投资 SaaSpocalypse：Heller House 的 Marcelo Lima

Marcelo, we are living in the SaaS apocalypse. You have been following and investing in SaaS companies for a long time. You've published 2 rebuttals to the SaaS-apocalypse narratives in the past couple of months. If you give me a link—you published them on Mailchimp, so I'm not sure if I have a link or not—I can include a link in the show notes if you want to make them public. I thought I'd have you on to discuss the SaaS apocalypse and everything.

I've got tons of questions and notes, but I'll start right now. I think your argument is that the SaaS-apocalypse narratives are overblown and there's a generational opportunity to buy some of the best companies the world has ever produced. I'd love it if you could talk about that—why you think this is an opportunity and why you think this is overblown. Again, you'd be shocked at how long my notes list is for this podcast.

Thanks for having me, Andrew. It's a long topic, but in short, I do think that—I'll give you some context. I've been following a lot of these companies since probably around 2018, and I've been following them very closely. I used to be a software engineer in a previous career, and I've been using Claude Code a lot recently because it got really good. I tried it last year when it was still not very good, but I've been using all the leading models in parallel because I send queries to all 4 and see which one is better. I've been doing this for several months now, since Gemini 3 Pro came out.

I've been seeing a lot of these companies—to name 3, maybe all 4: Adobe, Atlassian, ServiceNow, and Salesforce—to different extents, putting AI into their products for quite a while now. I think ServiceNow's Pro Plus SKU, which has AI, went GA, if I'm not mistaken, in late 2023. Don't quote me on that. I think that's right. They've been talking about the adoption of this AI-enabled SKU for years now.

It was very shocking to me that starting—I don't know if you remember the very beginning of the year, in January, when people came back from vacation and all of a sudden the trading days really began in earnest—you saw software as a service gap down and semiconductors gap up. These alligator jaws just went, and you saw semiconductors explode and SaaS get systematically liquidated. We heard even at the end of January from the Jefferies trading desk, “Get me out at any price.” There was 10-to-1 volume selling versus buying in software names. These things are uninvestable, et cetera.

That, to me, raises a red flag. Are people—did people not know what they owned? What's going on? Did they never study these companies? Are they not aware that these companies are putting AI into their products already and have been doing it for a while?

For somebody like me, who sometimes, to my detriment, is so deep in the weeds that I don't have the theory of mind to understand what other people are thinking, I couldn't understand the perception issue. The perception is, again, that this is something new, whereas to me it wasn't something new because I've been following these companies for a long time. To me, it was, “What is going on?”

Sure, Claude Code is much more capable, but I do think that what it does is make software developers way more productive. Software developers work inside software companies—Salesforce, ServiceNow, et cetera. These companies have armies of engineers who write code for a living.

I don't think that it makes any sense whatsoever that a company whose main product is not producing software—the example I pick is Volkswagen, because they had that project, CARIAD, where they were trying to develop software for their cars. It was a complete disaster, and they lost billions of dollars. I pick on PepsiCo because they sell snacks, but you can fill in the blanks with any kind of company. Their core competency is selling cars or selling snacks; it's not developing software.

It's not like these companies are going to all of a sudden get that DNA and start developing software just because Claude Code came out. To me, it was always a tool that would accelerate the roadmap of existing software companies, not turn a poodle into a lion, so to speak, by making these non-software companies all of a sudden geniuses at developing their own software.

Yeah. [Laughter.] I totally hear you on that. I'm a generalist, and I'm probably a pretty dumb one, to be honest with you. But let me provide some pushback and thoughts and things that are floating around, because as I've said before, I love to run into sectors where there's panic. I just want to be the guy who is running into—you know, somebody screams, “Fire!” in a burning building and everybody's running out, and my instinct is to run in, right? I really run in.

I will say I did banks in 2023, and I did busted biotechs in 2025, neither to the size I should have done with the benefit of hindsight. So, it's not like I'm saying I'm a genius, but I wanted to run in. I'll throw some random things out that worry me here.

Number 1, I think the overarching worry I have is that the AI tools we're using right now are the worst AI tools that we will ever use, right? They're getting better every day. I could point to a topical one that's come up since we planned this podcast: the Anthropic Mythos thing, which is apparently so good at cracking and exploiting zero-day exploits that the White House is literally calling in all the presidents of the big banks and saying, “You need to use this. Anthropic cannot release this publicly. You guys need to use this to shore up your thing,” right?

