Kubernetes CPU and memory requests are usually set once, from a guess, before a workload sees real load. Set them too high and the cloud bill keeps climbing. Set them too low and services hit out-of-memory errors or get throttled. The production data to fix this already exists, but nobody has time to tune every workload by hand, so infrastructure management stays reactive.
In this interview on TFiR, Shahar Azulay, CEO and co-founder at groundcover, walks through why Kubernetes requests stay a guess, how autonomous infrastructure and the Wand acquisition change that, and what AI observability and autonomous agents need from telemetry.
Guest: Shahar Azulay, CEO and co-founder at groundcover
Show: KubeStruck
LinkedIn: Shahar Azulay
Watch the discussion on YouTube
Here is what every platform engineer and SRE needs to know.
Technical Deep Dive
Q: How do you stop guessing Kubernetes resource requests and cut over-provisioned spend?
Shahar Azulay, CEO and co-founder of groundcover, says teams sometimes see up to 80 or 90% excessive spend on how they provision cloud environments. The fix starts with deep infrastructure monitoring that learns, per workload, how many resources it needs over time compared with what was provisioned. With the Wand acquisition, groundcover Autonomous Infrastructure realigns provisioning and requests dynamically across a wide Kubernetes environment to reduce spend and improve resiliency, work that otherwise takes endless manual effort.
“We see up to sometimes 80 or 90% excessive spend on how people provision their cloud environments.”Shahar Azulay, CEO and co-founder, groundcover
Q: Why is infrastructure management still reactive when teams already have so much production data?
Azulay says cloud brought innovation and abstraction, which in turn brought complexity and obscurity, so telemetry keeps doubling every year. In the current model, observability platforms effectively rent customers their own data back, so customers drop, sample and turn off features while value stagnates. AI accelerates the problem, because monitoring agents in production need even more telemetry, and coding agents now rely on that telemetry too.
“Telemetry is rising, value is stagnating because customers are forced to drop, sample, turn off features, pay more again, and go into that vicious cycle and not get more value.”Shahar Azulay, CEO and co-founder, groundcover
Q: What is autonomous infrastructure and why did groundcover acquire Wand?
Azulay says groundcover made its Autonomous Infrastructure generally available at KubeCon, as a result of acquiring Wand, a company from the Kubernetes optimization domain. It is part of a bigger vision in which observability becomes the grid on which autonomous operation is built. The starting point is Kubernetes optimization for cost and resiliency, and the next step is having an agent act in production when an incident happens.
“We’re moving from being a passive observability platform into a platform that can take action in production in the infrastructure itself.”Shahar Azulay, CEO and co-founder, groundcover
Q: What telemetry do autonomous agents need to optimize code and infrastructure?
Azulay says the telemetry has to span the verticals of observability: logs, APM, infrastructure monitoring and GPU monitoring, tied together with algorithms that can make decisions. Most customers do not use half of these features, and they turn them off or heavily sample them because of how current solutions are architected. That leaves most teams without the stored telemetry that agents need.
“Most of them don’t have the telemetry in place stored because agents require more raw and abundant data.”Shahar Azulay, CEO and co-founder, groundcover
Q: Do AI workloads need a separate monitoring tool, or full-stack observability?
Azulay argues AI workloads do not run in a siloed environment. They run alongside legacy microservices, managed cloud services, and the CPU, memory and log management that teams already monitor. Tools such as LangFuse emerged for AI monitoring, but he expects AI observability in separate tools to merge into platforms that have all the pillars of observability.
“If you only see AI observability and don’t know the context of the Kubernetes infrastructure it’s running on, you’re not going to get the full picture.”Shahar Azulay, CEO and co-founder, groundcover
Q: How do you monitor AI agents for cost, governance and hallucination?
Azulay says the basics are governance questions: which models are used, which tools are called, and where the budget is spent. groundcover’s eBPF sensor needs no instrumentation or code changes, so it can see AI interactions. It stores traces showing prompt use and tool calls in a bring-your-own-cloud backend more cost-effectively, which lets evals judge over a 2-hour session whether behavior was a hallucination or desired.
“It’s a superpower to see AI adoption inside your production you haven’t even known about.”Shahar Azulay, CEO and co-founder, groundcover
Q: Can you trust autonomous agents with production systems?
