Companies are spending billions on GPUs, data centers, foundation models, and AI infrastructure. But what happens when the network connecting all of it cannot keep up?
In this episode of Tech Talks Daily, I welcome back Avi Freedman, co-founder and CEO of Kentik, five years after our previous conversation. Avi has been operating large-scale networks since the 1990s, including more than a decade at Akamai, and brings a rare combination of founder experience and hands-on knowledge of how the internet actually works.
We discuss why network performance is becoming an important factor in determining the return companies receive from their AI investments. If organizations cannot move data efficiently to models or deliver inference reliably to users and applications, expensive compute infrastructure can sit waiting while performance suffers and costs increase.
Avi explains what technology leaders should measure to determine whether their network is helping or hindering AI workloads. This includes establishing performance baselines, synthetic testing across cloud and AI providers, understanding dependencies across the digital supply chain, and using observability to identify what changed when performance deteriorates.
The conversation also examines network intelligence and why collecting telemetry alone is not enough. Organizations need to connect network data with the applications and users affected, understand historical behavior, determine which problems matter, and give network teams enough context to act quickly.
Agentic AI introduces another opportunity. Avi explains how AI agents can increasingly perform the work of experienced network engineers by monitoring baselines, investigating alerts, troubleshooting problems, and recommending actions. But fully autonomous networks remain some distance away. Most enterprises currently want humans deciding whether significant production changes should be made.
That leads us into governance. As businesses give AI systems access to increasingly important infrastructure, credentials, permissions, guardrails, and oversight become major considerations. Avi warns about ungoverned AI systems gaining proxy access to corporate infrastructure and explains why companies need clear boundaries around what agents can see and do.
We also revisit a lesson from decades of internet infrastructure: individual components will fail. Rather than attempting to create networks that never fail, businesses should design for resilience through redundancy, over-provisioning, monitoring, and architectures capable of continuing when something inevitably breaks.
For founders, CIOs, CTOs, network engineers, and infrastructure leaders building around AI, Avi offers practical advice on observability, network resilience, autonomous operations, AI infrastructure, and knowing when networking expertise should be developed internally or brought in from elsewhere.
And we finish somewhere unexpected: how CEOs can use AI to make better decisions by explicitly asking it to disagree with them. Avi explains why turning AI from a sycophantic assistant into an argumentative colleague can expose weaknesses in an idea, improve communication, and help leaders test their thinking.
AI may be transforming software, compute, and business operations, but none of it works without connectivity. As AI becomes part of the operational backbone of the enterprise, understanding the network underneath it becomes increasingly difficult to ignore.
Useful Links
[00:00:00] - [Speaker 0]
Your agentic AI might not be secure even with real time data and proper guardrails, but Denodo makes sure your business has every avenue covered. By placing all your data platforms under one AI data layer, your business can reach semantic consistency safely and securely. So get your agents on the same page by visiting denodo.com, And you can learn more about how to start trusting your agents to make business decisions. Welcome back to the Tech Talks Daily Podcast. Now five years ago, Harvey Friedman last joined me on this podcast.
[00:00:45] - [Speaker 0]
And the world was a very different place. AI wasn't dominating every technology conversation and nobody was worrying about agents running their networks. And it's not until I reflect on that conversation that I realized just how much has changed in the last five years. But one thing hasn't, and that is none of this technology works particularly well if the network underneath it cannot keep up. So this time, I've invited the CEO of Kentik to explain why network performance could become one of the biggest bottlenecks in the AI boom.
[00:01:23] - [Speaker 0]
And we'll also look at what agentic NetOps looks like, especially when AI starts thinking like an engineer. And why CEOs might get better results from AI if they stop asking it to agree with them and start asking it to argue back. It feels great to finally put the network at the forefront of a conversation and it's always a pleasure to get my guest back onto this show. This is why I ask you to buckle up and hold on tight because I'm beaming your ears all the way to Seattle so you can join me and Avi in conversation. So a massive warm welcome back to the show.
[00:02:01] - [Speaker 0]
I discovered a moment ago we spoke, I think, almost exactly five years ago, but for people that missed conversation, can you tell everyone listening a little about who you are and what you do?
[00:02:12] - [Speaker 1]
Sounds great. Thank you for having me back. I'm Avi Friedman, CEO of Kentic. We make the Internet go. We take telemetry from all of the modern infrastructure and help everyone make their networks fast and secure.
