AI Agents vs. SaaS: The Future of Software

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone, I'm Rob May. I'm the co-founder and CEO at Neurometric.

Ryan Eppley: I'm Ryan Eppley, co-founder and CEO of Root Access, and this is the AI in NYC show.

Rob May: All right, welcome to the AI and NYC podcast, where we cover AI moving at New York speed instead of Silicon Valley speed, which is faster, by the way. I want to give a shout out to our, latest sponsor, Lorimer Ventures. It is a venture fund that is, has a network of a hundred operators who are hands-on, can help you with like real work tasks, all that kind of stuff. If you're raising money, definitely check out Lorimer. And, also go check out our other shows. We have a new newsletter. Subscribe to the newsletter. Other than that, let's get into into what we're doing today. We got a we got a great show. We have a partner at a large infrastructure fund in New York who has a really interesting AI agentic software thesis. So, we're going to talk a lot about that and some of the news that goes into that. So, welcome Kenneth Auchenberg.

Kenneth Auchenberg: Thanks for having me. Excited to be here.

Rob May: Yeah. Yeah. So, so let's let's actually start. Tell us a little bit about Innovation Endeavors. What do you guys do? And you're recently new there. What made you decide to join?

Kenneth Auchenberg: Yeah. I'm a partner at Innovation Endeavors. We're the fund of Eric Smith and Draw who started the fund more than a decade ago. And what we do is just backing technical founders. So I focus on software infrastructure and AI, but we cover biotech, robotics, deep tech, and really build on the premise that we want to back like the next frontier builders. And also we just doubled down and saying like we technical people and we love working with technical people and we think there's a lot of marriages in being able to focus and really just go deep with the founders we work with.

Rob May: Yeah. Tell us a little bit. So this show is interesting because we have some connections. You and I sit on a board together for another company.

Kenneth Auchenberg: We do. Yeah.

Rob May: And you the fund you were at backed Ryan's company. So there's a there's a lot a lot going on here.

Ryan Eppley: Yeah. You sat in one of our first board meetings too and we had a great discussion and you came by and Yeah.

Kenneth Auchenberg: So yeah, I mean New York is a somewhat a small ecosystem, right? [laughter]

Rob May: Yeah. Well, tech tech is still a relatively small ecosystem, I think. But so tell us a little bit about so how did you get to innovation endeavors? Like what was your background? You know, were you an operator, entrepreneur, investor and how did you get here and how did those things lead to you saying like this was the right thing for me at this point in time?

Kenneth Auchenberg: Absolutely. You know, I have a bit of a fun background is I don't have the typical kind of venture background. I started my career in Denmark in Europe as being 16 years old starting to write software and that's kind of when I joined my first startup. We were back then selling business intell intelligence which a fancy way of saying I wrote a scraper to call news sites and then we sold that to CM systems and it turned out back then we didn't know what we're doing so we went bankrupt. And then I kind of rolled on to the next thing. Back then it was a pre iPhone. We had our Nokia and Sony Xenon phones. I feel really old saying that. But we built a solution where you could back up your phone numbers to the cloud. That was really our servers in the office. But that became really popular. So we end up selling that to Vodafone.

Ryan Eppley: Yeah. Oh, I worked on a Symbiot the good old days.

Kenneth Auchenberg: And then after that the same crew, we did another company in kind of social collaboration space, end up selling that to CRIX. And then we did a company in the in the HR space, end up selling that to Workday. And that was kind of like back then I mostly did engineering, mostly worked on web platforms, developer tools and I realized hey maybe this infra thing is interesting. So I ended up actually getting hired by Microsoft on the product side of house back then. It was like a year after Satya became CEO and I joined to help them with the JavaScript strategy and how they thought about net new developers. That was the new new thing for Microsoft. So joined Microsoft on the product side ended up building a product called Visual Studio Code was the PM for VS Code for a few years.

Ryan Eppley: Yeah, that small product VS code, right?

Kenneth Auchenberg: Yeah. [laughter] You know, like it it was it was a fun experiment back then like we had we were experimenting can we build a high performance code editor in the browser and then we've experimented and really like my mandate back then was like let's do a lot of things differently when the old Microsoft would do. So we did open source development the open

Rob May: I mean a bunch of AI ids are built on VSD right

Kenneth Auchenberg: yeah all the good ones right like cursor cursor was a $60 billion outcome B built at least in the editor part on on the foundation that we open source back then

Rob May: that's awesome

Kenneth Auchenberg: yeah so long story short also had a stripe work on the developer platform there so spent the past two two decades building and then a few years back I kind of made it leap made the leap to the dark side of things or the more interesting side of

Ryan Eppley: I love how people say like the dark side when they say venture. Like I don't know. I have I have a good handful of investors that I think are super great builder still. They're just like add and driven, right? Like they're they're always curious about building too many things so they can't be a founder anymore.

Kenneth Auchenberg: I you know you know like I am I'm I'm probably a little old school but I think venture is a job you kind of graduate into after career building. I don't think it's job that you start your career in.

Rob May: I think that is the right approach.

Kenneth Auchenberg: Yeah. And yet like and I also think like you know like you're going to earn the right to work with founders and I think that you do that by building and learning.