Anthropic's Mythos—I don't know if Mythos just changed the enterprise value of a cybersecurity company, where it's going to be hugely positive and there's going to be a lot more consulting work, or if it just zeroed them out, right? But I can guarantee you that Anthropic's Mythos massively changed the enterprise value and terminal value of every cybersecurity company. I could imagine 15 different other things that Anthropic could roll out.

As these AI tools get better, I know Salesforce does a lot of stuff, but are you telling me to ignore PepsiCo vibe-coding its own Salesforce? Are you telling me Anthropic couldn't say, “Hey, we're going to code a perfect Salesforce engine on our own and release it, and it's going to be maybe 80% as good as Salesforce, but it's going to be a heck of a lot cheaper”? That's my number-one overarching worry.

AI gets so much better, and then all these AI companies can vibe-code replacements that are 80% as good, and you get into a pricing war. There's just a lot of competition coming. I've got other ones, but that's the big worry that weighs on my mind here.

Yeah. Look, there's no question that there are legitimate worries, absolutely. I think the compression in multiples is justified as a way to discount a lot of these worries. I do believe that the multiples have gotten so low that a lot of these companies are now priced for perpetual decline.

The way I define that is, if you just assume a 10% discount rate, you have your model: 1 over R minus G. R is your discount rate, and G is your growth rate. If you plug in 0 for G—0 growth—and your discount rate is 10%, 1 over 10% is 10 times.

So, let's just assume 10 times earnings or 10 times free cash flow, whatever your multiple is. A lot of these companies are there and priced for 0 growth, or even priced for perpetual decline on a forward basis—so lower than 10 times on a forward basis. I do think that the market very quickly repriced these things to reflect a lot of that risk.

So, yes, the barriers to building code have gone down. I'll give you the puts and takes, in my view. On the one hand, the barrier to generating code was never the real barrier, so to speak, because you could always go to different geographies. You could hire Indian consultants, for example, or consultants in the Philippines, or you could go to one of these platforms where you can hire folks to do jobs for you on a contract basis. They could produce code very cheaply in Eastern Europe, etc., and so you could replicate a lot of these things.

I'll give you another silly example. I don't know how many people know this, but there's a free, open-source version of Photoshop. I think it's called PTGui. It's existed for decades. A lot of different pieces of software have open-source equivalents, and yet consumers and enterprises choose to pay for convenience.

They would rather pay $10 a month per seat, $20 a month, whatever it is, to have software that is constantly improving, that has security patches, and that has compliance. When you get to enterprise software, it gets a lot more complicated, because now you have things like SOC 2 compliance, which requires you to audit every single line of code, and you're talking about a code base with millions of lines of code.

My understanding is that it would not pass an audit if you vibe-coded the whole thing, if you could vibe-code the whole thing. So, I think the barriers in enterprise software were never, in a way, the cheapness of the code. It was trust, governance, and regulation.

There's also a saying that a lot of enterprise software doesn't get bought. Some of it does, and some of it is product-led, but most of it gets sold, so you need a whole sales motion, enterprise teams, and system integrators, as you know—Accenture, etc.—to implement these systems.

Having said that, yes, at the margin, if you are a company—and usually smaller companies will do this, because there's less risk—you're probably not going to see PepsiCo vibe-code its own CRM, but you probably will see small companies vibe-code their own CRM to the extent that all they're using is simple functionality for customer relationship management.

Obviously, CRM—the company Salesforce—that's not what they are. They are a whole suite of things. They started with CRM 27 years ago, but it's evolved, and they've acquired several businesses and integrated them.

Until they change their ticker, I'm only going to think of them as a CRM. They've got to change their ticker if they want me to think of them as something else.

Yeah. Marc Benioff joked that he wants to rename the company Agent Force. You hit on one—there were a lot of questions and a lot of things I prepped that you hit on there that I want to talk about, but let me just start with the last one.

You hit on Salesforce, and I actually have an interesting anecdote. I was talking to a VC maybe a month ago, and they said—I might be getting the exact timeline wrong—but they basically said, “Hey, it used to be, when we wrote a Series B check into a company, that was basically the day where they would call up and get their first Salesforce CRM application.”

“What we're seeing now is that companies aren't calling up Salesforce and getting their Salesforce CRM until late Series C, early Series D.” I think the implication, as we discussed, was kind of like, “Hey, they still need the enterprise-grade stuff, but they're able to push it back right now, right?”