Azulay says groundcover dogfoods what it pitches: it uses agents heavily when coding, with its own telemetry, and is moving toward taking the human out of the loop in testing and verification. The goal is a fully autonomous CI/CD cycle, and he expects enterprises to get there in 18 to 24 months. Guardrails widen over time as AI proves its value, and the one thing teams can control is the quality of the data agents work on.
“The reality is that the only thing you can control is making sure that the value and depth and width of the data they work and operate on top of it is the best that you can get.”Shahar Azulay, CEO and co-founder, groundcover
Q: Can observability agents fix incidents instead of only detecting them?
Azulay expects observability agents to delegate tasks to coding agents within a few months, for example asking them to instrument a new service so it can be seen. For incidents, an agent that has reached the root cause could hand the hotfix to a coding agent. For infrastructure, it could change the machine type, drain the node or reconfigure the deployment, with the right context and guardrails. Trust will be built over time, and the types of problems handled autonomously will grow.
“I’m seeing a problem, I already reached the root cause, under some guardrails. Coding agents, please code the hotfix and push it to production right now. Why wait?”Shahar Azulay, CEO and co-founder, groundcover
Q: How is AI helping newer observability vendors close the gap with Datadog, New Relic and Dynatrace?
Azulay says groundcover is always in a parity game against Datadog, New Relic and Dynatrace, which are 15- to 20-year-old platforms. Coding agents cannot yet rebuild such a platform from scratch. They do help cover well-defined gaps, such as the database monitoring groundcover just released, with features that once took 6 to 9 months now taking 4 weeks.
“You can’t yet build an observability platform like Datadog from scratch using coding agents.”Shahar Azulay, CEO and co-founder, groundcover
Q: What is next for groundcover after its first acquisition?
Azulay says Wand is groundcover’s first acquisition and part of the strategy behind its recently raised Series C of $100 million. The company is growing its go-to-market, R&D and product teams. He expects more strategic and acquihire-style acquisitions, including around AI and the database layer, in the next few months and through 2027.
“Part of it is going to be strategic acquisitions that will strengthen our ability and widen the platform.”Shahar Azulay, CEO and co-founder, groundcover
Q: Why does groundcover say observability is reborn with AI?
Azulay says groundcover has always used fun, light themes to change how the observability domain is perceived. At KubeCon the booth celebrated the idea that AI is changing observability so much that it is becoming the grid for AutonomousOps.
“We’re celebrating the fact that observability is reborn with AI. It’s changing so much. It’s becoming the grid on top of which AutonomousOps is being built.”Shahar Azulay, CEO and co-founder, groundcover
Resources & Documentation
- groundcover Autonomous Infrastructure, groundcover’s Kubernetes optimization product for cost and resiliency
- Agent Mode, a groundcover product page
- Dashboards, a groundcover product page
- gcQL, a groundcover product page
***
👇 Click to Read Full Raw Transcript
Swapnil Bhartiya: Most platform teams run their production workloads on Kubernetes and they like the scale and elasticity that it offers them. But CPU and memory requests for those workloads are usually set once from a guess before the workloads actually see real load. Too high and the cloud bill keeps climbing. Too low and the services hit out-of-memory errors or get throttled. The production data to fix this already exists, but nobody has time to tune all those workloads by hand. So infrastructure management stays reactive and enterprises are looking for answers. Groundcover has been working on this problem, and to discuss autonomous infrastructure and what it means for engineering teams, we sat down with Shahar Azulay, CEO and co-founder of GroundCover at KubeCon and CloudNativeCon in Salt Lake City. Let’s go and talk to Shahar. Shahar, first of all, it’s great to have you on the show.
Shahar Azulay: Thanks for having me. Great to be here again.
Swapnil Bhartiya: I must say that at this KubeCon, your booth here is, I think, one of the cherry on the cake, you know, is popping out and no pun intended today. And also you folks made some acquisitions. So there’s so much to talk about, but I want to start from your presence at the show. You’re looking very colorful this time. Talk about the theme, the idea, and your presence.
