After SaaS

0:00Hey Reid, great to be here today. Always fun. So this week, we're actually going to talk about a theme across all of our questions today. And we're going to talk about software, or more specifically, whether or not SaaS, software as a service, is dead. Mark Cuban recently argued that software is dead because everyone is going to customize everything they do to their own unique utilization. Just a few weeks ago, a tweet about Cloud Code drove SaaS stocks down over 5%. No one is arguing that coded products are dead, but some people do believe that software may stop showing up as a fixed product and instead become something more adaptive, generated, and specific to each individual company's use case.

0:43I think some people are really excited about that. They've been annoyed that they're getting a one-size-fits-all software, so this new era of customization is pretty exciting. So when people say SaaS is dead, what does that mean to you, and do you agree? I think there's one sense in which I agree. And then there's a typical inference that people make where I disagree. So the thing where I agree is that the exact model for the last 20 years of SaaS companies is no longer sustainable. And the way that model was is that you build out a SaaS category, you make it detailed and ornate enough that it's like, you know, table stakes to play in it as a billion dollars of software development.

1:23And then you go out and you've acquired a bunch of the market and you're growing as you're going. And the problem is that closes the door to a startup competitor. You can charge pretty good margins on it because there's relatively little competition. You know, you need to have this like it's like, you know, kind of turns into a natural oligopoly. by the way and you know people are like oh is that terrible it's like well that's what happens with the auto industry and you know a bunch of other things it's like turning into a competitive oligopoly you know still ends up you know generating a pretty good consumer and societal you know output and that's kind of where where a lot of like sass categories were kind of heading towards you know the reason why you couldn't you you would pay the margins e.g.

2:06pay for the software It's not just like pay for it as a unit, but like at a 40 % margin or 50 % margin as a way of doing this is because that was the only thing that made enough of your requirements and was stable in terms of going into the future. And so the reason why people say, well, that business model is now gone is because they go, well, they're imagining that you go to Cloud Code, Microsoft Copilot, OpenAI Codex, et cetera, and you go, make me an HR system. Make me a CRM system. Make me an accounts payable system. And it just does it, right? Which is foolishness. And even foolishness as the models get substantially better.

2:49Now they say, well, but, you know, the AI will get so good, it'll figure out all the different details and what are the things that need to be running for me and, you know, da-da-da-da. And you're like, yeah, there's a bunch of stuff it does there, but I'm not trying to make my company the expert at doing HR systems or even HR systems or finance systems or accounting systems or CRM systems, you know, all the SaaS security systems for me. What happens is I think, and this is the implication, is like, oh, all these SaaS companies go away. There's no longer sale of software. I think that's incorrect.

3:23But the sale of software now has to be AI, like you have to have that AI generativity that you get from these coding agents as part of what you're doing. and um and so then when it when it gets to what is the place where there's an interesting you know kind of economic model for this it has to be within that loop it isn't just like oh it takes a billion dollars of human software time to build this so competitors can't enter it's got to be the okay what's the things that that make it work maybe it's an understanding of all the requirements, maybe it's the, you know, that the AI systems have been really tuned to work well, like when you buy a CRM system, and CRM system comes with a set of really well constructed AI, you know, kind of agents that help tune iteratively your CRM system for you.

4:20And that there's a new way that this system now works, both having powerful sets of libraries and tools back in, because by the way, part of the way the coding agents work is they figure out, you know, which systems build. And I think there will be a lot of like deep software libraries and different things, including how do you make this a whole lot better for doing it? And then there will be new lacuna. Now, is it a, like for example, one of the more subtle points when you begin to get into this, you know, is kind of like, is it a SAS model or is it a different kind of model? right like it is a seat model but so it's so i think sass is dead insofar as i if i came along to you and said i'm gonna build a seat driven new crm piece of software and everyone's gonna pay me for seats on this you're like okay you know not so much um but if you said hey i think the crm industry is being disrupted and here's a new economic model, a new way of providing it.

5:23And the way that customers, just like when you went from CRM on-prem to CRM in the cloud, now there's a new thing. Well, that's a new thing. Yeah. I mean, it's like people say, like Anthropic is still using Workday. Like there are going to be, you want to focus on your core competency. You don't want to be distracted by all those other things. And so amid this, obviously, we're going to have to see new business models, new market entrants. How does the new landscape affect the engineers and the builders? Does AI make them more powerful? Does it change the kind of technical talent that is needed within these SaaS companies?

6:00How should they think about it? 26 is beginning of the year where we're beginning to see both the quality of the models that kind of got there in November, December, and also the stampede to using them where the, I think the fastest changeover in doing a job we're going to see is in software development because they're technically sophisticated already. They know these tools, like as you and I know, when we go to other people and say, Hey, you should be using cloud code. I don't even know how to start, right? Like for non-software engineers, even though you can, the thing that I think will happen already now and is happening from software builders is that they're starting to use it intensely.