I thought that was really interesting, because what small companies do in the beginning is kind of where the larger companies go. Small companies—that's the advantage of being small: you're more nimble. If people are just pushing back from Series B to Series D, Salesforce will be fine if at Series D you're still getting it, right?

But you could imagine that if people are starting to vibe-code and use their own tool at Series B, maybe 2 more years from now, instead of Series D, it's Series E. Then, instead of Series E, it's when they go public—or not at all.

So, again, that's just another way of looking at it. I'm not sure if that's a bull or bear case there, but I thought that was an interesting anecdote, and that's one thing that's really stuck in my mind. I've been thinking about the delay here. The small companies seem to be moving away from some of the larger things, and is that where the puck is going?

Yeah, that's a very legitimate worry. A few thoughts. I guess number one is: Is the company getting the Series B check delaying the adoption of Salesforce because they are vibe-coding their own Salesforce in-house, or is it because they're allocating budget to Anthropic instead, to use Claude Code so that they can advance their roadmap quickly and get a minimal viable product in the customer's hands quickly? Do you know the answer to that?

Based on my conversation, I'd have to follow up with them. I think they were saying, “Hey, they're using internal tools to manage customer relationships at the Series B.” But I don't want to say that strongly, because, A, it probably depends on the specific company, and B, I wasn't really pressing on the details. I was just thinking—expanding my mind when he said that.

What I think—the overarching idea that I have—is that it is the incumbent's prize to lose. They have this amazing feedback loop: They have forward-deployed engineers inside their largest customers implementing a lot of the software, and they can see where the pain points are. They can then go talk to the mothership and say, “Hey, guys, we have to build X, Y, and Z, because this is not working, this is clunky, this is broken,” etc.

They have the enterprise go-to-market teams that talk to all these companies as well. They have armies of engineers back at headquarters improving the software, deploying continuously, and making improvements in real time. Atlassian, for example, is a company that pushes updates daily to its software.

So, given the fact that they have this feedback loop, it's really incumbent on them to improve their software. Now that they have this incredible leverage that Claude Code and Codex have given them in terms of being able to speed up their development—and these tools are amazing—the onus is on them, really, to improve that software as fast as possible to prevent what you're describing from happening, right?

Let me provide so much value to the customer that it becomes irresponsible for that customer not to pay the monthly fee for my Salesforce seat, for example. One more thing: Did you see the headless announcement that Salesforce just made—Salesforce 360 Headless? Marc Benioff tweeted about it today. It came out a couple of days ago. This, I think, is extremely important and interesting.

And basically, headless—what that means, for the benefit of the audience, if people don't know—is this: Let's say I'm in Claude Code on my terminal on my computer, and I want to interact with Excel. Claude Code can use a couple of tools. It can use Open Pixel, which is a Python library. You can use Excel Wings. And it does so in a headless fashion.

In other words, it doesn't have to open Excel. I never see the Excel user interface, and Claude Code never sees it. It's not like Claude Code is going in there and clicking buttons and editing cells. What it does is use these tools to go into Excel programmatically and change this and edit the cells for me.

So, that's what headless means. You can access a lot of different pieces of software that way. You can access browsers in a headless fashion as well, in the background, where it doesn't show up on your computer.

Okay. So, what Salesforce did is expose all the functionality on its platform, whether you're talking about Slack, Tableau, MuleSoft, etc., in a headless fashion through Model Context Protocol interfaces, which are kind of like an API—a way for your model to talk to their software.

This makes it dramatically easier to program on top of Salesforce, and you can now build any kind of user interface you want because you've got these Lego bricks, which are the MCP servers on Salesforce's side. You can now spin up a Claude Code window and say, “Hey, build me a React dashboard that allows me to look at my customers and look at my leads over here.”

A lot of the complaints you hear from people are, “Oh, Salesforce is legacy software. It's super ugly. It's clunky to use. I don't like the user interface.” Well, guess what? You can now have any user interface you can possibly imagine. So that's really cool.

And look, you were deeper into this than me, but I'm using Claude—specifically Claude Cowork—to build some really cool investing tools. We're both investors, right? One of the ways I prepped for this podcast was with my Claude Co-Work research function and my AlphaSense agent: “Hey, deep dive into these 5 stocks and just give me a report when you're done,” all that type of stuff.

But I would note one thing you said in there. People complain about Salesforce's functionality. It does strike me as if people are using Claude—to use a very loose reference—as the OS. If Claude is the thing they're doing everything on, and they're saying, “Hey, go to Salesforce, change it, and present it,” it does strike me as kind of worrying if they're saying, “I want to use Claude to create my custom interface.”