Shahar Azulay: I would want to say I wear it every day, but no, it’s mostly for the event. Basically, GroundCover has always been about fun and light themes to kind of maybe change the observability domain from being mature and all to something new. But we’re celebrating the fact that observability is reborn with AI. It’s changing so much. It’s becoming the grid on top of which AutonomousOps is being built. So it’s a birthday party and check out the booth later.
Swapnil Bhartiya: And you are celebrating with the whole community here.
Shahar Azulay: We’re celebrating with the whole company here.
Swapnil Bhartiya: We are here at the event. Talk a bit about any major announcement that you folks made here.
Shahar Azulay: So Groundcover has been releasing a lot lately, but in KubeCon right now, we have GA’d our Autonomous Infrastructure, which basically is a result of an acquisition of a company called Wand. That is coming from the Kubernetes optimization domain. It’s part of a bigger vision that GroundCover is basically believing that observability is turning into the grid on top of which autonomous operation is being built. We’ve talked about it multiple times in the software side before. Telemetry from observability is being used by coding agents every time to test code and ship code better, and also to remediate and reach root cause analysis better. We’re taking the same approach for infrastructure. So we’re moving from being a passive observability platform into a platform that can take action in production in the infrastructure itself. The starting point is Kubernetes optimization for cost and resiliency. The next step is actually taking different action when there’s an incident in production and you have to have an agent act on your behalf in production and remediate production. So it’s kind of a— Again, a concept of a bigger theme.
Swapnil Bhartiya: I want to just go back to the very foundational level. And if you look at, you know, infrastructure management, it is still very, very reactive. You touched upon that as well. You know, it is actually even more interesting because teams already have so much production data. So why is the reason that it’s still reactive? And then we’ll talk about how is AI changing it? Because now when we talk to everybody’s talking about observability, as you rightly mentioned, but let’s talk about the basic foundational
Shahar Azulay: The basic financial problem has basically been the same for years. And the thesis on which Groundcover was built was that telemetry was exploding and doubling every year because of cloud. Cloud basically introduced beautiful things into our lives. But one of the things that it— the 2 of the things that it introduced is innovation and a lot of abstraction, right? That was cloud brings in. They in turn introduced complexity and obscurity. Yeah. Right? So the amount of telemetry you need to monitor a cloud-native environment is just higher than before and rising and doubling every year with the scale of deployments in cloud. In the current way that observability platform operates, they’re basically renting our own data back. I’m paying for data volumes stored in the vendor. And that’s a major gap because what’s happening right now is that telemetry is rising, value is stagnating because customers are forced to drop, sample, turn off features, pay more again, and kind of go into that vicious cycle and not get more value. With AI, I think that the problem is just accelerating. AI is kind of pouring gasoline right up above it because, A, we need a new offering called AI observability. I want to monitor agents in productions. It requires even more telemetry because they’re different and hard to monitor. And the second part is that I now have new use cases like coding agents relying on that telemetry to code better. 100% of the customer base in groundcover are hooking their coding agents, Codex, Cloud Code, and whatever in the development stage, not in the production stage, to our telemetry. So they can get context from production as they code and build. Without that, you get a gap between you and your competitor, right? They’re building faster than you. They’re building better than you because of better telemetry. So I think same gap, not in the production stage, to our telemetry. So they can get context from production as they code and build. Without that, you get a gap between you and your competitor, right? They’re building faster than you. They’re building better than you because of better telemetry. So I think same gap, just bigger problems are caused due to that.
Swapnil Bhartiya: When we look at AI agents, I think there are 3 major problem areas that are emerging. One is, of course, cost, token, and that’s why the whole tokenomics is there. Second is the whole governance. And of course, the third is, I mean, you can use the term hallucination also, side quest also. Yeah. Now, you won’t even know whether there are any without observability. So how is GroundCover looking at These areas, or what are the— when it comes to AI agentic workflow, what are the problem areas GroundCover is focusing on?