[00:02:26] - [Speaker 0]
And the last time we spoke, not too many people were talking about AI. We're probably just coming off the back of a global pandemic, and what working at home from scale almost happened overnight. And we've had hybrid working, then AI. And fast forward to present day, AI conversations are focusing on models, chips, and data centers, while the network, though, still receives far less attention. So one of the reasons I was excited to speak with you again today is talk about why the network performance is becoming one of the biggest factors determining whether AI delivers results or becomes increasingly expensive.
[00:03:04] - [Speaker 0]
And we've seen so many big talk stories around tokenomics right now. But tell me more about that and the importance of network performance right now.
[00:03:12] - [Speaker 1]
Sure. Well, people are spending, billions and, really trillions on the physical infrastructure. But if they can't move the data to the models and they can't get the models inferenced into the users, then all that CapEx, it's going to waste. And modern applications can't run, and people will drive off cliffs and enterprises will stop working. So really being able to see what's connected where, what their performance is, and how to optimize it is one of the biggest challenges that both our enterprise and our AI infrastructure customers have.
[00:03:49] - [Speaker 1]
Because we have a lot of customers in that space that are among those spending billions. And believe me, when you have an outage and your customer has a $800,000,000 contract with you, it's a serious issue.
[00:04:00] - [Speaker 0]
Yeah. It really is. And I'm so glad that we're getting to shine a light on this today because when a network underperforms, model training can stall, inference quality can suffer, and costs quickly mount up. But for people listening here, what what should them and their business, and technology leaders, what should they all be measuring to understand whether the network is helping or holding back their high investments? The old IT adage of you can only improve what you measure.
[00:04:28] - [Speaker 0]
And over the last few years, many organizations have almost lost their way when it comes to measurement. But what should they be measuring here?
[00:04:36] - [Speaker 1]
Well, it's an interconnected world. Most people are not running all their entire AI infrastructure in house. So it really involves getting a baseline of what your networks are doing and the performance using something called synthetics testing between your infrastructure, all of your digital supply chain, whether it's AI providers, cloud providers. And with that baseline, you can see, is everything good? Are there issues?
[00:05:01] - [Speaker 1]
When did that happen? And then observability ties together all the different signals when there's an issue, what could have caused that problem. And those teams, you know, again, sometimes augmented by AI, which we can talk about, are always watching and making sure you know, need need to make sure that those network operations are going well so that the enterprise can function.
[00:05:24] - [Speaker 0]
And looking online, I saw you talk about network intelligence as almost the answer to increasingly complex infrastructure out there. So what does network intelligence mean in practice, and how does better visibility translate into faster problem solving and ultimately, those better business outcomes that everyone's chasing now?
[00:05:45] - [Speaker 1]
At the core of intelligence is having the observability. So it's taking all of the telemetry. So what the traffic is on your network, the state of the devices, the routing information that connects everything together, everything from your network, the ones you depend on, your hosts, all the way to across the Internet, your cloud providers, keeping it in a system that allows you or your intelligent agents to be able to ask questions about it, and then putting into context, making it make sense. So it's not just what are the numbers coming from the network, but what applications and users are affected, And then putting into historical context to say, this matters, this doesn't, here's what you should do about it. So really elevating the basic observability to intelligent insights that people need to actually run their infrastructure in a interconnected complex world.
[00:06:38] - [Speaker 0]
And the talk at every tech conference this year is AgenTiKi. That's the conversation on the show floor in the keynotes. It's all we're hearing about, but AgenTiKi is also beginning to move into network operations. So what would an AI agent that genuinely thinks and acts like an experienced network engineer, what what's that look like in practice, and what kind of decisions would a business be comfortable allowing it to make today? There's a certain amount of trust involved there, but what what would you recommend that they they should be comfortable allowing?
[00:07:12] - [Speaker 1]
So let me split that into two parts. Yeah. We built our our agent controller system, what we call Kentic's AI Advisor. That is something that we've trained with the knowledge of networking, how people use Kentic to run their digital infrastructure today, and most especially, what is possible and what you should not do. In our first phases of it, it's really something that you summon to say, help me understand this, debug this alert.