What is Salesforce really doing at that point? What Salesforce is really doing is serving as the back-end software. You're also probably pulling from a lot of data stored with them. But if Claude is the thing you're using to command and provide all of the front-end stuff, it seems like Claude is pretty damn close to taking over Salesforce.

Is it taking over Salesforce, or does the Claude team just say, “Hey, Salesforce is a $100 billion company. Everybody's coding their own front end on Salesforce. Why don't we just come and take them out? It won't take that much investment.” Does that make sense as a risk factor? I could be completely wrong in how it's happening, but it strikes me that if everybody's using Claude as the thing, the back-end thing can be switched off pretty quickly.

Yeah, that's a very common worry. I don't think it's right. I'll tell you why. This is not new. When Slack went public in June 2019, I attended the Slack investor day and was talking to Stewart Butterfield, the co-founder. Slack and Salesforce had nothing to do with each other back then. This was before Salesforce acquired Slack; Slack was going public as an independent company.

He said, “I never open Salesforce. When I go to a meeting, I'm in Slack. I just pull out my phone, type a command in Slack—a slash command—and pull up the latest notes on the customer I'm about to meet. It's like, ‘Customer XYZ has $1 million of ARR with Slack,’ and then all the notes.” When he's done with the meeting, he can quickly type up the notes in Slack, and Slack will go in the back end and update Salesforce.

So does that mean Salesforce was useless in 2019? No, because these enterprise customers have very, very complex sets of permissions. There's governance behind all this: who can access what, and which user can access what kind of data? There's tons and tons of historical customer information in there—15 or 20 years of it. You've interacted with customers and closed deals. There are call logs, call-center transcripts from recorded calls, contracts, and a whole lot of other things that sit inside this database.

To use the—I don't know if you remember when Satya Nadella said that SaaS is CRUD, which is create, read, update, and delete. He said that all SaaS is just a UI on top of a CRUD database. That was, I think, December 2024, and it caused a mini SaaS-pocalypse at the time. You could say that about a lot of things. You could say that about Microsoft. You could say that about it, right?

I'll give you the counter to that. And, by the way, I'd love to hear what you're building. I've been building some things using Claude Code, and the things I'm building need to use Excel because I build my models in Excel. I have all the financial data for all my companies in Excel.

What the software I've built internally has done is entrench Excel because it depends on Excel existing. If Excel went away, I don't know what I would do. Where are my models going to live? Where are my company models going to live, where I have the historicals?

And, by the way, Andrew, we're talking about a capability that Salesforce created today. It doesn't mean that tomorrow 100% of the installed user base never logs into Salesforce again and just uses Claude Code. This is sort of a spark. It's an idea. It's something that they launched, and now there's going to be adoption.

Of course, part of the interface where people are going to interact with agents and build stuff is going to be inside Slack. It's going to be inside Salesforce. Part of the interfaces are going to be inside Salesforce. Now, does that necessarily mean that Claude, just because it's updating the database fields inside Salesforce, can replicate the functionality of Salesforce and completely delete Salesforce? No, because Salesforce has all that governance information, all that business logic, and all the customer history that Claude cannot see.

Unless you gave Claude some mythos-like access where it could cyberattack your whole thing and read your entire network graph, permissions, and all that.

I mean, look, I'm a journalist. I'm not knee-deep in software all the time, but I think your Slack example at the beginning—especially with the Slack CEO saying, “I don't even log into Salesforce”—is a really killer example of why Claude can't do this. I think it's a very good example.

But let me give you some more patterns that run through my head. Chegg—are you familiar with CHGG? This is the company that got murdered by ChatGPT, basically. For those who don't know, this was mainly an online company that you would use for homework questions. It was mainly for college students, right? You'd say, “Hey, I've got a question,” and it would provide answers. It worked with a lot of textbooks and everything.

ChatGPT came out, and the stock was down quite a bit because people said, “Oh, you don't need that. You just ChatGPT it.” It's not lost on me; I had been studying the history of it. In November 2023, they came out and said, “Hey, we're an AI beneficiary. We're putting in an ASR. We're going to buy back our stock like crazy. If the market's giving us a chance to buy at a discount, we're going to take advantage of it.”