Rob May: Nothing's worse than pitching a 23-year-old newly minted MBA on your business who's an analyst at a big fund and they don't really even know the first thing about any of this stuff, right? And it's that that happens too much. And I also think I think the big venture funds a lot of times they hire a lot of the wrong people. And I do I do I think you want your background. I think I think you want builders, right? I think you have people that have seen the ideas that like sometimes you have to take a step back, two steps back to take two steps forward and whatever and all those things and you have to make hard trade-offs because you have limited resources. So you're you're constantly like the whole business is just thinking about hacks. How do I hack this thing to get more out of it, right? Like that's where a lot of your great ideas come from.

Ryan Eppley: Yeah. I mean we got to discuss this only a little bit, but this is a great format for it. Like walk us through like your first inclination when you were like wow like VS Code is really going to be something that's revolutionary cuz you mentioned you were part of like the first couple hundred people on boarding just like the Q&A like hey how how does this serve you as a developer like what else do you want to see from us like walk us through when that inflection point really happened for you personally

Kenneth Auchenberg: you know like like back then the team we were what were we on the core team like we could fit in the rooms we were like five six people trying trying to build this thing and I think for us you know VS code start started as an experiment on how we were building could we build a high performance code editor running in a browser yes we could we started it all inside the Azure portal turns out no one really were using the Azure portal not especially not for JavaScript it's like how can we take this technology we built and kind of build it for a net new audience and I really think it was an experiment of like how can we take the learnings of how we've been building tools additionally at Microsoft all the value we've provided to and other audiences and how can we bring that to this net new audience that was a whip developers

and as a part of that we did VS code we did TypeScript and we were taking all the learning learnings for how you actually building software at scale and applying to a net new audience and I think for us you know the first audience we're going after was like the full stack web developer that were building a simple front end with a NodeJS server and it's really about like how far can we go to make that developer successful end to end and once we kind of like saw we actually make them successful. We could write a better experience than back then it was like Sublime Text and Atom and then we were like, okay, we're on to something here. And then we started thinking about how can we make this product scalable into a platform, how we can go after new kind of workloads. And that kind of led us down this platform journey of building a really a code editing foundation. But in the early days, you know, like even though that I was on the product side, I ran the Twitter account and that was me at code, that was me managing that account, I ran the marketplace.

Ryan Eppley: Did you do any memes? Did you throw any memes in?

Kenneth Auchenberg: We did a lot of things, but also like one of my most proud moments was when we learned that we were the code editor in the HBO TV show Silicon Valley.

Ryan Eppley: Wow.

Kenneth Auchenberg: That was that was like that was that was the biggest moment internally where like I got access to a clip where I could take a screenshot and send that to the whole team. It's like we made it that we made it in into

Rob May: I didn't realize that's a great that's a great shout out.

Kenneth Auchenberg: So that was a that was that was a proud moment.

Rob May: Yeah. I had a I had a meme this week. The Mistral team or no somebody on Hugging Face, one of the Hugging Face's CTO put out a thing and he's like, "Hey, the new Mistral model dropped or whatever and it's called like they were calling it fat cat or something and it like crushed their servers and it was this whole thing and he's like now we're going to go out of business and I was like wait this is this is a joke, right?" And my team's like, "Yeah." And there all these like model memes and stuff. So, yeah, that's one of the great parts of tech.

Ryan Eppley: Yeah, meme culture is so important. So important. Walk me through the development of the marketplace because I thought that was so one it was a beautiful moment where people could be like wow I can serve my niche and my community by just building something bolted on right it wasn't revolutionary in a business model sense but I think it was revolutionary for developers to provide value to their niche communities like how did that experience of onboarding the first marketplace you know users go

Kenneth Auchenberg: I think like you know any kind of platform or product usually go through the evolution of being single product multi-product platform and I think for developer tools the reality is that how should I put this developers are so opinionated about how the tool should work and that really drives the need for customization and that means that you cannot provide a one-sizefits-all for every developer especially not when you're building a foundational tool that is a code editor debug or terminal so that really led us down a path okay we need to enable that medium to long tail of customization. What are the best way we can do this from a first principles perspective that is through an accessibility model and that was [clears throat] that was really like the driving lesson but also our team in the past have been building other code editors. A lot of the original VS code team has building was building IntelliJ so we kind of knew what were kind of working but how could we apply this in a net new secure kind of way.

And I think the other lessons for from back then is that we actually dog food a lot of the extensibility and the marketplace inside the product. So even though that you were just installing VS code everything was built with extensibility model everything was built with the protocols and we actually dog fooded the our full end platform for everyone and then when we were ready when you had been battle tested then we flipped the switch and then when you flip the switch the problem is always like how do you bootstrap a marketplace? Yeah. And I think the typical mistake a lot of people go for market places that they go to all the big logos. They want all the shiny logos. But really what you want is that you want to go to the long tail of people have like the obscure kind of use cases where you can give them distribution. And it also helped that we were Microsoft.

So I could go and knock on the door at Anders who was sitting next to me. He did TypeScript. And then I could go to the C# team etc. And ask them to kind of start building on building upon our marketplace and kind of bootstrap the ecosystem. But it's really about like how can we enable that long tail of experimentation and then that's how you kind of like nurture an ecosystem or garden an ecosystem is that you kind of like have a lot of sprouts rising and then you see what kind of works. Mhm.

