The 'SaaS apocalypse' narrative is incredibly overstated in the near term, though it may have truth over a 20-year horizon.

(00:00:00) The anxiety that I see is if you can generate an enormous amount of code and no one is reading it, you don't know the quality of the code. (00:00:11) Nobody deeply understands the code base and there's more fragility, right? (00:00:15) It's like the slop problem, vibe coding slop in my actual production code base. (00:00:19) But I think the broader problem that new company could go solve is like nobody knows how to manage that issue of (00:00:27) human attention to engineering. (00:00:29) I think it's like open season around this really, really big problem. (00:00:39) Hi, listeners. (00:00:40) Welcome back to No Priors. (00:00:42) Markets are melting down about the end of software. (00:00:44) Today, Elad and I are hanging out and asking, is SaaS actually dying or are people just projecting five-person startup behavior onto the Fortune 100? (00:00:52) We'll talk about what's real, incredible revenue growth, collapsing token costs, and faster turnover of vendors. (00:00:58) What's just hype and how to size the opportunity. (00:01:01) We also discussed the changing bottlenecks in building a software company and some parallels to the internet and cloud eras. (00:01:07) Let's get into it. (00:01:07) It's good to hang. (00:01:09) The market is freaking out around us. (00:01:13) So in all that noise, what are you thinking about? (00:01:15) Oh, you mean the SaaS apocalypse? (00:01:17) The SaaS apocalypse, the end of software. (00:01:20) Yeah, the end of software. (00:01:21) That's kind of interesting. (00:01:22) I feel like there's some meta trends that people are getting right and then a lot of specific companies that people are getting wrong. (00:01:28) And so I guess the basic premise is that SaaS software and Perseed software will no longer exist and everything's going to be replaced by AI and everything's just going to get vibe coded. (00:01:38) So why would you pay X dollars for a Salesforce instance when you can just vibe coded internally? (00:01:43) And all that stuff strikes me as incredibly shortsighted in the near term. (00:01:47) Over the long run, who knows what happens in 20 years or whatever, but there's lots and lots of companies that are quite durable. (00:01:53) I think an interesting example of that where I'm still a shareholder is Samsara where, you know, nobody's gonna Vibe code a fleet management app that will then be distributed through like what, Vibe sales? (00:02:03) Vibe, you know, enterprise sales or something. (00:02:07) And you're gonna build a Vibe like in-cab camera sensor that everybody will install in these fleets. (00:02:14) and then you're going to support them using Vibe agents or something. (00:02:16) It's just very overstated. (00:02:18) So I feel like it's one of those things where there's a massive market correction around something that in the long run has a lot of truth to it. (00:02:24) And maybe in the short run for certain types of companies has a lot of truth, right? (00:02:27) Ultimately, I think Decagon and Sierra are examples of companies where you're moving from proceed software to basically utilization-based customer support-related agents, right? (00:02:35) That is a real shift. (00:02:37) That may impact some of the prior wave of sort of perceived software companies, but this isn't going to be every single SaaS company. (00:02:44) So I view it as very short-term, overstated. (00:02:47) In the long run, who knows? (00:02:48) How about you? (00:02:49) How do you think about it? (00:02:50) I mean, I think the idea of Vibe Enterprise Sales is hilarious because we have portfolio companies. (00:02:56) with hundreds of millions of dollars of revenue, who are very committed to as much token usage as we can, as few great people as we can have, and today, they've less than 50 engineers. (00:03:08) And they went from zero to like, let's say close to 100 salespeople very quickly, right? (00:03:14) And so it's just a view from the growing AI natives that like vibe sales is not happening, right? (00:03:20) Oh yeah, vibe sales is definitely never good. (00:03:23) It's not happening anytime soon. (00:03:24) And so it's just, again, all this, it just seems like a very strong market reaction and market correction. (00:03:30) And it seems like it's very overstated, especially relative to a handful of companies that you're just like, why, like, how will you displace this company with (00:03:39) coding and in the fleet example, you're not going to have the fleet managers like writing their own apps to do all this giant surface area and stuff. (00:03:45) It just doesn't, it's just not