The $500B Nvidia-OpenAI Deal, Meta's New Data Center & Your Hidden AI Agents | Lightwork

0:00How are they actually going to stop a model when it's running without anyone really watching it? Nobody has built a single compute campus at this scale. Interesting that the moat is now a credit rating. I feel like there's notoriously a graveyard. Right, in life and in product development, timing is everything. I know, you don't want someone rolling back the tape on your personal argument. And it deletes your source code. And all of a sudden your application is gone.

0:33We're back with another rig week. more than a thousand researchers across OpenAI, Anthropic, DeepMind, and Meta have signed a single letter on the pace of automated AI development. The superintelligence land grab continues, and we break down deals between NVIDIA and OpenAI, NVIDIA SSI, and Meta and BlackRock. And then we discuss the ways that AI wearables are changing, including Granola's new Apple Watch feature, Nerva, and various others. We sit down with Axonius' CEO, Joe Diamond, on the question that no company can answer right now, how many AI agents are running inside your walls? Plus three new deals across the Lightweight universe.

1:19Let's get into it. So hot off the press, as of when we are filming last night, there was a really big letter from over a thousand researchers from big AI labs like Meta, OpenAI, Anthropic. In a first, they all signed a letter to call for the pacing of the frontier. Did you have a chance to see it? I did. I read the statement. it seems like they are trying to push the U.S. government and governments around the world to agree that they need to have some method of pacing AI. And I think it makes sense. And it's nice to see that it was bringing together Anthropic and Thinking Machines and Meta and all of the frontier AI labs, as you mentioned, over a thousand signatories of this statement to figure out how to pace AI.

2:11Yeah, this was a list of a lot of heavy hitters, just for some context. This isn't just academics or people outside the labs. These are pretty important people that, you know, we view as pioneers in the AI lab space. I've tracked a lot of these researchers' works. Who signed it? Jacob Pachocki. He's the one who's always at OpenAI's demo days. Mark Chen, their chief research officer. Boris Cherney from Cloud Code. Mike Krieger. Mm-hmm. John Shulman, who's an OpenAI co-founder. So that's just a few of them. You have folks from DeepMind, from Meta, Dario Amadei himself. And immediately after, actually, or within a few hours after, actually, OpenAI and Anthropic also corporately endorsed it.

2:54I think in particular, the language that I would emphasize here is that they are requesting the U.S. government to support an international effort to develop the technical and governance tools to pace the frontier of automated AI development. And I think one thing to note here is that this is not a pause AI statement. These are not researchers who are like, we want to pause all development or that AI is an extinction risk. What they're asking for is a technical and diplomatic capability to slow down if needed. And in particular, they're mostly focused on automated AI development. So this is having some sort of guardrails around AI doing AI research.

3:38And the best analogy I can think of with this letter is, again, that they're not necessarily asking for the industry to do a pause, but rather it's sort of like demanding that cars have brakes. It's not saying that cars no longer should be manufactured or that we should slow down the development of cars, but rather that you have some sort of brake function. I think this is really interesting because right now we're very clearly in a geopolitical race. But even within the U.S., you have kind of this intra-prisoner's dilemma between all the labs. The gist here is that, as we know of Prisoner's Dilemma, the only known escape is to have some third body to referee this arms race as sort of like an outside enforcer.

4:20So I think this is a really interesting effort. It's, I feel like, unprecedented, and I'll be interested to see how the government responds. It's definitely a very interesting thing. I feel like no one is going to be willing to sacrifice revenue or any of these metrics. We're all talking about all of these, you know, constantly accelerating AI metrics and like making sure you're at the top of the arena leaderboard and all of these other things. It's hard to have any internal pressure at any of the labs to slow down. So it's nice to see everyone coming together to unify around this like a little bit of a safeguard.

4:58One thing I've been thinking about, though, is there's been a lot of comparisons to nuclear. But the one thing that comes to mind for me is if this does go through and there is some sort of enforcing body here to potentially slow down and pace the frontier, it's hard because I think there's not a great way to measure and verify. I think with nuclear, it's very clear when, because these objects are so physical and so large, it's very easy to determine whether or not someone is breaking the guardrails that have been set forth, right? I think with AI, all of these exist in data centers and so in model weights.

5:48And so it's very hard to really enforce it. And I'd be curious to hear how, you know, if this does kind of follow through what that verification looks like. Yeah. How are they actually going to stop a model when it's running without anyone really watching it? So Claire, we have three big stories all around AI becoming basically a balance sheet business. We'd started a software and now it means land and power and debt markets and all of these interesting things. There's the NVIDIA and OpenAI reported deal. There's Meta and BlackRock building a data center and NVIDIA and SSI announcing a partnership.

6:30Can you break down these three big moves in the data center arena today? Yeah, pretty hefty. I think we should go through each of them one by one because I do think there's a lot to unpack in this super intelligence land grab that we're currently experiencing. And as I'm sure people have seen in the headlines, we're seeing this multi-trillion dollar data center and infrastructure build out happening. And now there's an interesting moment where we're figuring out how big tech and Wall Street are all paying for this. And as we know, compute is this sort of like new global currency. And there's kind of an entire infrastructure, both physically but also financially, being built around this.