Rob May: So since you've moved to innovation endeavors, you wrote a piece that is really really interesting called I a decade of software infrastructure. Yeah. And the things that you've learned and everything else and you developed this concept of a agent first infrastructure which I think I think is super important for where we're going. Tell us a little bit about like what that means. What's the insight that you have and like what you know how are people like us that are starting companies like how should we think about this?

Kenneth Auchenberg: Yeah. Like I think maybe we can take a step back on what's kind of happening and what has happened to software engineering. I think you know I just talked about VS code a product we built more than a decade ago. I'm not using VS Code anymore because my agent is writing the code. So we are a fundamentally a different kind of reality when it comes to writing software is that when you're building an infrastructure company today you're not really targeting the individual developer anymore but you're targeting the agent. And I think you know we spent the past 15 years talking about developer experience and optimizing everything for developer experience but we're now going through this reality where we transitioning from developer experience to agent experience and that's because we have a net new consumer of everything that we do. So I think that's kind of what we've outlined in our thesis about how we look for the next decade.

And I think you can kind of summarize into two things is that that's going to be agent first infrastructure that as we're pivoting all aspects of our developer tools and our infrastructure chain to be optimized for an agent first. And then there's the go to market side of that. And I think there's kind of implied assumptions of what it's going to mean to have an MCP server have some skills. But I think the other side that we're covering is like what does an agent first world mean for distribution? What does it mean for procurement? Because we're already starting seeing the early innings of that of like agents selecting tools making decisions for you etc. But if I were to just expand on the product pieces, I think today when we're sitting and writing code, we're using Codex or clo or open code, whatever flavor of the wigu you're using, but that's really producing the code.

So that means all all the abstractions that we provide have to work really well in that kind of agenda context. But it also means that how we think about end to end full stack software is changing. And I think the chronicle example is probably let's say Salesforce. Salesforce went out and announced at the last conference that they're going headless. What What does that mean? It basically means that the future of Salesforce looks more like an MCP server and an API. And the kind of the UI that made their business, they're kind of like fading that away because they know that people don't want to sit and click in a in a crappy UI, but they want the agent to do things. And I can even see that in how I'm using software. If I can get Claude to use it, I'll just turn and use something else or I'll ask Claude to build build me a version of it.

So I think even the consumer behavior is changing really really fast. So that's kind of one thing but like I think if the UI kind of goes away what does that mean for for typical software? I think another piece we outline is our thesis around like engines and kernels in the sense that if we reduce the UI you actually reduce a lot of like the surface area you provide to be more of a kernel like or rendering engine. I think we can take a lot of inspiration for gaming where you have like a unreal rendering engine with APIs you didn't consume and I think you can take software like a Salesforce if you reduce that to an engine of what Salesforce is and you and I can have different user experiences. Exactly. And I think that also means that systems of records are kind of up in the air etc. Or maybe they're going to be more important but how we typically think about like building in software owning a workflow landing and expanding is kind of changing in front of our eyes.

But I think more more interesting to your point about customization is that it's really about I think my base assumption is that the software I'm going to be looking at on my screen is going to look quite different than yours because it's going to be hyperpersonalized. Yeah. And I think if we assume most of the business logic is going to be headless software but also means that we're going to move to a world where we're going to have I would say like on demand hyperpersonalized UI and whether I'm consuming it from my car sitting and talking to claude or I am sitting in my B2B in enterprise interface and I want a widget stream to me because I want to look at this I don't know report. I think like the constant how we consume software is changing in front of our eyes as well. So that kind of leads us down a path of saying the next decade for for infrastructure actually means a lot more software but also a lot more software is going to kind of be infos infra spiked or shaped because you're going to end up exposing APIs.

Rob May: Yeah, this will really help people like me. So I've always been a complainer about UXUI because I don't like common things and what most other people like. So I'll give an example. If I can avoid watching a video online, unless it's this show, of course, but if I [laughter] if I can avoid watching a video online and read the transcript, I will do it. And I know I'm like in the minority that prefers that, but if you can give it to me in text, I will take it in text all day. And so I'm excited about this world because I'm one of those people that always wishes the UI was different than it is. And then I'm arguing with everybody who's like, "No, no, this is a great UI. This is the

Ryan Eppley: So, well, well, actually, you know, I think this is one of the moments where everyone always clumps UI and UX together." And I think for maybe the first time and I as long as I can remember the first time ever maybe where we're going to see UI and UX really have to bifurcate and who owns that cuz usually it's been clumped together. Let's just be honest people hire and it's literally in the job title. I think that is going to be completely different jobs now especially when it's so dependent on your experience being customized to you. It almost doesn't matter what your base UI is anymore. And I actually think UI might become more important with how simple you have to start just to get people going and momentum just to get it set up because onboarding now should be the most bland UI but making something simple as we all know is really hard and then customize the experience for that. How do you do that?

Kenneth Auchenberg: Absolutely. I think like the form factor in the surface area is going to change a lot, right? Like I think if you're still shipping a VIP app with a desktop view and a mobile view in 2026, you're kind of not getting what's happening right now that one day maybe Siri will work and we can use Siri or something. But you know but at least so far I'm using my open claw and my Hermes agent to kind of consume things and sign up for things and you know I think like how you think about the UX of all those experiences is going to make like you know what you're building much more comp complex and there's a lot of complexity you need to manage but I think the way you're delivering that user experience is no longer just a static kind of dashboard you log into.