[00:07:44] - [Speaker 1]
What should I do about this? And we are, through the year, making it increasingly proactive, but really as an augment to humans under the human control. So you could think of it as really a fleet of advisers that that you dispatch and control. So it can proactively be doing those things that I mentioned, which is watch the baselines, understand when something is diverging, do that triaging and troubleshooting, and then come to you and say, here's what I think is going on. Should I do something?
[00:08:16] - [Speaker 1]
And that is the state of the art. Today, very few workflows are agentically going complete closed loop. And I think this follows we're seeing this in our customers. Follow, you know, ten years ago when we launched DDoS detection. Our first for the first few years, everything would be tickets, someone human would watch.
[00:08:37] - [Speaker 1]
We would put what we call the big red button in it, which is basically you just click on it and it will do that action, and we're doing the same thing. So the adviser can make a recommendation. But today, most enterprises want humans in the loop. Sometimes even for reaching out to the infrastructure to do in the networking world, we call them show commands. Like, actually talk to the actual routing infrastructure.
[00:09:00] - [Speaker 1]
So no one's really ready for Skynet yet. It's all moving from human invoked agents to proactive agents to make recommendations. And I think we'll see use case by use case that opening up to more full agentic operations over the next few years. But I don't expect it to be automatically reconfigured the network, you know, in 2028. I think there's still gonna be guardrails and still comfort and making sure that humans are directing, validating, and in the loop.
[00:09:31] - [Speaker 0]
Whoo. A big sigh of relief if Skynet is not coming to fruition just yet. And, of course, autonomous operations, they promise to identify problems, determine root causes, and take action before users are affected, and any support calls have come in there. So what needs to be in place, though, before companies can safely move from just AI recommending actions to AI making changes in production networks? Anything else you can add to around that?
[00:10:00] - [Speaker 1]
Well, you have to have all the signals that are needed on a platform that humans or AI can ask the questions of. So a lot of the traditional observability systems, they do roll ups and throw the data away because they can't really handle the volume of not just the volume, but what's called the cardinality of network data. Meaning, number of unique values when you combine IP addresses and port numbers application IDs and all that. So you have to have the system so that your agents can go ask the questions to make the recommendations. And you need just need to have the guardrails to say these are the kinds of things that that we trust to be done, which in some cases is provision a little bit more infrastructure.
[00:10:40] - [Speaker 1]
Some cases, it's do more autonomous testing or do more on demand testing to make better recommendations. And we're seeing, you know, probably 25% of the basic network operations, especially if it's a we have confidence this is not an issue and close the ticket, are getting automated. But again, I think, at least this year, humans are in the loop on most of those, and you need to have the governance guardrails to make sure that people aren't using Clodbot on Mac minis under their desk to to run important corporate assets.
[00:11:14] - [Speaker 0]
You joke, but I that happens, doesn't it?
[00:11:17] - [Speaker 1]
Oh, it does. It does. But, you know, I mean, credentials Yeah. You know, a lot of these systems, you need to be very careful what you put in and what what what endpoints are exposed even just inside the enterprise. Right?
[00:11:34] - [Speaker 1]
A lot of people are building these ungoverned systems that have proxy infrastructure proxy access to pretty critical infrastructure. So that's a bit of a nightmare for CSOs and CSOs, and, you know, that's a whole separate, I'm sure, industry trying to help tame that.
[00:11:49] - [Speaker 0]
And, of course, you've had an incredible career and got somewhat of a unique vantage point here because you've been operating large scale networks since the nineties and have watched the Internet evolve from those early ISPs and dial up to cloud computing and indeed now AI infrastructure. I'm curious if you look back there, what what lessons from those previous technology waves are are companies forgetting about as they now begin to build AI? Because one of the things I'm learning as I get older is everything is cycles, and we see the same cycles returning time and time again. But what have you seen here?
[00:12:22] - [Speaker 1]
You know, it's interesting. I was recruited. I was at Akamai for ten years, and a a company that is now a large web scale company, they're a major technical major technology player, was recruiting me to come from their network, and they said, we need you to make a more reliable network for us because Google is more reliable than we are. And I said, you know that the reason they're more reliable is because they build around network failures, not because their network is more reliable. Right?