7:12I wanted to start with the NVIDIA OpenAI partnership. In particular, that one is going to be a 10 gigawatt data center project in southern Ohio. And for reference, 10 gigawatts is huge. That's roughly the electricity draw of New York City, which is insane to think about. That's huge. And that's much bigger than the meta one, which is only one gigawatt. Yes. Yeah. Nobody has built a single compute campus at this scale. And because of that, this campus could cost up to$500 billion to build. So insane number. Which brings us to the guarantee part and where NVIDIA comes in. So OpenAI wants to build and lease this thing, but it can't because it doesn't really have investment grade credit ratings and it doesn't have a history of profitability.

8:04It's a great company, but it's hard to underwrite this and guarantee that they can follow through on their payments. It's sort of like when you're trying to sign for an apartment, you need a guarantor. And so in this case, NVIDIA has come in as OpenAI's guarantor. And they've basically co-signed and said that, okay, if OpenAI ever misses out on rent, we will cover it. You might be thinking, what does NVIDIA even get out of this? I think for starters, it costs nothing to NVIDIA to technically co-sign this. A guarantee is basically a contingent liability. And so nothing moves, none of that cash moves unless OpenAI defaults, which is where there is a lot of risk.

8:48But where NVIDIA is investing its credit rating now, it's not investing actual money, it's lending its credit rating because it would eventually unlock a$300 billion chip sale into the facility. So you can kind of see it helps co-sign leasing of this 10 gigawatt data center so that eventually OpenAI can fill this data center with NVIDIA chips and send$350 billion. So that's, I think, the gist of this partnership. So interesting that the moat is now a credit rating. I imagine there will be similar partnerships to come, especially between these hyperscalers or chip providers and AI labs that don't necessarily have the financial reputation.

9:35But this would be one of the largest infrastructure commitments in corporate history, actually even rivaling a Manhattan project. So that's why it made a lot of news. I think what a lot of people are calling out or what I mentioned around just the kind of financing dynamics here where it's like NVIDIA is lending its credit rating and then the building gets financed and then OpenAI gets somewhere to put the chips and then OpenAI buys a chip from NVIDIA, which goes back to them. So it's a little bit I think people have been calling out the loop nature of that. At the same time, how I'm viewing it is that you're very clearly seeing demand here.

10:14It's just that you have a financing structure here where it concentrates that risk. And standalone, you know, all of these different links are very rational and make sense. It's just that, you know, when you have all this capital circulating between this very small number of players, that's where I think some of the concern comes from. Yes, of course. And it is just interesting to see the like financialization of intelligence, like before it's fully monetized, maybe. Yeah, which I think could also bring up the meta BlackRock deal. Absolutely. I think on this similar concept of the financialization of compute, this one is slightly different.

10:52This is not meta or BlackRock becoming co-signers. It's a different situation. Here, Meta and BlackRock are creating a standalone company, a JB joint venture. And the job of this joint venture is to own this El Paso data center campus. In particular, they are building a one gigawatt campus. It's going to be about a thousand acres, and it will be powered entirely by renewable energy. It'll probably cost around, they're estimating,$14 billion to build. But what's interesting here is that there's an 80-20 split between Meta and BlackRock here. And BlackRock's owning 80 % and Meta's 20%. Yes. And then there's a lease structure too, right?

11:37Meta gets a four-year lease with an option for a four-year renewal. And so what Meta is putting in here is it's contributing about$2.3 billion in land and the construction assets that they've already put in here. And in return, because they've already begun the construction, they're actually getting a billion dollar distribution back. BlackRock is putting up about$4.9 billion in cash and putting up a bond sale. This is$12.5 billion to fund the rest. And so lots of numbers here, lots of financial terms. But the thing that matters here is basically for the last few years, hyperscalers have really been funding, self-funding the compute built out.

12:18So and it was you know, it made sense. Meta, Google, Microsoft, they were building these data centers out of their own balance sheet and profits. And so in BlackRock's case, they manage all these pension funds and insurers who need a safe place to park their money. And they want to put money in something that grows. And in this case, that would be a 20 year lease to Meta, who technically is one of the most reliable payers on Earth with all these guarantees attached. So, yeah, in many ways, you can think of it a little less as like, you know, where NVIDIA and OpenAI were a kind of like co-signer situation.

12:57This is more the creation of a new institutional asset class around compute. Which makes sense because it seems like AI has moved off of our computers and into all the physical worlds where we're seeing it in all of these new applications. Let's talk about maybe the third and last and perhaps the most mysterious of these announcements. What happened with NVIDIA and SSI? Very mysterious. We barely know anything about Ilya Sutsgevr's SSI. It stands for Safe Superintelligence. And, you know, for people who are less familiar, Ilya Sutsgevr's really big name in AI was pretty consequential in OpenAI left-to-go build SSI.

13:42And so far, we don't know very much about the product. But what we do know is that they raised at a$32 billion valuation in February 2025. And this was priced mostly around a bet on Ilya as talent because this was pre-product and pre-revenue. So this compute deal is interesting because it's one of the first few pieces of news that we've been able to get around SSI. And what I think we're seeing here, and this is just guesses on what is happening. We don't have very many details on all this. My guess is naturally SSI is presumably attempting a scaled training run. And there is a sort of compounding mode where it makes sense to have NVIDIA's Verirubin chips to power that.

14:34And then if Frontier Labs continue to choose NVIDIA, their research improves the next set of chips. So I see that. And I think the other interesting trend here is that, you know, NVIDIA is basically, as you can see, creating this kind of ecosystem-wide equity portfolio around its own demand base. And you actually see this pattern with a lot of other companies too. I wouldn't say like SSI is the first company that has benefited from a similar sort of partnership. But now you can see how NVIDIA has all these positions, whether with... It's funding AI, it's building AI. all of it. Exactly. It's holding positions across equity, credit, and supply.