going to happen in the short run. (00:03:48) I think a lot of it is actually driven by some assumptions that, you know, persona close to my heart, but engineers and builders are making about like the rest of the world, right? (00:03:59) Because there's this, there's this implied belief that like everyone will want to make their own software. (00:04:04) And I think it's like software is eating the world. (00:04:07) Is that what you're trying to say? (00:04:09) I am not. (00:04:09) I think like we're we're still. (00:04:12) Time to build, Sarah. (00:04:13) Time to build. (00:04:13) I don't think that everybody wants to make their own software. (00:04:17) I think some set of people want to make it and others will want other people to do it for them. (00:04:21) And like sometimes like what's a what's a like if you think about a good example of this, engineers sometimes have a like my personal labor (00:04:32) focused picture of the world. (00:04:34) So if you like, should you build JIRA in most engineering organizations? (00:04:38) Like, is that a... (00:04:40) Yeah, it's not the best use of your time if you're focused on product. (00:04:43) I mean, the other piece of it is the examples that people use. (00:04:47) Oh, my five-person startup built our own CRM, vibe coded it, blah, blah, blah. (00:04:52) Yeah, of course. (00:04:52) I mean, before that, you just did it all in a spreadsheet, and that was fine too. (00:04:55) You'd have to vibe code anything. (00:04:57) And so for very limited niche applications where it's a technical team doing something really quick because it's useful and custom and bespoke, amazing. (00:05:05) Of course, that's going to happen. (00:05:06) Does that mean that a Fortune 100 company is going to displace their CRM with some internal thing that got bipoded over the weekend? (00:05:13) Probably not. (00:05:14) And so I think it's also extrapolating or projecting behavior of very small technical startups onto the world's biggest enterprises. (00:05:23) And that's the second thing people are getting wrong is they're misunderstanding the (00:05:26) the moment. (00:05:27) And I think the internal software stuff that people are building is amazing, right? (00:05:31) It isn't impressive that you can do that. (00:05:33) It's incredibly impressive. (00:05:34) It's just extrapolating that behavior so aggressively so early just doesn't make that much sense right now. (00:05:40) I think to your point of the five-person company versus the very large enterprise, if you ask that same engineer who's pissed about paying $10 a seat for Jira, if you asked him or her, do you want to do the change management in Bank of America of getting everybody to do this the way you think is right? (00:05:57) And then dealing with all the security considerations and managing other people's opinions about potential changes to the story management workflow and then maintaining the system. (00:06:06) The answer is like, probably not, you know? (00:06:08) And so I think it is focused on, I actually think the idea that actual production of code becomes not the bottleneck for, if you know what the spec is, not the bottleneck is like incredibly interesting, but I (00:06:26) I do think it overstates like how much of the overall software vendor problem that is. (00:06:33) Yeah, I think people also misunderstand how much demand exists for software products. (00:06:39) And by software products, I mean everything. (00:06:41) I mean AI, I mean... (00:06:42) Is software eating the world? (00:06:43) Is AI eating the world? (00:06:45) AI is eating the world. (00:06:47) So I think that is actually true. (00:06:49) And I think Mark's post on that was really thoughtful and poor thinking on it all. (00:06:53) I think that fundamentally, (00:06:56) you know, there's so much demand for software, and there's so little supply of engineering in reality, relevant to that demand, that as you add this enormous boost of productivity to software engineers, it just gets soaked up, right? (00:07:12) Because there's so much more stuff to build and to do. (00:07:15) And I don't see teams, you know, startup teams continue to hire engineers for a reason, you know. (00:07:20) I think the nature of the work is shifting, and I think some people are going to have real issues with that shift. (00:07:25) Because fundamentally you're shifting from, in some cases, there's a few different types of mindsets around engineers. (00:07:33) And one of the mindsets is the really bespoke craftsmanship. (00:07:38) I'm going to do the aesthetics of the thing that I'm doing really well, and I care about the code quality and the artisanal version of what I'm doing. (00:07:48) And then there's people who write