Shahar Azulay: So there’s a new emerging field called AI observability, which we’re heavily in. And I think that GroundCover is leading that domain from a perspective of how far we take it. Basically, the basic questions are like you said, right? People don’t have the basic governance of understanding which models are being used, which tools are being called, Where is the main budget being spent and how is AI even being used inside the platform? 2 main factors that we always do in different things that we do impact why we could do it better. One is the eBPF sensor. We have a sensor that doesn’t rely on instrumentation or code changes. So we can see AI interactions without developers running after the tails and instrumenting code. And the fast that AI is going through and the amount of AI adoption happening right now, It’s a superpower to see AI adoption inside your production you haven’t even known about, right? And the second is that we can store all this telemetry, including traces from AI showing prompt uses and tool calls in the bring-your-own-cloud backend that Groundcover has more cost-effectively. So then we can layer on things like evals that can basically let AIs adjust, determine over a 2-hour session, was, like you said, was it an hallucination or an actual standard behavior that you wanted or a desired behavior? So we have the ability to store more data correlated to everything else we do, from infrastructure monitoring to regular APM, and then allow AI to decide if it made sense. So I think the foundations GroundCover has built help us to create a very interesting AI observability offering.
Swapnil Bhartiya: And when we look at GroundCover today, and when you look at when it came to existence, the focus of cloud-native Kubernetes, are you still looking— I mean, of course, AI is running on top of Kubernetes is running on top of a Linux kernel. So when we look at groundcover today, are you focusing, purely focused on AI workloads or you’re still focusing on the— I mean, I will not use the word traditional for Kubernetes, but if I ask you, where is the focus?
Shahar Azulay: I think that our belief is that AI workloads are not running in a siloed environment, right? We saw it in the past, I think, when a part of the market is running so fast, Then what happens is that tools are being developed for that specifically, right? Tools like LangFuse and other tools that are basically saying, okay, there’s a new player in town called AI. Let’s build an AI monitoring tool for that. But the reality is that AI is running as part of your stack, right? With legacy microservices, as you say, right? And usage of managed services in the cloud and the need to monitor infrastructure and CPU usage and memory usage and log management, right? So Basically, we see an observability platform, a mature observability platform, being able to build all these layers and then reach AI observability. So we’re definitely focused on the entire stack, and we think that’s our superpower because if you only see AI observability and don’t know the context of the Kubernetes infrastructure it’s running on, you’re not going to get the full picture. It’s not running in void, right? It’s running as part of your stack. Your other services are using it. So GroundCover is trying to create that whole experience, and we also believe that AI observability in those separate tools is merging into tools like GroundCover and even Datadog, of course, right? These type of solutions that have all the pillars of observability can create a better understanding of how you use AI. So that’s kind of how we see the world. So it’s not a separate world for us.
Swapnil Bhartiya: And when we look at from where you sit, you know, when we look at cost, for a lot of vendors, they focus more on the token cost and all those, but you’re also looking at cloud cost. Because, you know, they just said, you know, CPU memory request for Kubernetes workload. So you’re helping organizations in mitigating these 2 cost areas. Yes. How do you enable teams to also manage that? Now we don’t talk about cloud cost at all. Now every discussion has shifted to AI these days. We are not— the discussion we had with you last year or before that was totally different than we are. But as you said, fundamental technology is still the same. So basically
Shahar Azulay: we see up to sometimes 80 or 90% kind of excessive spend on how people provision their cloud environments. It starts by a very deep infrastructure monitoring layer that understands per workload, you know, on a very wide Kubernetes environment, how much resources it actually needs over time and the trends it actually consumes when users actually use it compared to what you provision for it. So GroundCover knows right now with the Wand acquisition and our new feature we’re GAing today of Autonomous Infrastructure, how much resources it actually needs over time and the trends it actually consumes when users actually use it compared to what you provision for it. So GroundCover knows right now with the Wand acquisition and our new feature we’re GAing today of Autonomous Infrastructure, we know how to realign dynamically all of the provisioning and requests of the different workloads across a very wide Kubernetes environment to reduce spend and also improve resiliency. So with the eBPF sensor and with all the data we already have, with a click of a button, customers right now can enjoy a new SKU inside the groundcover offering that allows them to basically optimize their spend and resiliency in Kubernetes. In a very large environment, that can sum up to a lot of spend and a lot of money being spent by these companies. That otherwise would require endless manual work that is really hard to achieve eventually, right? For us, it’s the first stage toward even deeper configuration of allowing an agent to reconfigure deployments in a deeper way and controlling incidents. But the basic layer is I’m spending too much and my system is not reliable due to a lot of kind of busy neighbors and interdependency between things. I’m deploying on Kubernetes. That’s what groundcover is GAing today.