Rob May: What do you think the impact is of the complexity of the ecosystem? And I say that because when you when you look at earlier things like you probably remember the LAMP stack right in the early web days and it made it really easy to build because we're like we're all using you know Linux Apache you know PHP and my SQL and then you know started to fragment or whatever. I don't feel like the AI ecosystem the infrastructure pieces are moving that way. I feel like it's getting more fragmented people using all kinds of things. Some things don't work with other things. I think it makes it a little hard to build like you know where do you get your models? Do you go to model provider? Do you go to open router? You host them on your own together AI like there's and that's just like one layer of the stack and you have all these layers of the stack and they're all changing and innovating a lot like I feel like it's really hard to do like what do you think the impact has on on people building companies?

Kenneth Auchenberg: Yeah. Yeah, you know, as a as an infrastructure investor, I'm feeling this every day in terms of like my job is working with founders and making making bets on like what do we think is going to be the doable kind of pieces as we kind of in this AI buildout. And I think the reality is that like you know we're coming from a world where let's be fair the past 20 years we've had a somewhat stable kind of stack. We kind of knew roughly you need to have a UI so there's some there's some storage some compute and now we're in this net new era where we don't know what the app layer is going to look like and it's going to is right up in the air. We don't know how we're building agents. You know up until like what is it like 12 months ago an agent was workflow automation. Yeah. So we in this very interesting kind of area where the bar is also moving incredibly fast and you know when I talk to like people at the big labs like most of the labs don't have product road maps filled out in 3 months because there's going to be a major state-of-art model dropping in 6 to 8 weeks.

Yeah. So we don't know what kind of capabilities that are emerging. So we are in this very very fluid kind of period and the bar is kind of changing constantly and what capabilities do we have available but also how we're going to architect these systems. I think if you if you if if you look at like the value right now, the most hot thing you can do is a microVM based compute company building sandboxes and probably a knowledge graph, right? That's kind of like the current vibe that's going to look different in two months. So, so we are definitely in a very interesting time. But I actually think like you know the counter to that is I actually think technical leadership is much more important because you need to navigate this changing landscape even faster where in the old era you could actually get pretty far far by not having technical expertise on your team because you kind of knew what the broad strokes were.

Now the landscape is changing so fast that you actually need to have that expertise as a as a founder on your team to make sure you're kind of staying on top of the capability unlock curve. So it it's very exciting but also very like you know I think every day all of us are feeling that the acceleration and something something has materally changed now right.

Ryan Eppley: Yeah, I think you said something earlier that we really didn't hit on and maybe it's like my love for gaming as well, but just recently there was an experimental MCP for Unreal Engine now. Yeah. Released. And I had a lot of fun with it already. I think game engines are the cornerstone of where a lot of infrastructure is going to be headed. As every domain starts realizing, they're going to have to figure out how to become an engine for some type of agent to go touch. I think if you have any game developer on your team, you should promote them to technical leadership immediately and be like, "Hey, remember you built those game engines a couple years ago and then you joined our company because we gave you a pay raise and things like that. What you learn over there. You know, I think game engine developers are going to be probably some of the most hotly contested talent over the next years.

Kenneth Auchenberg: I definitely think that that's I agree with that. I think there's going to be a lot of retooling how we think about software and what kind of components and how we think about the boundaries of software is that like you kind of just do like a full stack Rails app and then call it the day. But to your point about like all the change, I actually think like if you're having well- definfined interfaces for what you expose, I actually think a lot of that responsibility is going to be pushed to kind of the more generative kind of layer in the sense that I'm actually glad I don't have to think about like what a bar chart for MR metric is going to look like because just going to generate that for me, but I can focus on providing the data. So I think that there's pieces here that we can kind of offload to the agents so that they can figure out where we can kind of focus and so I actually think that's a net good.

But at the same time I do think that there's going to be a lot of experimentation and you know I'm an infrastructure investor. I care about like the primitives. I'm not an app layer investor but I would be a little scared if I was an app layer investor where I'm investing in this this company like you know h am I using Lora hardly for legal or so clude going to eat that market is it just going to be the data provider et what's what's the future of like how you own a workflow how do you add value how you deliver a valuable outcome what are the means and how you kind of get there

Ryan Eppley: yeah we didn't have it on our our list of topics to talk on for the news, but it it was published maybe three or four weeks ago about can investors even invest at this time without it just going to the labs because the labs are just going to eat everything. It feels like right now it it's hard to see them not bulldozing their way to monopolies for the app layer, right? They'll just see what you're using and be like, "Oh, you're yo 10 millionaire, 20 millionaire. Actually, we'll build that product, too. That that's a tasty market. You're right. Thank you for being the frontier, but we will go settle that land now.

Rob May: I mean, this came out during the cursor acquisition discussions, right? Which is Anthropic originally told cursor. Yeah. And it was a research project and then launched Claude Code.

Kenneth Auchenberg: So, it's interesting like like I think I have a two-part kind of answer to that. I think just on on the curse on the codegen kind of piece, I think a core thesis of mine is that every agent will be a coding agent. I think code will kind of be the substrate of how we get knowledge work done. So in that lens, I think that's kind of what all the big labs realize that code and be really good at code is really the unlock for all these other things they want to do. So it kind of makes sense that they they they did that. Also, it's interesting kind of seeing OpenAI pivoting from being going down the consumer path to now doubling down on B2B and Codex because they kind of realizing the same thing. And in terms of like the app layer thing, I think putting on my investor hat, you know, I think if you play that, if if you're asking the same question like 10 years ago, then the question would be why wouldn't Google do this? And then you go further back. Why wouldn't Amazon? Why wouldn't Microsoft?