15:16So yeah, very interesting. Interesting to watch because I think when you read these headlines, it's like, okay, I see a lot of big numbers. I see the data center build up. I think what I always try to dig a bit deeper in is like, okay, but what does this actually mean about the economics of building this infrastructure? What does this actually mean for the labs? And what does this mean for NVIDIA? So we'll continue to see. And, you know, I think the, to emphasize the NVIDIA OpenAI one is still all rumored. So that's another thing I would caveat is like, this is all still developing in real time.

15:53And we'll see how it plays out. You know, I think it's an interesting motion and interesting way to partner. But it could all play out a different way. So we'll only have to see. Yeah. It seems like a lot of these are very nascent and we'll learn a lot more in the months and weeks to come as we unpack some of these relationships and where these go. Claire, did you see that Granola actually launched an Apple Watch app? Yes. I do not have an Apple Watch. I have been, I think, for the last year, stubbornly avoiding wearing an Apple Watch. Well, no one wants more notifications on your body, right?

16:33I think I have one in a drawer that I might actually just pull out because I actually want granola on the go. One of my pet peeves is I used to have a friend, and I love them to death, but they would wear their Apple, because of the data tracking, would wear the Apple Watch everywhere, which was fun. And, you know, it's great, I think, you know, to each their own. But what was crazy is when we would go to, for example, friends' weddings or friends' parties, she would put them on her ankle to continue to track the data. It's a no Apple Watch at a wedding. That is like the number one rule of weddings.

17:06So she was walking around with this Apple Watch because she was like, I need that data. She's like Anna Delvey at the wedding with an ankle bracelet. That's what I was like. You look like you're wearing an ankle bracelet. But I do think there There's like a cohort of people who are, you know, obsessed with tracking. Yes. And I mean, I I'm not wearing my Oura Ring today, but I've been a big fan of Oura Rings and what it's been able to unlock for me from a data standpoint. But I do think that's actually an interesting point is like, why do I love Oura Rings, but not like Apple Watches? But anyways, Granola is coming to the Apple Watch.

17:38Which is very exciting because I think that we all have those meetings where they're in a coffee shop or in, you know, these unexpected places and you want to take notes on it. It happens to me all the time. I'm sure it happens to everyone watching. And I think also it's one of those things where what granola is doing is really smart because it's something that a lot of people have. But to backtrack a little bit, do you want to explain why Granola is different? Because I think one big question I've gotten from a lot of people is like, why Granola is any different from like Zoom Notetaker or Notions AI Notetaker?

18:12Yeah. So Granola, backed by Lightspeed, is a product that I think a lot of people use on their computers now to record either a Zoom or a Google Meet call. And the benefits are that it creates a transcription of your entire call, which you can then take to your LLM to then follow up on important details in the call or in the meeting. And I think it's a really valuable tool for most folks in the business world. But there's been this increasing need for, you know, having AI keep track of life outside of a Zoom meeting. And I think that's where this big launch comes in. And I think there was a lot of excitement on the timeline around people being like, oh, you know, I have an Apple Watch, but I don't use it that often.

18:59And now because Granola's on it, I might actually use it. Yeah, I was like, maybe I should get an Apple Watch for meetings. As much as I've been, you know, anti-Apple Watch for a while, granola might finally the thing that gets me over. Also, isn't it like just kind of journaling your life? I feel like a lot of people are not just interested in this for the work side, but, you know, there are parts of our lives that we forget. And there is something nice about thinking about, is there an opportunity to journal moments that I otherwise might have forgotten? Like what use cases are you thinking?

19:33I think if I was in an important, if it wasn't taking me out of the experience and I was in an important moment that I wanted to remember, there are things where, you know, like actually maybe it's at a friend's wedding where like there's a script, like a speech that I loved or something that I remembered that I would have loved to remember after that I maybe otherwise would have forgotten. I think there are those moments. I also think people are constantly thinking about journaling as a way of like reflecting on their lives. And so I think there are a lot of use cases that are more personal, that are not business, that this could work for.

20:11So I'm interested to see whether that's as big of a need as I think it might be. I mean, I think that's the entire premise of that company called Friend, I think, which is the pendant. There's another company in LightSoup's portfolio that also has a pendant form factor. Sandbar is another one that raised, and it's like a ring. And I think those are more so focused on this idea that you can ambiently capture someone's life and have audio voice notes. And I think it's playing on that same desire of like, can I bring in all these memories and give an AI all this context on my life? And it might be one to truly just capture these memories, but two, it's a lot of great context.

20:55And I think what's interesting for me with the Granola note-taking app coming to Apple Watch is like you have a note-taking app coming to this form factor versus I think a lot of wearable startups have started first with like. Yeah, they've started with the product. Yes, they started first with capturing all the data. And that was humane back in the day, right? Like there was Rabbit. There's been all of these pins and different, you know, pieces of equipment. But I think they were all, frankly, a little early to this moment. Yeah, I'm curious. I feel like there's notoriously a graveyard. Right, in life and in product development, timing is everything.

21:37I think granola has met the moment. And I do think that there's going to be, maybe back to your point about the girl that wears the Apple Watch at the wedding, is like there's going to be a moment when fashion and form and function kind of supersede that like as long as it looks nice enough that that people wouldn't notice that it's a recording device I think there's a big market for that. Do you think Apple Watch x Granola will fare better than these kind of purpose-built you know rings or pendants? I think they fit two kind of like archetypes of human beings, right? There's the tracking person who's going to be perfectly fine with just hitting the granola button on their Apple watch and have, and recording a conversation.