Fast-forward to right now. Salesforce—we've mentioned them a couple of times—comes out a couple of weeks ago with literally the largest ASR in history: $25 billion, if I remember correctly. monday.com comes out, and if you read their 20-F, they've got a little kernel in there that says they bought back 15% of their stock. It says up until the end of March, but if you think about blackout periods and stuff, it's probably from the middle of February to the end of March. That's 15% of their stock in roughly a 6-week window.

I mean, these are aggressive things. You've talked about this in at least a few others: CEOs and boards of directors are coming in in a big way and buying their stock on the open market. On the one hand, I see, “The insiders are signaling this is an opportunity.” On the other hand, I say, “Chegg thought it was an opportunity 2 years ago, and they were dead.”

And on the insider buying—actually, I can't remember the exact price—but a lot of these insiders started buying at $250 and at $225, and now the stocks are at $180 and $160. I say, “I love insider buying. I love big buybacks. I love these signals.” But when I see that, I say, “Do these guys really have a read on value, or is the landscape shifting so quickly underneath their feet?”

They're kind of the guy who, in 2024, says, “Our stock's trading at 10 times earnings,” and then earnings are falling off a cliff and everything's going terribly. So, I threw a lot at you there. I'd love to hear your thoughts on that.

Yeah, that's a very fair point. You remind me also of—I can't remember which crisis it was, Andrew—where bank insiders were buying their stock all the way to zero. It might have been the GFC, where they were just buying their stock all the way down.

Lehman?

I think both Lehman and Bear Stearns had insider buying literally the week before they went to zero, if I remember correctly. First Republic had insider buying right before it went down as well.

At the same time, there were about 10 other banks where there was insider buying and the banks made it through. But banks in particular have CEOs who like to say, “Hey, this is great.”

Everything’s fine. And then there’s a bank run—zero. It’s all over. But it’s very easy for an insider who feels comfortable; things shift really quickly. They say, “Our stock’s a bargain, down 30%.” Bed Bath & Beyond was buying stock at $45 per share in 2021, and 15 months later it was in a death spiral to zero. Every situation has to be evaluated differently, right?

My mental model of the world is that things tend to have a power-law distribution, where you have the Pareto principle: 20% of the people have 80% of the income. You see this distribution so frequently in the natural world. Something like 5% of the videos on YouTube account for 95% of the watched hours. There’s a lot of skew.

As you know, Hendrik Bessembinder’s paper found that 4% of all stocks in the last 90 years generated all the net wealth in the U.S. stock market, if they were held during that period. I do think that in the universe of SaaS, there’s going to be a tail that’s probably not going to survive very well, or not survive at all, because these are companies that perhaps have point solutions or things that are more easily replaced.

I do think that the companies that have much broader platforms, are much more entrenched and mission-critical to their customers, and are difficult to replicate are going to be the ones that survive. By the way, they also need good management. You need management that is not asleep at the wheel. You can’t be the guy driving off the cliff without knowing what’s happening.

I think Salesforce is certainly one of those companies. It has Marc Benioff at the helm, and to his credit, he’s very open-minded and flexible. He has this mantra: “In the beginner’s mind there are many possibilities, and in the expert’s mind there are very few.” He always wants to have a beginner’s mind when he approaches any problem, and to his credit, he has demonstrated that over and over again, in my view.

I do think that they are adopting AI, and the fact that they’re opening up Salesforce to allow any agent to talk to Salesforce, again, in my view—maybe I’m wrong—entrenches Salesforce’s position. It increases usage of the platform and increases its attractiveness.

If I’m on the fence and I don’t want to pay for a Salesforce license because I hate the user interface and think it’s clunky and old, I’m like, “Great. Now I can have all the functionality that it enables. I can have access to the entire suite of products. I can have Slack, MuleSoft, Tableau—I can have all the things that they offer—but I can just pick and choose my modules. I can code my own user interface.” To me, that’s very compelling.

I do think it has to be evaluated on a case-by-case basis. One exercise I did, Andrew, in that February memo—which I guess I called the March memo because I released it at the beginning of March—was go through the earnings of all the SaaS companies that I thought were relevant.

I tried to figure out which ones were on the front foot when it came to adopting AI and putting AI into their products, and which ones were already seeing benefits in their products. A lot of them were. There’s a list at the bottom of that memo with all these companies, so I guess we will see. It remains to be seen.

I hear you on the Chegg example, but I do think that, so far, the best companies are adapting very well. You can see this in RPO growth, subscription-retention metrics, and all their financial metrics. But to your point, we don’t know what’s ahead yet.