Swapnil Bhartiya: Perfect. And since you mentioned, of course, autonomous infrastructure, which runs on, of course, as you once again mentioned, eBPF sensor with no sampling, what kind of telemetry it needs and what does it mean for the future of how observability platforms are built? Perfect. And since you mentioned, of course, autonomous infrastructure, which runs on, of course, as you once again mentioned, eBPF sensor with no sampling, what kind of telemetry it needs and what does it mean for the future of how observability platforms are built? Not only because the workloads are changing, so it has to evolve.
Shahar Azulay: The telemetry has to be built basically from the entire different— these are verticals of observability, right? Log management from the application layer is still important. APM is still important to understand API and cross-network interactions. Infrastructure monitoring is still important, right? GPU monitoring is still important, but tying it all together to algorithms on top of it that can make decisions is the next step. I think that part of it is allowing people first to store all of this telemetry. We have to remember That even though observability signals are pretty mature, right? You don’t have to explain to the average customer what is infrastructure monitoring or what is APM. Most of them don’t use half these features. They turn it off or heavily sample them because of the way these solutions are architected today. So even if, you know, from outer space, there’s a great new technology of AI that can optimize your, you know, code or your application or your infrastructure, Most of them don’t have the telemetry in place stored because agents require more raw and abundant data. So we feel that pull from the market right now.
Swapnil Bhartiya: How much AI agents is GroundCover actually using internally as well? to build the systems? And since some of the systems they’re working on, very critical, how do you also ensure that you can actually trust that? Because these agents are autonomous, they’re taking action on your behalf. They’re not just assessing changes. So they’re trustworthy and the system that they have access to, you’re fully confident that they will not break it or they will not hallucinate and go on a side quest as well?
Shahar Azulay: First of all, groundcover is dogfooding a lot of the things we pitch, right? Because these agents are autonomous, they’re taking action on your behalf. They’re not just suggesting changes. So they’re trustworthy and the system that they have access to, you’re fully confident that they will not break it or they will not hallucinate and go on a side quest as well. First of all, groundcover is dogfooding a lot of the things we pitch, right? I think of course we’re ahead of the curve. We use agents, Heavily when we code, of course, and we use the ground cover telemetry on ourselves as we code. So there’s a lot of autonomous code and code writing and testing right now happening in ground cover, and also moving towards getting the human out of the loop with a lot of agents we’re writing internally that basically dogfood the telemetry even in the testing and verification stage before you know code is being shipped. Eventually, we want to create a fully autonomous code and code writing and testing right now happening in ground cover, and also moving towards getting the human out of the loop with a lot of agents we’re writing internally that basically dogfoot the telemetry even in the testing and verification stage before you know code is being shipped. Eventually, we want to create a fully autonomous CI/CD cycle when when code is being written and pushed into production autonomously. Of course, we’re ahead of the curve, but the market is getting there 18 months from now, 24 months from now, the enterprise is going to get there. And the reason we believe they’re going to get there is because the difference between company A and company B right now with better telemetry is just that you are more reliable than me as a product. You have less downtime. But if you’re writing code better than me, if you’re shipping features to your customer faster than me, because your coding agents have better data and you’re writing it fast on a better harness, then you’re going to get better at me as a product and let them operate inside them, right? But as time passes, since AI brings so much value, these guardrails are widening. and we let them take more and more actions on our behalf. The reality is that the only thing you can control is making sure that the value and depth and kind of, you know, width of the data they work and operate on top of it is the best that you can get. Because the more data, the higher granularity, the deeper understanding of what they have on, for example, production, if they code, will make them more productive. And I think there’s a lot of, you know, doubt right now on the ability of a coding agent, given the right context, to code better than the average engineer. I think given the wrong context or a partial context, yes, we will see a lot of problems and a lot of gaps. That is exactly where observability is going to play such an important role in that, and that’s where we’re moving. So I think that’s the only thing you can control at the moment.