22:29I think there's a lot of people that will prefer the pendant. And I think it just depends, like, you know, are you the kind of person that cares about the fact that Nerva's going to have these kinds of fashion drops and like changing it with the season and like essentially making it invisible jewelry or visible jewelry but invisible technology i think there'll be a case for both um i think what matters though is the technology behind the pendant as good as granola is when you use it on your computer or how it interfaces with clod and open ai and all of the products that you like to use because at the end of the day as we've talked about, all of the data that you get from your aura ring and all of these sources, eight sleep, whatever, if you can't bring it to this single source that you can then process and get more value, you know, what is it really doing for you?

23:24And I think that was kind of the problem with the humane and the rabbits is like, it wasn't yet in that space where you could take all that data and process it and get these like deeper learnings. Or maybe, you know, maybe you realize from having the pin on all day or the necklace that you say like too much and you want to reduce to the number of likes that you say in your day. And so you make a conscious effort to say like a little bit less. It could be these little small improvements for yourself that I think are really valuable that people will see a lot of value in. Yeah, I think that's a really good point in that the r1s and the humane pins of the world one i don't think nailed the the aesthetics part the form the i don't think so you know the the form part but even just functionally to your point i think where granola has an advantage is that it plugs into your work context too and so i think i i can see a use case where i'm taking uh granola to a work meeting um and it might just be a walk and talk and then that feeds into the same research that I'm doing from my laptop on my zoom meetings I guess one question that comes up to mind for me with a lot of these is just even looking at the backlash that meta ray-bans have had I think culturally there's a lot of angst around this idea that we're always going to be recorded I think there's been a lot of discussion around etiquette around recording too.

24:56So I'm curious, do you have thoughts on whether or not these audio recording apps will face similar backlash slash how they navigate this? I feel like we're already in the moment where we feel like we're recorded all the time. You get in an Uber, you see that there's a camera recording you. That is now a precedent that I think is set across a whole bunch of barriers. I think the line that will make people uncomfortable is the like when you're not sure that you're being recorded, right? When it's quite unclear. So we'll cross that. But I don't think that it's, at least on the timeline with the reaction to granola, it seemed universally positive that people have these coffee meetings or these encounters that they want to record and just don't have an easy way to do it right now.

25:44And I think to talking about what we always talk about in AI with like workflows being solved by products. Granola has a really elegant workflow solution to be able to tap that, get the notes, transcribe it, and then allow you to act on it in a way that I think many other products don't yet. Yeah. I think where I personally stand is I, you know, I hear, I hear the feedback and I do think there is like a gray area with this always on, constantly recording society. And I wonder how we navigate that. I would say I'm comfortable with like session based recordings, like this idea that we have a meeting and I need to record notes.

26:30And it's similar to bring an intern to formally ascribe or an intern to take notes for a meeting. I think where I see it being maybe even a bit more problematic is the ambient capture space where it's like, OK, I just have this constantly on. And for people on the receiving end, they might not know that they are participating in a recording. So I think that's probably a distinction. I think the industry will have to make is, is this a session based recording and session based device? Or is it something that is like ambiently there? I think that's where people have issue with something like meta Ray-Bans is that people haven't consented to being on camera versus I think for granola.

27:11Oftentimes, someone will say, can I granola this this walk and talk? And I think that workflow makes sense. I think the other thing I'm curious about is just I saw an article on this idea that we there was this woman who was recording even her dates. and yes and maybe it's like an extension of the idea of having an apple watch track all your data it was like she she just wanted more data on how her dates were going which i think comes it's like well intentioned but i think you know it is maybe definitely crosses a boundary yeah exactly i think there's a lot um here and it's an interesting cultural conversation but for me my sense is like granola is strictly a product.

27:59Feels very much like a productivity work product. Yeah. Like I don't find myself reaching for granola to record my my a day or hangouts with friends. I know you don't want someone rolling back the tape on your personal argument versus I think where I see a lot of value for granola is if I'm having a coffee chat with a founder. And, you know, I think if I'm able to just use an Apple watch to I'll say, like, can I granola this conversation. I see that as a great use case. I would say my personal sense is that's where they're also seeing a lot of their usage. I haven't met many people who are granolaing their personal lives.

28:36Yeah, but that is where all these other devices come in. And I think this area is only going to get grayer and grayer and grayer. Our guest today is Joe Diamond, the CEO of Axonius, which is the company that created the asset intelligence category on a very simple premise that you can't secure what you can't see. This year, Axonius crossed$200 million in annual recurring revenue, doubling in two years. And this month, they recently introduced an AI agent and an MCP server so that the AI tools already running inside an organization can query verified asset data instead of guessing at it. He's here today to answer a question that every leader is asking.

29:21how many AI agents are actually running inside your computer right now. Joe Diamond, welcome to Lightwork. Joe, it is so great to have you on here. Axonius plays such a critical role in our ecosystem, and you all have created this space called, or category called, asset intelligence. We're really excited to have you on here to explain that for us and also demystify a little bit of what's happening around all things agents and security. Yeah, there's a lot to impact there. So first and foremost, thanks for having me. It's great to be here and great to talk about asset intelligence as a whole. So it actually started as a category that was called CHASM.

29:59So that's Cyber Asset Attack Service Management. As you know, our space loves its acronyms. So no security conversation is complete without discussing many an acronym. But basically, it started under this guise of organizations needed to have better visibility and actually understand what was in their environments, what existed. and basically just give me an inventory. Because organizations at the time, this would have been around 2017, 2018, really had no idea. And actually still to this day, they really struggle with answering the question of what exists across my environment. So what are my laptops?