code because it's a utility that allows them to build product. (00:07:53) There's some people who really like aspects of the math. (00:07:56) There's lots of different motivators for people to write code, and I think a subset of those people are going to be less happy in the new world. (00:08:03) It's kind of like the indie game developers who'd make these handcrafted individual games for themselves and then for their friends, and then they'd launch them on the Apple Store or whatever, versus the people who'd work at EA. (00:08:14) And they each had their own version of craftsmanship, but it was just a different type of thing. (00:08:17) I think we're going to see a lot of these really great engineers who care about the (00:08:22) bespoke craftsmanship about everything they do, they're going to be unhappy working at larger companies as these coding tools get even more accelerant, because it goes against their approach of how they like working and what they enjoy out of the work. (00:08:35) And for other people who are really focused on the utility of just building product, it's going to be freeing in some ways. (00:08:39) So I think there's also like a variance in terms of the reactions to this stuff, depending on the type of utility function that you have relative to the work you're doing. (00:08:50) Yeah, I think related to that, the one thing I've seen is that if you have an engineering identity that's based on like a value-based ranking of difficulty or skill, like the specific types of engineering that are considered, you know, impressive or high status can actually be like less hard for agents, right? (00:09:12) So I think there's an enjoyability like element and then an identity element. (00:09:17) And actually, one of your founders from Applied Intuition wrote a good blog post where there is an essay where he says, like, keep your identity small. (00:09:26) I think that's like wonderful overall advice for this period of time, right? (00:09:30) You're like more adaptable if it's true. (00:09:33) But I think your overall view of there are a lot of unsolved problems and like making an abundance of software can better address that. (00:09:41) I strongly agree with. (00:09:43) And one thing that (00:09:45) actually is near and dear to the audience that is really unsolved is like we've broadly been thinking about what happens if you have abundant code generation. (00:09:55) And in like, I think in all of our teams, agent-first engineering management and thinking about code quality is an unsolved problem. (00:10:03) Yeah, and we'll get there. (00:10:04) It'll be your kill work and we'll get there. (00:10:06) What do you view as the major problems? (00:10:09) Well, the anxiety that I see is (00:10:14) Like if you can generate an enormous amount of code and no one is reading it, you don't know the quality of the code. (00:10:22) Nobody deeply understands the code base and there's more fragility, right? (00:10:26) It's like the slop problem, but instead of it being like vibe coding slop for random websites for non-technical people, it's vibe coding slop in my actual production code base for... (00:10:37) every lazy engineer, which is every engineer. (00:10:38) I think people are like looking at some problems of actually do you think ticketing, ticketing systems are like at risk, but I think the broader problem that Jira could go solve or new company could go solve is like nobody knows how to manage that issue of human attention to engineering. (00:10:58) And there's a bunch of ideas like testing and smart review, just let agents do it, formal verification. (00:11:04) I think it's like open season around this really, really big problem. (00:11:08) I think the one other thing people are bringing up that I don't quite buy is that agents are already making big decisions for vendor purchases and things like that. (00:11:16) I think somebody near and dear to your heart posted about that. (00:11:18) And I think that there, the statement was, oh, agents are increasingly making decisions about what software people are using. (00:11:27) And really what that is, well, you have a partnership, your Cognition or your Cloud or whoever, and you have a partnership, and it's part of that partnership, you spin up a Supabase instance and you use very specific tools because you have a partnership to do that. (00:11:39) And that's always happened. (00:11:40) If you're using Airtable and they're on AWS, you're spinning up an AWS instance without knowing about it in the background. (00:11:46) So I also think that whole notion that in the short run agents are making these choices is also (00:11:53) overstated. (00:11:53) I think in the long run it's true. (00:11:55) But