Swapnil Bhartiya: And you and I have had this discussion for a long time that, you know, with observability, it stops at telling you what is there, but with the autonomous agent, they can actually take action now. So talk a bit also about how you’re also seeing The evolution of observability, which maybe even you are leading, where it can do more than just identify a problem. And because of now AI agents, automated, it can actually go and fix the problem itself. Of course, it depends how much you trust it. Do you want to keep a human in the loop? But now we have that capability as well, which we always wanted.
Shahar Azulay: Right. I think in software, it’s already easier to understand, right? You will see in a few months from now, observability agents delegating tasks to coding agents and saying, you know what, this is a new service, it’s not instrumented, I can’t see it well as an observability agent, please instrument the code with telemetry so I can see it. That’s going to happen very soon. Same for incidents, right? I’m seeing a problem, I already reached the root cause, you know, under some guardrails. Coding agents, please code the hotfix and push it to production right now. Why wait? Right. If the observability agent together in collaboration with the human or the coding agent reached the problem. Same for infrastructure. If there’s a problem and I’ve reached a root cause analysis, I might be able to remediate the infrastructure as well. If I need to do something else, change the machine type, drain the node, reconfigure the deployment, I can take action before an engineer gets to that with the right context and the right guardrails. We’re going to get to that soon again. And that level of trust is going to be built over time, right? What type of problems we’re going to do from day one is not going to be the type of problems we’re going to end in we’re gonna handle in a year autonomously. But we already saw it with software development. It’s gonna get there really, really quickly.
Swapnil Bhartiya: How do you see, based on your own experience as well, where you’re seeing that AI is going to accelerate the work that you folks are doing, where you, in past, it would take like 2 or 3 years to achieve something for groundcover. Now you’re like, we are moving so fast, we are breaking things so fast. The things you wanted to do, not able to do it.
Shahar Azulay: I think one great example is that groundcover has been, and we will still be in the next few years, there’s always some parity game against Datadog, New Relic, Dynatrace. These are 15, 20-year-old platforms, amazing products, very wide, and kind of COVID everything in the book, right? So we will always be in some chase around another edge case feature that we don’t support today. We see the gap shrinking, right, with the ability Yes, it’s built on the right foundations, right? With the sensor and the backend, a lot of things we’ve been doing, it’s not magic, right? You can’t yet build an observability platform like Datadog from scratch using coding agents. But the ability to cover specific gaps like database monitoring that Groundcover has just released and features that will take us maybe 6 to 9 months in the past and we can do right now in 4 weeks, we can definitely see the gap closing fast on just good foundations and the ability to close gaps of things that are pretty well defined, like the parity race we have in the observability race. So we at GroundCover kind of enjoy that right now and kind of close the gap faster and faster against these incumbents, which had years to do it. So we definitely feel that firsthand.
Swapnil Bhartiya: I will go back to the first question. We talked about the acquisition. Is this your first acquisition or?
Shahar Azulay: Yeah, it’s our first acquisition.
Swapnil Bhartiya: So you have opened the door. What is next when we look at groundcover? If you can give some visibility into your own roadmap.
Shahar Azulay: First of all, I think it’s part of our strategy as in the Series C of $100 million we raised recently. Groundcover is growing. We’re growing our go-to-market team and also our R&D and product team. Part of it is going to be strategic acquisitions that will kind of strengthen our ability and widen the platform. We started with autonomous infrastructure as kind of expanding our ability to go proactive there. Groundcover has a lot of challenges where it comes to AI, the database layer of how we operate, where we would want to expand more teams in acquihire-style acquisitions, bringing in more talent and more knowledge into the team, in our R&D and product team. So definitely in our roadmap to do more acquisitions and go more into that direction. Part of the strategy right now that Groundcover will probably go through in 2027 and kind of the next few months.
Swapnil Bhartiya: Shahar, once again, thank you so much for joining me and congratulations on the acquisition. And also, as you rightly mentioned, that you are kind of ahead of the curve since I’ve been covering you folks for a long time. So I have been through that journey as well. So it’s great to see companies evolving. But I really appreciate your time today, and I would love to have you back on the show. Thank you.
Shahar Azulay: Thank you so much. Great to be here all the time.