30:35What are my networks? What are my SaaS applications? What are my IoT and OT devices? And this presents obviously a huge security challenge when you don't know what's where, how do you actually go off and protect the risks that you have throughout the organization? So it started as Chasm, but there was a very clear need for that to evolve in what we think of as asset intelligence, which is now, don't just give me the stuff. Don't just give me the inventory and the things. Also help me understand how they interface with each other. Like there's this laptop, what SaaS applications does it have access to?

31:07What is it doing when it accesses those SaaS applications and so forth? Obviously, this has become of even more paramount importance with AI agents and the rise of the play of AI in general, because what is AI at the end of the day? It's an NHI effectively. It's a non-human identity. So you need to understand what are those AI agents doing with those endpoints? What is it doing with those downstream cloud applications? What is it accessing? What does it have access to? What can it do? What did it do? And so forth. So asset intelligence is very simply put, give me and help me understand all of the things that exist across my organization and what is it doing across my entire organization?

31:44And then help me respond when it's doing things or my assets are doing things that it shouldn't be doing and help me understand all the risks that are happening across my company. Yeah, I love this framing of real-time visibility into everything happening across your organization. Just to make it tangible for folks who might sit at a higher level, haven't really been in the room trying to track all this, before Exonius, what was that super fragmented process? What were people doing? and then you all came, you know, I think in the last pre-AI created this category of solution. And then now today with agents, it feels like there is almost another evolution.

32:20So walk me through how that's evolved over the years. Yeah. So in the past, people would use a different approach to try to adjust this. So it was either done with network sensors or with agents that are actually installed on the devices. I think it's pretty obvious why those approaches don't always work particularly well because you're not going to have an agent on every device, right? And not every device supports agents first and foremost, and then you're not going to control every network that people work off of. So when you go off and you rely on an agent-based approach or a network sensor-based approach in its entirety, you may only be getting 40 to 50 % of what your entire ecosystem, what your footprint looks like.

32:58So I often refer to this as like, okay, so then you're accepting this notion of 40 to 50 % of my environment might be the equivalent of dark matter. I don't know what's going on across that. And in a world in which AI is now as pervasive as it is, is that okay? And now with Mythos exposing vulnerabilities across all of these different constructs across our organizations, we could be talking about organizations that were historically managing hundreds of thousands, if not low millions of vulnerabilities across its ecosystem to now millions upon tens of millions of vulnerabilities across its ecosystem.

33:32just as a result of discovering that extra 40 to 50 percent that exists across its footprint. So what we've done is we've gone with an API-based approach to do all the integrations with the downstream IT systems and security controls. So we have complete visibility into everything without relying on a network-based approach or an agent-based approach, simply put. You mentioned Mythos, and I think for a lot of people, that was one of the more tangible data points around what is happening in cybersecurity is like the first aha moment of, okay, we've had so much cybersecurity infrastructure. It's amazing because it works seamlessly and it's invisible.

34:08And most people don't even touch that layer. But obviously with the rise of Mythos and these systems that now have insane cybersecurity capabilities comes also the risk of all these vulnerabilities not being patched. So in that sense, what role does something like an Axonius play for companies that are looking to prepare themselves for an era where, you know, a mythos level model is out there in the public? Yeah, it's just making it more critical to make sure that you understand what your environment is and that you understand all the risks and vulnerabilities that exist across that environment.

34:41So it starts with you have to know what's out there. And then what Hexonius enables people to do is help me prioritize all the vulnerabilities that exist across all those different aspects of my environment. Now, again, if we're talking about a world where you have millions of different vulns that you have to go off and respond to, you have to really be good at prioritization, right? Because you can't triage and respond to all of those things particularly quickly, even if you're deploying AI yourself to go off and respond to them. So you need like your top 10 list, your top 20 list of like, these are the most critical.

35:14You may have what's perceived as a critical vulnerability, but if it's on an air gap network that doesn't have access to anything and it's not in communication with any devices, then its risk of actually being exploited is actually pretty low, right? So that might be a lower priority than something that is a non-critical volume that is sitting on a printer that's on the open internet, as an example, right? So being able to pair together this notion of, I understand my entire ecosystem with, these are the vulnerabilities that I have with, here's how I need to respond to them in what order, and here's what I do.

35:47It became amongst the most critical things, and it's been a massive tailwind for our business. I love how you laid out how it all flows into each other. You all have now recently, though, launched some new AI security capabilities, it sounds like. Tell us a little bit more about those. So first, it started with us making sure that we launched in MCP, a model context protocol, which gives organizations the ability to take their frontier models that they already have, that they've deployed, and leverage the data that exists within Exonius to pair and ask questions of the Exonius data. Exonius has asset context data that you can't really get from anything else.

36:24So now they're able to take their, you know, what they're asking of CrowdStrike paired with what they're asking of Exonius to draw a conclusion for how it is that they need to maybe respond to a forensic issue or a particular compromise that they're trying to chase down and then respond to. So that's one clear example. The MCP launch is coming out effectively now. The other aspect of this is what we call AI agent within Exonius. So what this is, is you can go into Exonius and then rather than relying on search or a query language, you can ask anything of the platform in plain English and it will tell you exactly what is going on across your environment.