Rob May: Yahoo.

Ryan Eppley: Maybe let's get all the way back.

Kenneth Auchenberg: Yeah. Yeah. But yeah, you're right. But I like I kind of think it's kind of like the lazy straw man here in terms of like I think one of some of the more interesting questions is that is this at the top of the priority list for these companies? Because these companies have infinite money, they have infinite talent, but is it really a priority for them to go after like a $50 million kind of market? And are they going to continue to be interested in that? I think that's kind of like the classic of like if you look at like killed by google.com and look at the Google graveyard of all the things that they tried. I think we kind of seen how history kind of been been playing out. But I definitely think that there's a world we're moving into where there's going to be much more software in the world. There's going to be much more room for bespoke vertical specific software because it's going to be much cheaper to kind of build.

And I think the real question is really about like why would you buy something if you can generate it yourself and the continue and the cost kind of continues to drop. I still think that there's a piece that is kind of like in the whole SaaS apocalypse kind of narrative is that turns out most people don't want to maintain They don't want to be responsible for it. [clears throat] Mhm. And I think that's a big part of like the narrative that we haven't really talked about or maybe we've forgotten is that the reason for why I would like to buy a issue tracking system from Linear is because turns out they're really good at doing that. They have really good taste and I trust them being able to handle that problem for me. And I don't think that has changed. I don't think that has changed even with code going to zero. And that's why I don't think it not everyone will kind of vibe code or generate their own CM system etc. They don't want to do like their own workday even the work day is not a great product but they will probably just buy it right

Rob May: I mean so I can tell two two stories about this so first of all when when we had this like drop of claude code like Thanksgiving time last year where it was just a it could solve the problems of your gaps in knowledge when I'm like I don't know how to do that piece it's like here's how you do it I can do it for you I vi I spent the whole first half of this year like vibe coding the out of everything I could think of and I came to the same realization where it's like you know I would put do it I'd put something out there and I would go to use it, I'd be like, ah, like this I got to do a package update because or whatever. And then I was like, I don't want to maintain this, right? Like it's not worth it. And so I think that's yeah, I so I think this is going to happen more and more and more because I agree with you. I think I mean I do think SaaS valuations are going to come down, but I think people don't want to don't want to maintain and do these things. It's a lot of work.

Ryan Eppley: I mean on the infrastructure side is that also inclusive of you know like sandbox AQ and like LQM have you like gone into that like path of thinking as a part of your thesis as well

Kenneth Auchenberg: in terms of what specifically of like

Ryan Eppley: so like their large quantitative model right you can now access it via claude are you familiar with sandbox AQ

Kenneth Auchenberg: yeah yeah yeah

Ryan Eppley: you know they're spin out of Google as well right raised tons of money I think billions of capital raise and they've built their own foundation model, right? So, but they're not like having people access it directly as maybe I think their future roadmap now it's like again back to the engine, right? Are you thinking about other companies have to think of themselves as foundation models to even be able to play in these games anymore? Is that where your head's going?

Kenneth Auchenberg: Yeah. You know, like I like for example, we working with a with a company called Extend here in New York. I know as an investor I shouldn't be and just talk my book but I think this is this is a good example. Extend is doing document processing and you can kind of make the argument that maybe I can just use a foundation model to pass my PDF do my OCR and kind of get that done. Turns out if you want to go from like 60% accuracy to 99% that's a lot of work and I think most kind of problems looks looks like that when you really dig in the surface how you pass a PDF how you do the right conversions etc. So they're building agent first infrastructure and doing document processing and they're powering all the agents when you like use the ramp agent to do your your your reimbursements is using extend. So I think that there's going to be a lot more infrastructure and if you think about like that that in the old day it might have been a full stack app now it's an infrastructure kind of component.

So I think a lot of software will continue to exist in those kind of worlds and these kind of flavors. And I think that to what I was trying to say before, I think that there's going to be a lot more bespoke software that solves very niche specific problems because there's going to be a lot more software in the world. What used to be spreadsheets is now going to be a bespoke piece of software for the dentist or the vet that they couldn't have before and that's going to require infrastructure. So I also think that we're going to see kind of net new kind of pieces emerge that might not be a ventureback business but it's probably going to be a really good business because it solves tangible kind of problems.

Rob May: Yeah. So, I want to transition from from talking about some of this stuff to some of the some of the news we have for this week because we have about 10 minutes left. And the I think this article that you sent is fascinating before the show, which is Matthew Prince from Cloudflare tweeted that AI agentic traffic grew 7,851% last year in the past year. And for the first time, bots outnumber humans on the open web. And now when they say bots, they don't mean like web scrapers and automated programs. Those have outnumbered humans for a while. They mean things going out and doing tasks, right? Doing research, finding things and everything else. And so it's and Prince apparently in the tweet in the tweet was like, I thought this was going to happen into 2027 and then early 2027 and then it's happened now, [laughter] you know, so it's clearly accelerating. What do you think this means? What does it mean for people like us building things? What does it mean for the internet for this to happen?