6:43Now, here's an interesting current milestone. Software groups that are starting from scratch are seeing a huge acceleration from current AI tools. Software groups that are going, I have 500 million, you know, or even 100, but, you know, like, let's call it real, 500 million code base, and I'm refactoring. That's actually much slower. Like AI does not like just plug it in, do it, et cetera. Like you go here, I've got this bug in this software, fix it. And put as much compute behind it, it still doesn't kind of really get there. It isn't to say that it won't. It's to say that this is a current milestone in terms of how this is operating.

7:23And how it gets there is interesting. Like, for example, part of what this is is like, okay, so construction from scratch, you know, if you're not doing an orchestration of a whole bunch of AI agents, you're like kind of like it's like you're you're using your punch cards but if you're doing the kind of remodel refactor yeah you're still struggling your way through it you're figuring out what are the what are the right deployments and where is the ai building to now one of the things that'll get to be interesting in this is what does it mean because part of what happens with code is not just you build it but you maintain it well what's ai going to look like for that?

7:57Are we going to have a bunch of like, you know, what's the role between, you know, kind of human and AIs and doing that? And, you know, people tend to think, oh, because it's going so fast and exponentiating, eventually the only role for humans will be metacognition and awareness and kind of orchestration. And that's, by the way, possible. It's also possible that in, you know kind of planning this larger context base that there is also you know kind of various forms of human ingenuity in it and also in goal setting and like what is the real understanding of what goal you should be trying to do like if you're actually doing this stuff the people who I know who are like living in this go frequently they have to go unblock it because it doesn't understand hey if you shifted this goal a little bit and you called this library and did this all of a sudden you'd be doing a lot better and it can get caught in that.

8:54And that's with, you know, millions of dollars of compute going into the tokens, kind of realizing it. And so there's still this kind of role for that, but that's the reason why to engage. Engage primarily as a director and orchestrator. And so I think that's the question that people have. We talked about this, that at the top companies or the companies who sort of started post AI, their developers aren't writing code. It's the AI that is writing 70%, 80%, 90 % of the code that's coming out of these places. And so their question is like, what are you talking about? Then what do engineers do? Then what do they do all day?

9:27And so you were talking about goal setting. Is it also deciding what should be built? Is there a trust component? When to actually ship? What are the other things that you think are going to become more important for software developers now that they're not spending most of their time writing code. So here's kind of like the detail. It's almost like what's happened is a lot of code before was management of this very specific set of like detailed tools, like how to call the API, how it worked, you know, writing the code and the right syntax and language and, you know, kind of blocking all that together.

10:01And that stuff, you know, AI is already great at. There's ways in which it's still not like fully in the kind of human code or capability a la large code maintenance. But that's the reason why people go to, well, actually, in fact, it's strategy, it's taste, it's metacognition. And by the way, now that you have these tools, like, well, you say, well, previously, if I had to write a connector to a different kind of computer system, I'd have to research, understand a whole lot, that would take me weeks. Now that's probably like an hour. And it's an hour with a, you know, like, okay, I set a bunch of, of, of long, you know, kind of co-pilot codecs, you know, clawed tasks going, it comes back.

10:52I, you know, I look at certain things, the trust thing that you mentioned, because, you know, you know, it's like trust, but verify. And by the way, that's how human coding works too that's part of the reason why we have like agile programming pair programming etc you know it's kinds of ways of doing this because they understanding that verification is an important thing it's an important thing even before we had ai and so you know and maybe with ai we'll get to you know more cracking of codes of a verification coding like it's verified to work, which has been too expensive for most human systems in terms of doing it.

11:34But maybe there's ways of doing that with AI. So that's why I'm saying that the surface opens up. So yes, I think there will be the kind of metacognition, the strategy, the understanding goals, the trust in what's happening. But also like, for example, when I started with a connector, it's like well maybe what we should do is figure out how to have you know not just the connector but what are the other things that should actually really go with the connector should there be a monitoring system should there be a fail system should there be a modularity to doing multiple connectors like like when you do this because the the cost of doing certain kinds of tasks gone so far down it now opens up the space of what you're doing so not just do i build a connector but i also then say, okay, we're going to have a monitor.

12:22It's going to play in, oh, we have a whole monitoring system. This is how the monitor is. Oh, well, this is the way that we should be actually checking safety and cybersecurity in this. This is why the Jevons paradox is when there's infinite demand for stuff, it massively expands. And in software, there's tons of areas where there's infinite demand. Now, back to your original thing, it's like, well, what should you be doing? It's like, well, you should be at the forefront of using and deploying and leveraging these tools because that's how software is going to be constructed. It's like, if I think I'm going to be a modern mechanic for cars and I show up and I've got my wrench, you know, like, okay, you know, like, no, no, no.