[00:12:51] - [Speaker 1]
It's about the architecture that you have overall and resiliency and and and redundancy in a hybrid world. And I think that's the thing that people still need to remember is a little bit of over provisioning, a little bit of redundancy goes a long way. Because ultimately, we affectionately refer to them as the butt cracks will will cause, you know, will cause outages. You know, things get dug. Fiber goes down.
[00:13:14] - [Speaker 1]
Router router vendors have bugs. So it's really about the whole architecture that you deploy, which assumes that you're you need to over provision and and plan for redundancy and failure. And of course, you need to watch it and keep it healthy. Right? A lot of networks for for, you know, almost almost half a century, people have been discovering that if you have redundancy and then half of it's been down for a month and you didn't notice, then then and then the next link dies, all of a sudden you're really scrambling.
[00:13:46] - [Speaker 1]
So it really does need continued watching and grooming, you know. But if you do that well, then you could build a reliable infrastructure like Google and, you know, some of the largest companies in the world have.
[00:13:58] - [Speaker 0]
And many startup founders that could be listening and they're building AI products, Understandably, they're focusing on models, applications, and customer experience. But what do those startup leaders listening need to understand about network architecture, infrastructure economics before scale starts exposing some expensive weaknesses in their business. Because, again, it's something that you can get blindsided by the technology and and miss some of the basics there.
[00:14:26] - [Speaker 1]
As I said, you need to understand that network components individually can fail. Yeah. So you need to have redundancy in your infrastructure. There's no shame in admitting that that that networking is not a core, especially if you're a startup in another space. Networking is not a core competency.
[00:14:41] - [Speaker 1]
There's companies where, you know, we work with those groups. They have some of the most sophisticated people in the world at at at CoreWeave and Crusoe and Lightning. And the companies that are dedicated to running that AI infrastructure, a networking is a key part of what they do. So if you don't have that as a core competency, you can develop it. You can work.
[00:15:01] - [Speaker 1]
You can outsource it. And then if you use those kind of providers or the web, you know, or or or the large, you know, cloud companies, then you just need to make sure that your connectivity to them is reliable, redundant, and that you're actually monitoring, watching, and and grooming it to make sure that it's performing.
[00:15:22] - [Speaker 0]
And I think one of the big criticisms of AI is it's almost a sycophant. It will go out of its way to make you happy and agree with you. And what reasons I bring this up is I was reading online that you've said more CEOs should use AI to challenge their thinking, even ask it to argue with them and push back. So how do you personally use AI as a leadership tool, and what have you learned about getting better decisions rather than simply faster answers that just agree with everything that you say?
[00:15:51] - [Speaker 1]
Well, I I sort of evolved that technique in part because I'm I'm rather well educated and pedantic about the use of the English language. Yep. So when I would go into ChatGPT, Claude, or say, please review this document and find any errors. If it didn't find any grammatical errors, it would just suggest a whole bunch of things that were stylistic that I disagreed about. So I sort of I adopted a different technique.
[00:16:18] - [Speaker 1]
I would say, I only wanna know if you can prove to me that I have a grammatical syntax spelling error. Please then argue with me. And then I found that it refined my use of that. And then I really extended it. I was actually we we just finished a book for O'Reilly, and I would do the same thing and discover that it would it would still keep trying to to help me.
[00:16:45] - [Speaker 1]
And some of the ways that it suggested I found really useful. It said, would you like me to pretend that I am a pedantic IETF? That's Internet Engineering Task Force. Network engineer critic critiquing your your the last pedantic details of your discussion about direct telemetry. Would you like me to pretend that I am an application engineer that does not think network is important, and why should I care about this bug?
[00:17:07] - [Speaker 1]
And so I actually discovered that asking those kinds of questions and assuming, you know, inform informing it that I wanted to be arguing with me, like, you're probably you're old enough. You're probably familiar with the Monty Python arguments, Gabe. Yes. It it is very valuable. And as a CEO, often when it's wrong and it thinks that I'm wrong, but I'm right, it's that I didn't really explain my thesis well.
[00:17:30] - [Speaker 1]
And communication is such a key part of being a CEO that even when the AI is wrong when it argues to be, it still can be instructive. So I find it to be a very valuable technique, and it also eliminates those sycophantic tendencies.
[00:17:42] - [Speaker 0]
Yeah. It's been so good to bust out at that echo chamber. I would imagine that once you find yourself arguing with an AI that is very resemble or resembles many teams that you've been a part of over the years, you you almost think, oh, yes. That's better. I'm home now.