37:00So as an example, you can say, show me all of my Windows endpoints that have a severity vulnerability level of eight that is a high risk and on the open internet, as an example. And it will just give you, hey, here's the list of the things that you need to go off and respond to right now. And here's the recommended actions in order to go off and respond to it. So we're pairing sort of the ability to interface with a pretty complex set of data in plain English, but then also give the here's what you should do about it in plain English. So it's super actionable right away and they can go off and actually address it.

37:34And historically, without something like an agent in Xonia, it's like how long would that workflow have taken or how much friction was there? To be blunt, you might just not be able to do it at all. You would probably going off and trying to pull that information from maybe 30 or 40 different systems. You would have to manually curate and stitch that together. And the way that organizations are oftentimes trying to solve this right now is through that raw approach and through the combination of spreadsheet dumps and VLOOKUPS and manual curation that way. And there's a lot of hope for the best that comes with that sort of paradigm.

38:09But in a world in which maybe that was acceptable six months ago, nine months ago, 12 months ago, by the way, it wasn't acceptable then. It was just we were a little bit more okay with accepting that risk at that point. Now, the impact can be quite dire. So we need to, as an industry, have our fundamentals in place. our industry in general has this problem of, I call it shiny object syndrome, where we're always moving to the next shiny thing that we think that we should be solving while forgetting that a lot of our fundamentals from a security program perspective weren't necessarily in place properly either.

38:47It sounds like it's things like agents inside Xonias or this MPCP that are massive step changes and unlock some value because they're just simply core workflows, fundamental to the industry. I guess switching gears, one of the things that stuck with me as I look at Xonius' company is this idea that you all have created this consolidated space where people can get full visibility and everything happening in their company from a cyber standpoint. And one thing that stuck with me is that this idea that nobody can tell you how many AI agents are running in your organization right now. Before we get into how you all approach that, I think it would be helpful to explain a little bit more about why it is so important to even track this and why is it different from tracking an agent versus tracking a human employee or tracking traditional software?

39:36Yeah, I think of this as similar, but at a much grander scale than what we were talking about in the past with shadow cloud adoption, shadow IT, you know, 10 years ago, 15 years ago or so. Everyone was off adopting cloud applications that weren't particularly sanctioned. And we had to figure out how to deal with it as an industry. And everyone was terribly scared of that at the time, you'll recall, right? We all remember those moments. Maybe I'm aging myself and maybe you don't remember those moments, but I certainly do. but from an AI perspective it's that but much worse because you know back then you couldn't it was harder to go off and deploy you know a cloud application but in this case anyone can hop in a browser and actually do some pretty some pretty interesting or crazy things with just basic AI chat prompts that may be exposing very sensitive corporate data so being able to actually understand what the AI agents are, how these tools are being used, where they're being used, and how they interconnect with every single aspect of your organization, whether you control the device or the network or not, is the most critical thing that organizations have to face today.

40:47The reason why that's so critical is, again, they have access literally to everything. And why this is different from the human, to answer your question, is how much more quickly they can move. a human is going to bounce around and maybe do something accidental through negligence or, you know, so the blast radius is much smaller. Whereas with an agentic approach, it can literally touch everything within minutes, seconds across your entire organization, depending on what it has the permissions to do from just a read perspective or a write perspective, or even a delete, you know, standpoint, the damage it can do is in some cases can be irrecoverable.

41:26Yeah, I was just thinking about Demis' recent piece of like, you know, as you think about the progress that we've been able to achieve, he almost describes it as like 10x the industrial revolution. But the scary part is at 10x the speed. And so I think dealing with that speed and dealing with that scale, maybe to help people understand just like the gravity of this, what would like us at scale AI agent security breach even look like? Like, kind of give an example of a worst case scenario for a company, for example. Yeah, I mean, the worst case scenario can be as brutal. And we've actually heard of scenarios like this where it goes in, you're asking it to do something that is, you know, theoretically relatively straightforward and it deletes your source code.

42:10And all of a sudden your application is gone, right? It has access to your HCM or your financial systems and that it does something it shouldn't do and all of your customer data is gone or something to that degree. And remember, it touches every single aspect of your environment. And we are all still learning as an industry, including the frontier model developers themselves, like how far this is going to go and the exact possibilities of where it's going to go over time. So there's a lot that we're learning together as an industry. And I think the worst case scenarios are oftentimes not fully understood.

42:44it. I know that there are great safeguards that we've put in place, but there are a lot of cases where they have moved around these particular safeguards that we've had in place. So there's a lot of unknown. There's some that is known. But the line that I've been using to describe this is like you talk about 10x or 100x of the industrial revolution. I often talk of it in the context of like, this is as big, if not bigger than the invention of the internet itself. So we are not fully understanding the possibilities of where this is going to go in the same way that When, you know, TCIP and, you know, the internet first came about, we did not know where this was exactly going to go over time then either.

43:20This just has potentially and theoretically much greater potential. I mean, truly just building the plane as we fly. In the same way that I think, you know, to your point with the internet, are you seeing, you know, uptick in adoption in your customer environments around tracking these AI agents and kind of new forms of assets? folks are trying to figure out how they use existing tools to solve the problems that are in front of them. So it's like, hey, we bought these solutions from these platform players. What can we get information-wise from the platform players that are already in place? The challenge is that typically ends at the network layer and the endpoint layer, right?

43:59So it's like, what about everything in between, right, becomes the big problem. And then what What do you actually do about it? Like what becomes the common fiber or the control plane that brings all of those things together? So when you think about the evolution of Exonius, it isn't purely for us to just be an asset intelligence platform that helps you manage your exposures. It's actually to be sort of the common control that brings your IT stack and your security stack together in one place to basically manage your entire security posture. And that's something that is obviously a, it's a very broad vision for where the company will be going over time, but it's a necessary one because there isn't a particularly good solution that exists in market to help us solve these problems as an industry.