then you get into all sorts of agenic commerce decisions and do they understand your persona and what you're doing. (00:12:00) actually want and need and all this stuff. (00:12:02) So I just feel like we're in a little bit of a noisy moment where people are potentially, and I'm somebody who's very pro-AI progress and a believer in all the changes that have happened and are coming, but I think we're having a lot of overstatement now of what's actually happening in the world. (00:12:15) And part of that is a SAS apocalypse and this giant reconnection. (00:12:20) And part of it is extrapolating that the future is here already, when in many cases it's just say we did a BD deal or whatever. (00:12:28) So (00:12:29) I just think people kind of need to, or the multiple stuff where you're like, yeah, that seems human generated in terms of the emergent behavior. (00:12:37) So I don't know, we're in this odd moment where I feel like this was the month of hype in a way that we haven't seen in a while, where a bunch of stuff got overstated in all sorts of ways and people believed it. (00:12:48) And by people, I mean like mainstream media and others are like, oh my gosh, look at this behavior of (00:12:52) these agents trying to cut out humans from their forum where it's Reddit like and blah, blah. (00:12:56) And you're like, okay, maybe you should see where the posts are coming from in some cases. (00:13:00) And it's exciting, by the way, don't get me wrong. (00:13:02) I think it was very exciting behavior that's happening. (00:13:04) I just think a subset of it was planted for marketing purposes. (00:13:07) Yes, certainly. (00:13:07) I think people are also figuring out, there are things that tap into deep emotional reactions that people have to their view of things that feel very human, right, from a marketing perspective. (00:13:22) And like, that's clearly one of the things that's happened around them, the Multbook stuff. (00:13:25) I also think that like, one of the things I actually think happened was like the idea that demos are different from the reality of the full software that you need, like has not quite arrived in many of the equity research people's desks, right? (00:13:40) And so like, I'm like, guys, like your whole job was to think about (00:13:45) these, like the structural advantage of your businesses and what is going to compound. (00:13:49) And the theory of competitive advantage didn't just like poof, disappear, right? (00:13:53) Like software markets have been a fight about how to do things and how to distribute to customers as well as a battle of how to produce code for a long time. (00:14:03) So I feel like that has been missed a little bit, but I do think long run, the fundamental thing that the (00:14:13) bottleneck on production of, you know, expensive to produce software being loosened is really cool, right? (00:14:23) It just means like if you think of there's a lot of embedded points of view in software on how to solve a problem, right? (00:14:29) You know, if it's engineering or enterprise sales, not a very software problem or general productivity, right? (00:14:35) Like Notion is a way to do things. (00:14:38) It's a building block system, but it's definitely got a point of view. (00:14:41) And so if you (00:14:42) reduce the cost to express that point of view in software, I think it's cool that we're gonna see a lot more ideas. (00:14:48) That's amazing. (00:14:48) And again, I think it's a revolution. (00:14:50) So don't get me wrong, I've been involved with coding companies really early on, and I'm very excited about everything that's happening. (00:14:58) And I think it's transformational, and I think it's revolutionary, and I think it's really important. (00:15:03) I just think we had a month of kind of bullshit hype. (00:15:05) Okay, so if we ignore the noise of the last month, (00:15:10) where people got a little like frantic. (00:15:12) What do you think is a signal that people are not paying attention to enough in such a noisy landscape? (00:15:19) You were telling me that like growth pace is like of the biggest companies is still underpriced. (00:15:26) Yeah, one thing that Jared on my team put together that I thought was super interesting was he pulled data from Capital IQ where they just like predicted. (00:15:36) some projections on OpenAI and Anthropic, and they looked at, and then he's sort of graphed out, and maybe we can share these graphs as part of this episode. (00:15:47) He graphed out how long it took different companies and years to go from a billion in revenue to $10 billion of revenue. (00:15:53) So for example, ADP, (00:15:55) took 20 something years to grow from a billion to 10 billion in