13:04You are not ready for the 21st century. Yes. So it's like, you have to be doing that. And, and look, it'll be a little vertigo. It'll be dizzying. It'll be changing rapidly. People wish you could just say, oh, it's just like, well, you just have to learn to use, you know, Rust now. He's like, no, no, no. It's, you know, or, you know, I'll go old school for some of the older people on the thing. Java, you know. And so it's like, okay, it's a fast moving and iterative tool set. But by the way, that's what's going to make it fun too. Yeah, absolutely. So talk about something that's a little less fun.

13:39It feels like every week we're seeing a new headline about layoffs that are coming. And whether it's because of AI or not, We saw recently that Atlassian said they were cutting 10 % of their workforce, and a lot of it was due to AI. We also saw reports that Meta may cut 20 % of their workforce. Obviously, some of this is AI. Some of this is broader political and geopolitical things that are happening in the world right now. Obviously, we don't want to see job loss. But the other side of that is people talking about software moats. You said yourself that one of the reasons in the past is you had barrier to entry.

14:10People couldn't compete because you had this big company that was so far along that a startup entrant couldn't compete. When we think about moats, you've obviously talked about network effects for a very long time as one of the best moats around. In the new AI world, are these the same? Are they different? Do the moats disappear because you can have a two-person team taking on an incumbent? Where do you see those sort of lasting and what are the new moats post-AI? Well, the new moats are essentially in company as a theory of the game investing in or co-founding like Manus and other things. So it's kind of like a question of theory of the game in terms of how you're constructing those.

14:55And that's not proven yet. I do think it will be network effects. I do think it'll be your position with a significant set of customers, whether those are consumers or enterprises, because that tends to have, there's various ways you can have that as a compounding effect. And that gets, you know, all the way back to why investing in models is important. Or not starting models, apps, but potentially models too. But apps was what I was thinking about there. And I do think that the notion of the fact that you have to kind of retool your workforces for being kind of AI, kind of native AI heavy, et cetera, is going to be extremely important.

15:49And I think when you begin to get to this question is you begin to go, well, what's the cost of tokens is one way of looking at this. What's the cost of tokens produced by humans? What's the cost of tokens produced by AI? What's the cost of tokens produced by the combination of them? And of course, it's the right tokens, right? Like one of the things in this kind of usual token economic discussion is, well, actually, in fact, if you're just producing tokens, well, the monkeys can create Shakespeare. but it's actually in fact it's the right tokens and it's the the right edge the higher value and and you know and that kind of thing and that's part of you know what we're going into now that token production is do you have the right unique data sources for producing it um i do think token production will be part of it so like as opposed to like the kind of classic SaaS model of, you know, per seat, I think there will be more of a compute consumption, you know, kind of as a way of doing this.

16:53Now, how we're going to work that out exactly, because it, you know, you need to have, it needs to be, it's like predictable, right? So for example, like part of the reason why you had the previous seat model with SaaS was because it's like, well, well, okay, I know how to predictably, you know, do my cost line to model it into my business and my revenue line and my margin line and how to be investing in the future and so forth. Well, how am I going to do that if you're charging me on tokens? Like, and by the way, the answer is we will figure out ways, right? You know, a classic old school way is, well, you know, you get a baseline token budget for a subscription and then you get pricing depending on you know how much you're like almost like utility or energy like you're you're you're pre-paying for a bunch of tokens or a bunch of per unit you know hour unit unit uh day unit week unit month etc because you know part of token production comes from the compute infrastructure which is you know chips and energy and in data and all the rest and so all that comes anyway so i think there'll be something new around tokens and there'll be an interesting rationalization relative to, you know, kind of human capital and compute capital in this.

18:14And I think there'll be a kind of an intersection point between the two of them. And I think we'll begun understanding what compute capital looks like, you know, in addition to human capital, in addition to, you know, kind of energy capital and, you know other kinds of things as as as this is playing out now where does this get to you know kind of where are the key places where there is you know good operating margin and all the rest and and i think that's what all of us investors are trying to figure out with our investments right now um i think almost always a bunch of the model the the the moats of the past will stay in certain ways, but the game gets really shifted with a platform shift like AI.

19:00And some of those modes will matter a lot more than others, and there will be new ones. Awesome. Reed, thanks so much. Appreciate it. Always a pleasure. Possible is produced by Pallet Media. It's hosted by Ari Finger and me, Reed Hoffman. Our showrunner is Sean Young. Possible is produced by Tanasi Delos, Katie Sanders, Spencer Strassmore, Imozu, Trent Barbosa, and Tafadzwa Niemorundwe. Special thanks to Surya Yalamanchili, Sayida Sepieva, Ian Alice, Greg Beato, Parth Patil, and Ben Rallis.