DeJoy

>> And again, this is fairly ubiquitous across our customer base. People very quickly evolve from data engineering to AI engineering. And it's pretty energizing-

>> That's interesting progression, Pete, because one of the things I observe is I met with this big group, I won't say their name because I don't have permission, but they do a lot of business data and B2B data. We were having a conversation like use AI to first figure out what you have because their aperture for services has grown because they have basically everything about businesses out there. Their products are changing because of what AI is doing. So the first wasn't, I mean, they had the workflow stuff, that was low hanging fruit, they got some nice workflows, but their big aha moment was we got more value in the data here than we ever thought. And that opens up again, to your point, you're in the weeds. Then you go, okay, but look at this. You're smiling because that's probably happening, right?

Pete DeJoy

>> Oh, a hundred percent.

Andy Byron

>> I mean it happens in our company. That's how we make a lot of product decisions.

>> This is like generational real-time value. This is proof points that the gen AI is going to unlock new value that's never been seen before because you got faster compute. You have a agentic coming around the corner, but the precursor is just pipelining and doing plumbing.

Andy Byron

>> That's right.

Pete DeJoy

>> That's a hundred percent right, John. And I think that's what gets us super excited about the future specifically for our company. But at the end of the day, there are going to be world-changing outcomes unlocked by AI and also of course specifically generative AI. There's no question about that. But the real gap today is a data engineering problem. And that's what gets us really excited about the future because we know how to solve those problems, and we're-

>> Is that a shift right or a shift left conversation?

Pete DeJoy

>> It's a good question.

>> Because I think that in the CI/CD pipeline, developers will be dealing with data decisions with either voice activation or coding in line. Like security. When you start to get into these clean pipelines and you have these guardrails with governance, I think data will be free to fly all over the place and then manage from governance central co-located intelligent OS. You can provide developers real-time coding or enabling an agent to either check another sub-agent to make calls on the fly and change the outcome for the user experience.

Andy Byron

>> That's-

Pete DeJoy

>> We see it going in both directions, shift right and shift left, because we're pulling it into the developer workflow itself and making it much easier to deploy these things. But also what then becomes important as data explodes is the unit economics of this data, right? We talked earlier about cost of CPU and productivity. That's nothing compared to the cost of these exploding data platforms at the end of the day. So as people really invest in these data platforms to drive these differentiated outcomes, understanding the unit economics of all of your data infrastructure is super important.

>> Well guys, it's been great to have you on theCUBE again here in our new location. We'll maybe see you at re:Invent. Certainly let's keep in touch. Final word, let's give a quick plug for Astronomer. What are you you guys looking to do? You're hiring, you're building out. What should customers, why should they call you? Is the house on fire? What sign would tell a customer to call you guys? Give us the plug.

Andy Byron

>> So first, we are hiring always. We're always looking for the best talent. Second is anybody that wants to make use of their data as a competitive advantage or anything associated to making their company better, call us.

>> Guys, power in the next generation. I'm John Furrier here in our studio in the NYSE Wired community. We're building an open network, Silicon Valley connecting into Wall Street with NYSE, Brian Baumann and the team. theCUBE will have a set here. We'll continue to program as much as we can at any time all day. If you want to come on, give me a ping on DM and we will connect. Thanks for watching.