revenue. (00:15:58) And then the next wave of companies like Adobe took about 20 years to go from one to 10, and then you fast forward in time and you have things like Salesforce or SAPs for an even more modern cohort, and they took eight or nine years. (00:16:10) Microsoft took seven-ish, eight years. (00:16:13) Google and Meta and AWS took a couple years, three, four, five years, but the AI labs did it in roughly a year. (00:16:22) And then if you look at the projections- It's a wild chart. (00:16:26) It's a wild chart. (00:16:26) And so we should add it. (00:16:28) But you just see it go from like 20 something years with Adobe to like a year for the AI labs. (00:16:33) And then if you look at the projections that are sort of the public projections, they aren't necessarily the company-driven data, but the public projections on where the labs will end up or how long it'll take them to go from 10 to 100 billion in revenue. (00:16:45) For Microsoft, that was something like 27 years. (00:16:50) For Google, it was over a decade, same with the AWS, roughly the same for Meta. (00:16:55) And then for the AI labs, it's like three, four, five years. (00:16:57) It's very fast. (00:16:59) And so we're seeing the fastest time to real massive revenue that we've ever seen in the history of software. (00:17:04) It's just these insane curves. (00:17:06) And again, we should just post them. (00:17:07) Part of that, I think, is just the internet has created this global pool of liquidity and then something of your customers online. (00:17:11) It's much easier to distribute than it's ever been. (00:17:13) So that's one piece of it. (00:17:14) There's more people with access. (00:17:16) There's higher GDP. (00:17:17) There's lots of drivers for that. (00:17:18) But then simultaneously, you're just creating enormous business and user value at massive scale simultaneously. (00:17:25) And these capabilities are so rich that you're seeing this take off in terms of revenue. (00:17:29) And so it's unprecedented. (00:17:31) It's really impressive. (00:17:32) And I think people are ignoring the revenue and usage side of the equation. (00:17:35) The other thing that we actually put together was the collapse in token pricing for equivalent models. (00:17:41) I think this was done initially by David, who worked for me, and then Shrin. (00:17:45) And so for example, we looked at the cost of a GPT-4 level or equivalent model. (00:17:49) We looked at that a year or two ago, and basically in 21 months, it went from like 37 bucks for a million tokens to 25 cents. (00:17:58) And so pricing dropped by 150x in 21 months. (00:18:03) And then we tried to accelerate that curve, but obviously people aren't really using GPT-4 level models anymore. (00:18:08) even though, you know, they're two, three years old. (00:18:10) And so we looked at O1 equivalent models and the cost of a million tokens on an O1 equivalent model in December of 24 was about 26 bucks. (00:18:17) And then in November of 25, it was 30 cents. (00:18:20) So we saw another 88 X drop, not 88% or 88, you know, 88 times cheaper in 11 months for that next generation of model. (00:18:30) So we're having pricing collapse on the token side while we're having revenue ramp insanely on the (00:18:38) usage side. (00:18:39) And so that's insane if you think about that, just this pace of shift of cost, of revenue, of utilization, of everything. (00:18:45) And this is back to like, I'm incredibly bullish on everything that's happening. (00:18:50) And so it's more just modulating it against this, odd over extrapolation of what's actually happening or actual capabilities or, what these things are really doing. (00:18:58) Yeah, I think one thing that people miss in the like bear case and all this stuff is, as you said, like revenue numbers, which is hard to miss, but, and then (00:19:09) Just like actual like token inference count, right? (00:19:12) If you look at one, if you look, where's the inference happening? (00:19:16) It's either happening in inference clouds, right, based on mobile fireworks, or it's happening at the, like, the very large model providers. (00:19:24) And it's happening in a lot of spring, which is still much more 2 magnitudes more. (00:19:28) And humanity in general. (00:19:31) Yeah. (00:19:31) Yeah, that's true. (00:19:32) In terms of power utilization, human brain is really impressive. (00:19:35) What is it like? (00:19:35) Tens of watts, 20 watts? (00:19:37) How much? (00:19:37) Like what's the power utilization of a human brain? (00:19:39) It's look at it right now. (00:19:39) It is 2 magnitudes. (00:19:42) It's like 10 or 20 watts, I thought. (00:19:44) I think to the point of like real data, the inference