44:41Yeah. And it's helpful that you all are playing at that visibility layer, but then you're actually going a step further and thinking more broadly about action off of these identifications of issues. Speaking of that incident growth driven by the tailwinds that we're seeing around agents and making sure we're building a security layer around AI agents, you all have just crossed a big milestone of 200 million ARR, 100 % growth in two years. Super impressive. You're at the helm of this as the CEO. What is next and what have you largely seen as you've seen this kind of insane growth? Yeah. The phenomenal thing is we saw a lot of this growth really, I would say, before we caught what I would call the tailwind or now the jet stream of what's happening right now from an AI perspective.

45:28Because again, asset intelligence went from being like certain types of companies really need to have this to now everyone needs to have this as a part of the foundation of their cybersecurity program. So that is a phenomenal thing for us to catch, right? You know, from this tailwind standpoint. So we're obviously, you know, anxious and excited to tap into that. And we are as we speak. The other aspect of this is just every cybersecurity company always has to ask itself, is it in the right place in the cyber stack to solve a specific problem? We are in the right place in a stack as sort of like this neutral agnostic API layer that doesn't care about the network, that doesn't care about control over the device to basically secure AI from end to end.

46:11So what's next for us is making sure that AI is deeply embedded in the platform itself, deploying agents on top of all the unique context and the data that we have within our platform, but also how it is that you look at AI as an asset class. So it's no differently than you look at cloud applications, software that's installed in devices, IoT and OT as asset classes, AI is now an asset class as well. You have to be able to catalog all of your AI agents, what they're doing, where they are, how they interface with your entire organization. And that's our next movement. From a vision perspective, like we talked about, over time, you'll see us become sort of like that larger control plane, literally for everything.

46:53And driving action is sort of like the end state, if you will. We are the platform that makes it so you have visibility into everything, but also makes it so you can orchestrate and respond to all the things to drive action that is an issue across your organization. I want to definitely spend a little time on the vision towards then. But before that, I actually want to also hear a little bit more about your story, because I do think we have a lot of founders, aspiring founders, operators who tune into our show. And I think it's awesome to hear about your trajectory. So we'd love to hear more about, you were a CMO in 2024 president 2025 ceo this year um curious about your own trajectory of axonius and how that impacts how you think about the company and you know your your journey alongside axonius yeah i get this uh i get this question quite a bit uh in particular of the the atypical journey from cmo to ceo um maybe it's not the most common thing but just from a historical standpoint i never really believed in swim lanes like from a role perspective i always sort of carry that you know that ownership mindset of whatever the company's need is, is my need.

48:00Whatever problem needs to be solved is my problem. So I never saw swim lanes in terms of roles. First and foremost, I came up as a technologist. So I have an engineering background and then moved into product and then moved into marketing and then basically had run virtually every organization possible for the most part in a software company. And that was purposeful by design because I wanted to make sure that I understood every in and out of a company in as much as I possibly can. Everyone at the end of day as an archetype, I think. By background, by design, I would say that I am a product guy, but understanding all of the dimensions and all of the makeup of a company, the technology side paired with the go-to-market side, paired with everything from a G &A perspective tends to be pretty helpful in this role, I found.

48:42And it's been a lot of fun. In many ways, I feel like a lot of alignment in that I sit in this investment role, but I do a lot around media and it feels like just narrative and storytelling and being able to drive these narratives is a big part of my day-to-day job. And so those blurred lines definitely resonate. I'm curious, you have just seen the evolution of the platform going through this big AI paradigm shift. You're a CEO of a company that is constantly and needs to build the plane as it's flying. How do you keep up with all the changes in the space and make sure you're building in the right direction?

49:16There is no keeping up with all the changes in the space, I would argue. because it's moving so quickly. It feels like it's a new model a day, a big shift a day at this point, if not multiple. I personally probably spend at least two plus hours a day just on AI. I personally deploy and play with local LLMs all the way through playing with the frontier models myself at the deepest possible levels. So I don't think you can do any of these jobs effectively without being an expert in them yourself. And so I make that a big part of my job. So that's, that's aspect number one. And then aspect number two is, you know, just staying very close to the customer, staying very close to what's happening in our market as a whole, in terms of the needs and like what it is that organizations are doing and what they're primarily concerned about.

50:06And then of course, staying very close to the investor community, because they have a really great pulse in terms of like what it is that they're seeing across the board as well. And, And everyone is feeling kind of like the full spectrum of the paradigm of one end of the spectrum of gold rush versus the other end of the spectrum being bubble and figuring out which one is this. I tend to think it's more on the end of gold rush and not just a point in time or a hype cycle. But there's a lot of ground to cover. I don't think it's possible to stay up on everything. But if we're not as aggressive in spending time and getting our hands dirty, you just won't know the details.

50:40Of the infinite ground that could be covered in this moment, whether it's focusing more on product, keeping up with the agent capabilities that you all are releasing, how are you thinking about breaking down your time and where the individual teams or functions are pushing more? Yeah, the Venn diagram of overlap might be more than what people think. We do have a lot of focus on how it is that we drive operational efficiency for all teams across the business from an AI perspective, because I do think there's a lot of lift to be gained. I think that one person can very easily become tent if they know how to leverage AI properly.