Ryan Eppley: It means browser base is crushing. [laughter] you know, like making making your website also very easy to digest is incredibly important. Now, I've started to seeing a lot of companies popping up like using that as like a hey, we will make your your website super super great to now just not only search but even navigate, right? You know, so again, it's your whole agent first thesis, but yeah, I think you're the one that sent that article in. Like walk us through like what does that mean in your eyes?

Kenneth Auchenberg: I mean like we can kind of go all galaxy brain here depending on what kind of like level we want to apply. I like I think my lens is that I think it's fascinating that we're kind of one and a half year ahead of like the adoption curves. To me that that's really interesting. H but it's also kind of like it's the end of the human internet like the internet that kind of I grew up or we grew up with was really built with the humans as the main consumer and you know I remember like seeing hamster dams the first time and those kind of websites those like world yeah animated gifts like those those kind of things and now we actually building the incident that's emerging very very fast is really built for for the agents and I think like there's going to be very interesting kind of like second order effects of what that is

I think like a second part of my thesis is agent first infra but also sovereign infrastructure and I actually think the sovereign piece is very due top of mind for for a lot of people but I actually think it's important to kind of like factor that in is that the internet we built for humans in the past 25 years has really been in a globalized world it was all about having the internet we all like were nice to each other etc and now in a different geopolitical reality so even the agents are going to grow up and use an internet that is going to be the American internet, the Chinese internet, the European internet. And so I think that's kind of interesting that maybe agents will have access to less data because the internet is going to be more balconized. I think that's one interesting piece. I think the other piece is also the regulatory kind of lack like you know we in technology we always push technology forward and then technology kind of catches up.

If you every time I go back to Europe I look at goddamn cookie banners and it's depressing. But the reality is a lot of our compliance and privacy and how we think actually is built for human first world. GDPR doesn't work in an agent first world. It assumes there's a human intent. Even we think about like socks 2 compliance that's assume that a human has written the code and human has reviewed the code. Now we right now we're in a world where agents are generating the code and agents are are reviewing the code and just deploy it. So I think like even the regulatory kind of lag is interesting but also what does it mean if there's this kind of parallel internet where agents are kind of like talking to each other and doing things and then there's still like this kind of like bespoke human internet where we still talk to each other etc.

I think it's very interesting that we're kind of seeing these net new rails emerging. And I'm not sure we kind of like have fully understood what this transition means, but to me like this is the very beginning. I think you know, was it last week or the week before that that that Cloudflare saw that data point? Yeah. And it's kind of like we're entering a new kind of era of how to think about building for the internet.

Rob May: Yeah. I have a funny hamster dance story which is [laughter] my my So I came out of school during the web 1.0 bubble, right? So, I got, you know, if you had a degree in a pulse that was technical, like I had I had a friend that had a 26 and he got like a great job, right, which you would would never happen today, right? And the job I took with Harris Corporation in Melbourne, Florida. I was a computer chip designer and I go down there and like my third day there, this guy quit because he got a 40% raise somewhere else cuz people were throwing money, everything. So he he cranked he went to hamster dance on his computer, cranked up the volume, locked the door and shut it and locked the key inside. So like for like two hours it's all it was just blaring and I thought, man, what what is going on here? And maintenance had to come and get it in and shut it down. Wow.

Kenneth Auchenberg: That's a that's a good.com kind of like Yeah, that's a story. [laughter]

Rob May: What else we got on the top? So I want to talk about what do what do you guys think about? So Amazon has decided to sell its training chip Trainium chips to other people. So if you know if you don't know this, Amazon makes a chip that you can use to train AI models similar to what you can do with the GPU from Nvidia and Nvidia has been the king of all this like the I mean the runup in that stock over the last five six years has been utterly incredible. They're owning this market. And so I'm curious what you guys think. We've touched on this before if this might happen. It looks like it's starting to happen now. And so I guess two questions. Like number one, if you own Nvidia stock, do you care? Like is this a market where you're like, "Dude, it's so big. Nvidia's supply constrained, like there might as well be other people." does this make you pro Amazon? Does it, you know, provide optionality to people like us developing? Like what what what are the impacts that this has when you look out 12 or 18 months?

Ryan Eppley: As someone that loves hardware, I'm just excited to see more hardware entering the market, right? I think you know summer you know summer well but you know she's nerding out about this moment of GPUs versus CPUs and CPUs seem to be like where the frontier is actually at now but obviously Nvidia and GPUs have dominated for so long so I'm I'm excited to see different architectures really start to come into their own even like FPGA on the edge now right like people are really optimizing edge models right I just went to the conference in London for edge AI and a lot of that stuff was not being run on GPUs of course right heavy a lot of thermals things like that you know so what do you think about Amazon now selling its chips as like a do you think it's a revenue operation or is it more for them to optimize it for their own data centers

Kenneth Auchenberg: like like I you know I think my my view is first of all I'm I'm not a hardware guy so this is my my uninformed kind of take I think we're so early in the kind of the compute journey when it comes to a gentic systems is that it's probably what like 1% of the world not even half a percent is using agents running 24/7 and we so early in the kind of the compute spend and the token spend

and I think I love this metaphor of like you know tech up until this year has really been like the day shift is that we go to the office at 9:00 and then we go home at 6:00 and then work kind of stops and with agents, we kind of introduce the night shift. Yeah. Similar to what happened in manufacturing. So now we can kind of work 247. I spoke to our founders like my job in the day is to plan all the work for the agents when I come back. And that's kind of like how how he's kind of telling his team to kind of work. But I think the true unlock is really not like I don't have a better metaphor, but like is it like the parallel shift or what you call it? Because the key thing is the decoupling between like work and time that you can spin up parallel kind of workloads.