clouds are growing 1000 X in terms of consumption, right? (00:19:52) And then they're getting more efficient. (00:19:53) So revenue grows at some. (00:19:56) lower rate than that, but it's wild. (00:19:58) It's 12 to 20 watts of power, which is comparable to a dim light bulb or a computer monitor in sleep mode. (00:20:03) It's not even like a computer monitor. (00:20:05) It's when your monitor is sleeping, that's the amount of energy that your brain is consuming as it does all these crazy calculations. (00:20:11) It's one blade of 1 GPU fan in one of these data centers. (00:20:16) Yeah, it's nuts. (00:20:17) I feel like Noam Shazeera's brain though is probably consuming like 1000 watts. (00:20:20) Well, I think that's great. (00:20:21) I think like we have a lot of efficiency work to go. (00:20:25) I kind of meant it the opposite. (00:20:27) he's so smart, but he's probably consuming more energy. (00:20:28) But to your point, maybe he's more energy efficient. (00:20:31) Maybe he's at like 1 watt and I'm like at 1000 watts or something. (00:20:34) I meant for the computers. (00:20:36) We're all stuck without the, you know, brain computer interface work improving. (00:20:41) But I'm just interested in how much efficiency we can get out of the models. (00:20:45) Yeah, it's probably obviously just based on the human brain, there's a lot of room. (00:20:49) You know, one thing I do think about, I was talking to a friend who leaves a bunch of purchasing and (00:20:55) traditional large enterprise this morning. (00:20:57) And he was like, oh, well, the incumbents can, this whole thing is overstated. (00:21:02) We're so committed to all these big enterprise vendors, whatever, a lot of things that we've been talking about here. (00:21:10) And his other view was that the incumbents have the money to buy and go fight back on these dimensions. (00:21:16) One thing I immediately thought of was just like, (00:21:22) Reflexivity in markets is such a good concept. (00:21:25) And here it's like, well, they do, unless they don't have the market cap to do it, right? (00:21:29) With these companies that, to your point, you know, first the labs, but then a series of the very best application companies, if they're growing to a billion of run rate rapidly and valuations grow in concert with that, then I do think there's a question on whether or not you, (00:21:49) have the currency to compete too. (00:21:51) Yeah, I'm already seeing that in the SF housing market, right? (00:21:53) Where SF housing is starting to rise again in part due to, I'm assuming, outcomes from the lab tenders and things like that. (00:22:01) Because suddenly you have these companies that are worth hundreds of billions of dollars out of nowhere in a few years. (00:22:06) And as employees are selling into tenders, there's this new sort of influx of cash in the ecosystem. (00:22:12) So, and there's also Nvidia going from, you know, (00:22:16) tens of billions or a hundred billion to trillions in market cap. (00:22:18) Like there's just this shift happening right now in terms of scale. (00:22:22) And there's an interesting question actually where this is one other thing that we looked at as a team and maybe I should just publish all these slides. (00:22:28) We basically asked what proportion of GDP is tech, right? (00:22:35) And just the US economy at least. (00:22:37) And how has that grown over time? (00:22:40) And also, what does that meant in terms of market caps, right? (00:22:43) And so if you look back to 2005, Google was worth $100 billion and Exxon was the world's most valuable company, $400 billion in market cap. (00:22:54) And then it took until 2018, Apple was the first company with a trillion dollar market cap. (00:23:02) market cap ever. (00:23:03) Everybody was shocked that anything could get to a trillion. (00:23:05) And at the time, tech represented about 30% of the S&P. (00:23:09) Before that, it was, say, 10%-ish back in 2005. (00:23:15) And now the top eight tech companies are about 23 trillion of market cap, and they make up well over 50% of the S&P in terms of value. (00:23:25) At the same time, they went from basically 4% of GDP (00:23:31) in 2005 to about 12% of GDP today. (00:23:33) And so then the question is, what proportion of GDP eventually just becomes tech? (00:23:37) And AI is a driver of this, right? (00:23:39) Because you're taking services and you're taking certain types of jobs and you're augmenting them with AI and you're converting them into effectively software spend or tech spend. (00:23:47) And you can make different assumptions about growth rates. (00:23:49) And then based on that, you can end up with anywhere between 15, 20% of GDP to 30% of GDP in 2035.