51:18You know, there's a lot of pressure in our industry, as you know, and a lot of organizations are figuring out like, oh, like, you know, do we have to make cuts in certain places? Like, how do we make those cuts? I don't personally look at it in that context. I look at it more in the context of how do we magnify the team that we have? How do we make them more streamlined? How do we make them more efficient? How do we make them, you know, more capable? How do we make them faster? Like all of those things, like how do we maximize what it is that we have? The other side of that coin is how do we build robust AI capabilities into our platform?

51:48And do you have a unique qualifier for why that has to sit in your platform? Like we were talking about before, AI at the end of the day is a data game. If you have unique contextual data that you can't get anywhere else, you're ahead and you can be considered a winner. If you don't, you're kind of right for pretty profound disruption. So that's how I look at it. And I spend a lot of time thinking about both of those dimensions, a really significant part of the day. It just like translates through the camera that you were like energized about and curious about this season, constantly learning.

52:20I feel like that's a common theme of so many great builders and founders in this space is this constant needs and you can't there's no one playbook. And so much of it is this ability to learn and adapt in real time. You know, I think last two questions, you have this very unique window into the state of security and AI. What does securing an org even look like in five years, 10 years? Obviously, there's no predicting the space, as we know, but in the best way possible, how can we make sense of what it's all going to look like? It kind of goes back to a little bit of the fundamentals that we were talking about before.

52:55Are the fundamental principles for cyber programs going to be that different in five to 10 years as they were even 10 years ago from now, right? The reality is, is like, again, the fundamentals from 10 to 15 years ago, we still don't have in place today to the degree in which we need to have. go sit down with any CISO and ask them when they don't have Exonius in particular, what they know about their environment, what sits where, who has access to it, et cetera. More often than not, they're going to say, we don't, or we have a confidence score of like five or four or three out of 10, something to that degree.

53:30Now, it's not something necessarily that people want to speak super proudly about because it shows a hole in their overarching security posture, but it really is always about getting the basics right and the fundamentals right. the things that are going to evolve, the things that are going to shift over the next five to 10 years, to your point, are like, what new attack surface areas are opened up? Like cloud years ago was a new attack surface area that we had to figure out how to go out and protect. AI is a new attack surface area that we had to go out and figure out how to protect. What new emerging areas are going to come?

54:03That is yet to be seen. And we're going to have to deal with the notion of quantum. Is quantum going to become a reality? And how much more efficient are these agents and these AI and compute going to be as a whole in terms of how quickly it can move? And what effect is that going to have on our overarching attack surface as well? So those are the things that I think of. I still think it's going to come down to fundamentals, but there are going to be new emerging threat factors that we're going to have to consider as a part of it. Insane to think about, but I'm glad there are people like you thinking deeply for a lot of these companies.

54:35And last bit of advice for the founders and operators who are watching, I'm sure everyone's, you know, certainly after this conversation, thinking a lot about how they can properly deploy AI agents. Do you have any advice about how to think about all of this? Yeah, do the basics, get the fundamentals right, not just for how it is that you secure your environments, but also for how you run your business. I'm a big believer that you can do three things well at any given time. Make sure you're doing the right three. Put the other 500 on the back burner and realize that you can only do so much at any given time.

55:09You know, maybe AI is going to help us get to being more effective at doing the three, prioritizing the three, or maybe we can even get to four. But we're not particularly great at multitasking, you know, so we have to embrace that as a reality, do the right things and nail the fundamentals. Thank you so much, Joe. This has been such a fascinating conversation. It's really made me think a lot deeper and take this 30 ,000 foot view about the future of AI and security. Because, you know, to what I pointed out earlier, it is something that feels so invisible and everything runs so smoothly because of companies like Exonia.

55:43And it's just interesting to peel back all the chaos and potential chaos that AI might unlock and the ways that you all are making sure that everything we're working on is keeping safe. That's right. Thank you. It's been a pleasure.

56:00Andara, which is the AI-native audit platform, just raised a$37 million Series A led by Lights Feed. Andara's agents can read unstructured evidence, whether that's spreadsheets, PDFs, and screenshots, and produce these audit-ready work papers. And what the company says is about a third of the time of a purely human team. Right now, Audit is one of the last big professional services categories still being built on manual evidence review. And their bet is that agents can do the greeting while humans keep the judgment. Cusp AI raised a$450 million Series B. It's a search engine for materials that don't exist yet.

56:42You describe the properties you need and it generates candidates to match. The hard part was never the idea. It's the proof. So Cusp.ai built a foundry network of more than 45 partners, including NVIDIA, Samsung, and LAM Research, to move candidates into real labs. Materials sit underneath everything. Chips, batteries, carbon capture. And it's one of the few corners of AI for Science where a discovery has an industrial buyer already waiting. Harmony raised a$34 million seed round led by Lightspeed. It's an AI platform for employee support across IT, HR, finance, and procurement, and to replace ticket queues with agents that work inside the tools that people already use.

57:28Right now, internal support is one of the few AI use cases where the buyer already knows the cost of the status quo down to the ticket. And Harmony's entire angle is meeting people in Slack and email rather than asking them to start and learn another portal. This week was all about the financializing of AI and it truly growing up into this brand new industry with all of these asset classes that are emerging to shape the future of our next generation of AI products. Yeah, and all while people are still figuring out what all of this compute and AI actually looks like in the hands of real users, I feel like we had many different takes on what wearables should look like.

58:11And if anyone has thoughts on what your most optimal AI wearable looks like, let us know. But see you next week on Lightwork.