You can have a whole like factory a whole army working for you. I think that's going to be the true kind of unlock but if people are just running you know open claw on a Mac mini doing some some code kind of queries etc and we are compute constraint now to me that's like we are so early in this kind of compute spend journey we are on both most of because our models are super inefficient etc. But it kind of makes sense that that Amazon would be like, you know, we have these customized chips, we have our own GPUs, why not sell them? And I also think it's more interesting when you look at like I think based on what I understand with with Nvidia and CUDA and like the abstractions is that C like Nvidia really had a monopoly with CUDA.

But what has happened in the past year is that you are seeing OpenAI you're seeing Anthropic you're seeing Google Google training on on training and doing infants and TPUs you're seeing Anthropic doing even even rock from AMD got better with their CUDA version but went public right and cerebras went public so we are really in a world where like we actually have like very diverse kind of compute h that is being used by a lot of different info info providers and Nvidia actually don't have the same kind of like monopoly at least on the abstraction layer anymore because everyone has kind of been forced to diversify because they need the compute. So, so it to me it's just like a reflection of where we are in the cycle that of course if you have good GPUs or CPUs you want to sell them, right?

Rob May: Yeah. Part of what might have made them make that jump is like there's getting back to the coup argument. Cud is the most mature software stack and we at Neurometric about a year ago tried to use some Trainium chips and it was like like the you know but they've they've probably invested in it and gotten it up to par to the point where you know I think they also signed a big deal with meta a couple of months ago. I may we covered it on the show and so I think they're the software stack's probably improving. It's becoming you know debugged and stable and probably time to go 100% right.

Ryan Eppley: It's just like, you know, if you go to the car lot and the Ferrari is not there, but you still need the car, you'll still get the Honda even if it doesn't have the libraries and all the optimizations built out cuz you still need to drive something today, right?

Kenneth Auchenberg: And I think like talking to founders working on like compute sandboxes and kind of like like to your point about CPUs is that some of the stories I hear and how people out hunting for CPUs. I was grabbing dinner with the founder the other day and he was like as we were sitting there he was getting access to some more GPUs and he took them on the spot h but also like how people like if I can get like a thousand CPUs here I'll take them and then I buy fiber interconnect between the data centers to get latency down people are out there hunting in the markets to get capacity and we are so constrained and when I talk to like people in the neo clouds and even building like in the kind of the software layer people are on so diverse compute environments and all the customers is like I need more so it's just a very kind of fascinating time we're in this kind of buildout

Rob May: yeah and the needs are accelerating right we're using more and more and more and so that leads into I think the third article I want to talk about which is the New York Times wrote this article on token minimizing and so I'll tell you a story because my company was featured in the article and the way that came about was when all this token maxing started because I am a person who likes to push people's buttons I went out and bought domain tokenmmining.com. I [laughter] was like, "Someday I'll do a site about using the minimum amount of tokens." And when it actually started to get a big thing, I was like, "We should actually launch this." So, we wrote a manifesto, made a you know, email newsletter and like Reddit forum and it's like, "Hey, here you go. This is this is about token." And the New York Times found it and wrote, I mean, we just launched this like a week ago.

New York Times wanted to write about it. But my question for you guys is now that the world has become aware of like wow we don't need frontier tokens for email summarization and customer sentiment analysis and these things and so we're going to see a fragmentation that's going to you know make it more efficient but I don't know to the chip point I don't know if that efficiency is going to outweigh the 10 20 50x yearly jump in AI usage. So like what how does this play out in terms of you know moving from tok moving away from token maxing to more efficient AI infrastructure versus the fact that we're still wildly supply constrained.

Ryan Eppley: I don't know I just thinking through this on the spot right now. I just feel like we as humans have gotten good at organizing other humans and it's just going to be like we think about our payroll. It's like payroll optimization. I might for certain tasks need to hire someone that has a certain salary and I give them a certain title. And then you give them a certain sliver of the company payroll to say, "Hey, you're going to manage this and I need you to pick what salaries and what titles they're going to need." And I think that's just sub agents. It's like managerial agents are now going to have sub aent kind of tokens and access to certain models that are optimized for those workflows. And you know, I think we could go down the I like your term of galaxy brain, right? Like I think we're going to get down to this moment where we're going to give a ton of autonomy to the best models to go figure out what to do underneath them and we're just going to check in with the top models about everything underneath them.

Kenneth Auchenberg: Yeah. Yeah, I agree. But also to me that's like we're so early in kind of the token spend. But it's kind of interesting how we gone from token maxing to now token what token mining. That we kind of gone there in the matter of like what 2 months 8 weeks 6 weeks. And I definitely think like like you know I remember the days where I could drive like in a lift across San Francisco for like $4 and we were kind of like in the golden era of Uber and Lyft. In many ways, we're kind of in the same golden era of cheap tokens. Yeah. In the sense that, you know, when you think when you look at some of the margins from some of the companies, I think it was cursor that had like negative margins on the tokens that that they were selling etc. We are definitely in that kind of era. But I also think to your point about edge inference etc.