[00:17:59] - [Speaker 1]
In in well, I I I was really surprised. It's hard to comprehend how much knowledge is encoded in these models. Yeah. And so what we call our, you know, the unintended consequences, the side effects, the really just the the the expect unexpected surprises. You know, it really does suggest some things where you're like, wow.
[00:18:20] - [Speaker 1]
It really understands me. But, you know, it's really just all prediction underneath. But it was a really valuable technique, and I'm glad it suggested that. Because I have spent a lot of time in my life arguing with pedantic people who wanna go into endless detail about things that don't actually matter. It's true.
[00:18:37] - [Speaker 0]
Yeah. It's part of being in the industry for sure. And looking ahead, anything that excites you about the future of AgenTic NetOps and everything you're working on, working towards, anything that makes you wanna jump out about in the morning and excites you about where we're heading?
[00:18:51] - [Speaker 1]
I think it's really exciting that that enterprise systems are all getting decomposed. And finally, we've been talking all the way back to Corva and web services and all this stuff, and it just wound up being rooms of people making documents and doing more talking than coding. And with with MCP, which is really just APIs, we're actually seeing a tremendous amount of fluidity and actual benefits that enterprises are getting from connecting their systems together. I still think, again, you know, it's not gonna replace humans. Humans that use AI better are gonna replace humans that don't use AI well.
[00:19:26] - [Speaker 1]
But we're seeing a tremendous amount of innovation and pace. And pretty much every customer that's trying to connect things together to get insights and and help humans run their businesses, run their infrastructure is getting benefit. And I think that's amazing.
[00:19:39] - [Speaker 0]
Yeah. Me too. I think it's a great moment to end on. But before I let you go, big thank you for bringing a dual perspective of a more forward looking founder, boots on the ground network operations expert to your role as CEO. It will resonate with so many people listening around the world.
[00:19:56] - [Speaker 0]
Anyone that wants to connect with you, learn more information about Kentik and anything we discussed today, where should they go?
[00:20:04] - [Speaker 1]
Kentik.com, kentik.com. I am Avi@kentik.com, and I'm Avi Friedman on, the usual usual social places.
[00:20:14] - [Speaker 0]
Awesome. I will add links to everything that, you mentioned there. I encourage people to check that out. It will be in the show notes. Great chatting to you again, and we cannot leave it another five years until we talk again.
[00:20:26] - [Speaker 0]
But thanks again for joining me. Really appreciate you, Tom.
[00:20:28] - [Speaker 1]
Absolutely, Neil. Thank you for having me again. Look forward to catching up in a year or two.
[00:20:33] - [Speaker 0]
Wow. It's incredible to believe that it's five years since Avi last joined me, and today's conversation felt like a great reminder that technology changes incredibly quickly. Time moves incredibly fast, but some lessons remain remarkably consistent. Yes. AI companies can spend billions on models, chips and infrastructure.
[00:20:54] - [Speaker 0]
But if data can't move reliably between systems and users, then that investment will quickly lose value. And I also loved his advice on AI as a leadership tool. Instead of surrounding yourself with another digital yes man, ask AI to challenge your assumptions, push back on your beliefs, argue from a different perspective, and tell you where your thinking might be going wrong. That is so much more valuable than surrounding yourself in an echo chamber with an algorithm that just agrees with everything that you say. And even when the AI gets it wrong, you might discover that you haven't explained your own argument clearly enough.
[00:21:38] - [Speaker 0]
So give that a try. And as Agentic AI moves into network operations, perhaps the future isn't Skynet autonomously reconfiguring everything while network engineers head for the beach to drink cocktails. Instead, it's humans with increasingly capable AI advisors that are helping them spot problems, understand what matters, and make decisions faster. I'd love to hear your thoughts on anything we covered today. Would you trust an AI agent to make changes to the network and keep your business running, Or would you always want a human finger hovering over that big red button?
[00:22:17] - [Speaker 0]
Let me know. Techtalksnetwork.com. And if you're attending any tech conferences from September to December, have a look at the event page. We might be able to meet and have a hot drink or cold beer. But I've taken up far too much of your time today.
[00:22:33] - [Speaker 0]
Time for me to go now. I will meet you here same time, same place tomorrow. Bye for now.