Is that maybe we shouldn't need like a PhD level model to do like fix these typos kind of thing. So, I definitely think as kind of the token spend kind of compounds, continues to grow, I do think there's going to be a more diverse kind of set of tools we're going to use to solve those problems. But I think one thing that I find really interesting is that so far people seem to be willing to always pay for the latest intelligence. Yeah. And they're willing to pay that premium. Like if when there's a new OPUS dropping or new GPT, everyone seems to migrate very very fast. And I think there's an element of like just competition here is that if your competition have more intelligence than you, you're going to fall behind. And I also think even when we run agentic kind of workloads, the one argument is that if I run an agentic workload can run at night, I it can take longer.

But the reality is kind of like a arms race is that if my my com competition can run faster, I want access to the same kind of thing. But I definitely think it's just so early and it kind of feels like you know using a rocket launcher with everything where we should use much more bespoke vertical specific things. So I'm actually very bullish on what's going to happen to edge inference particular when it comes to the industrials etc. And I think there's a lot of workloads we're going to move to that but I think there will always be a massive need for state-of-the-art really really great intelligence and that's going to be cloud inference.

Ryan Eppley: There is unlimited need for intelligence.

Kenneth Auchenberg: Yes. Unlimited need.

Ryan Eppley: Yeah, you know, but now it's at some point to your point, the subsidized period is going to come to a [clears throat] stop at some point.

Rob May: IPO [laughter] at some point.

Kenneth Auchenberg: I mean, I think what's very interesting like, you know, like I think Uber CTO has been out there in the media, but I think like, you know, I think they now said like is it $1,200 or $1,500 per engineer per month as that kind of engineering token budget. And you know, I actually think that's good news for developer tools. That's actually really good AR per se. Yeah. That you can charge if you're selling tokens. And the fact that we've gone from like capping a junior engineer being 250 a month for max subscription to now being 1,200 is actually a really good kind of t expansion there in the market. But it kind of shows where the demand is and where the value is kind of being perceived. But I also work with other founders that you know if my engineers are not burning a year salary in tokens, they're not pushing hard enough. Yeah. So, there's definitely different kind of mindsets out there as you kind of tap into like should I just have agents running 24/7 or should I go hire a team? I think there's different philosophies that is playing out, but it's just so early.

Rob May: I mean, I can tell you what we see in the market in our customer base is it's actually divided along a line that you wouldn't expect. Interesting, right? Because, you know, because people are always like, oh, like what what industry or what title, whatever. No, it's is your token spend opex or COGS? If it's OPEX, you kind of don't care. You want Frontier. If it's COGS, you have to show margin improvement. And so, you're looking at SLMs, you're looking at cheaper models, you're looking at open source. It's a it's a hard thing from a marketing perspective to scream for, of course. But that's but that's the divide in the market that we see.

Kenneth Auchenberg: I mean, it's interesting, you know, like for the first time in God, how many years we actually looking at a significant cost to produce software. Yeah. And like in the old days, it was kind of like the intangible economy. You produce it once and you can sell infinite kind of copies.

Rob May: Marginal cost matters again.

Ryan Eppley: And manufacturing companies.

Kenneth Auchenberg: Exactly. Exactly. We are in kind of a new era as long as there's not like a revolutionary kind of like unlock in model architectures or anything like that, right?

Ryan Eppley: Well, as we wrap up, we have to ask, right? You see the $60 billion exit of Cursor and you were instrumental from day one in the backbone of what they've used to get to that 60 billion. Do you ever feel like Microsoft or even you personally would do you ever feel like maybe that maybe you should have had some type of tangible takeaway from that or were you just proud to see your baby in your product developer role see something come to fruition because it's a massive exit?

Kenneth Auchenberg: Yeah, like you know like I think the cursor team has done it incredibly well but I also think if you look at what cursor has been building they never built a code editor they built an agent and really built like an agentic software engineer and we were not building that back then like you know our teams back then we're building VS Code we had a team next to us that was doing the very early versions of co-pilot

so I would be like if I were the one that invented co-pilot I would be in a different kind of position looking looking at at the market but I think you cursor for them it was always about them using a code editor to reach developers where they are and help change the perception and I think it's kind of ironic that you know even cursor is now getting acquired as an amazing outcome but most developers are probably using clo or Codex they're not sitting in that kind of traditional kind of ID service and that you know for me it's a kind of it's a decade ago I worked on that time really flies and I'm gl I'm actually glad that we open sourced a lot of technology kind of enabled this foundation but hey I'm I should ask myself why didn't I go and build build that I think that's more like the real question

Ryan Eppley: but that's the lore for for our viewers you know you're so instrumental into this bridge we'll call it right this bridge from VSS code was the backbone of cursor which is now the backbone of how people have developed this agentic future so no we really loved having you on the show this is really fun

Rob May: yeah thanks for coming on Kenneth

Kenneth Auchenberg: thanks for having me

Rob May: and if people want to reach out to you or pitch you on something what's the best way to do that

Kenneth Auchenberg: h kenneth.io IO is my website and also Auchenberg, my my weird German name. You can add that in the subtitles below or whatever. At Auchenberg on Twitter. So I'm my DM and email is right open.

Rob May: All right. And everybody out there watching, thank you for watching the show. Please give us a like, please subscribe, send us guests, send us ideas, send us sponsors, subscribe to our newsletter, and we'll see you next week. Close the door, please. The next