Just building off that, that’s one question I had. It’s really interesting because the market adapts really fast. Again, I mentioned the Chegg example, where Chegg’s earnings came out and the stock was just down, down, down right away. You’ve seen these companies, and the sell-off starts at the beginning of the year. Most of them report earnings in mid-February or early March.

I’m not saying it’s across the board, because, obviously, the point-solution businesses, as you labeled them, have been getting crushed. But for the most part, the really good software companies you’re talking about—the ones that span beyond just one product—not only have I not seen them report, “Hey, AI’s taking our business. We’re changing our outlook,” but I’ve generally seen them say, “Hey, we integrated AI into our business, and we think it’s accelerating. The outlook looks great. We’re seeing a lot of inbound. Our forecast is coming in at the high end.”

I’d love you to tell me: Am I wrong? Are there any of these bigger software companies—and again, not point solutions, but these bigger, better software companies—where they’re saying, “Hey, AI’s adjusting our outlook”? Has any of that changed? Maybe in February they were saying everything was fine, and then by the end of March—or, we’re talking mid-April—they’re starting to say, “Hey, things are looking great.” Have you seen any of that so far?

It’s funny, right? People joke that this is like asking the barber if you need a haircut. It’s absolutely not lost on me that all these guys are saying, “Hey, AI’s going to increase our business.”

The story I’ve said multiple times is that these were more point-solution businesses. I talked to a few customers of a couple of smaller, not-Salesforce-level software businesses in October, and a customer was like, “Hey, we’re increasing our usage of this product because of AI. Their AI tools are great. We’re increasing our usage.” I was like, “Oh, that’s all clear.”

Then I talked to them 3 months later, and they were like, “No, we stopped using it completely. The AI tools evolved, and now we don’t even need them. We’ve cut them out.” I was like, “Oh, man. If the customers can change that much in 3 months, I have to be really careful here.”

Yeah, that’s so interesting.

Look, I’ll give you 3 personal examples, and then we’ll talk about what the companies were saying in March, et cetera. Very quickly, I use Adobe Lightroom because I take a lot of photos as a hobby, and these are very large, 100-megapixel photos in raw format. So it’s not like your typical point-and-shoot camera.

I need software to pull up this catalog—it’s hundreds of thousands of photos—and I need a way to organize them. I also need a way to develop them, so to speak, because I get them in raw format and need to adjust the lighting, shadows, contrast, and all that stuff. This is real photography, not me taking little photos of my kid with my iPhone.

I got you.

Over the last year, Adobe has been adding a lot of AI features into Lightroom. It’s amazing because now I can go back into my catalog. I’ll give you a silly example: Let’s say you went on some trip and took an incredible photo—an amazing landscape with a monument, whatever—but there’s some pesky person standing there who wouldn’t leave. This tourist has invaded your frame, or maybe there are a dozen of them.

Now you can press one button, and it detects everyone. It detects the distraction. Then you press another button, and everybody disappears. I realize that this is table stakes now, but it wasn’t table stakes a year ago, and it certainly wasn’t table stakes on a raw file that’s 100 megapixels and very high resolution, where you have to keep everything consistent.

This has made my back catalog more valuable because now I can go back into it and reevaluate photos that I would have thrown out but that are now more useful.

A second example is the one I told you about with Excel, where I’m building this internal software now. It uses Excel a lot, so it’s more entrenched. I guess the third example, counterintuitively, is that I recently got out of Bloomberg.

I got out of Bloomberg because I used it primarily to pull data into Excel. The problem is that a lot of their data was wrong. I had a group chat with a bunch of Bloomberg employees. It was me and 19 others—about 20 people—in this group chat.

I kept pointing out errors: “MSCI—these are large companies—your cash flow in Q4 included the whole year’s cash flow in 1 quarter instead of taking Q4.” It was endless mistakes like this. Every time I pointed something out, it would take 2 or 3 days to fix, and they would loop in somebody else from another data team in another country if it was a foreign company. It was very cumbersome.

Now I have a script on my desktop, and I tell Claude to go get the latest financials. I have another script that updates my spreadsheet in the background, and it’s flawless. I can have all the KPIs that the company reports, and I can have everything exactly the way the company lays it out. It’s not in the standardized Bloomberg format, which means you lose a lot of nuance there.

I know exactly what you're talking about.

Now, am I typical? No, because the typical Bloomberg user values the network effect of the chat. But I don't use Bloomberg Chat, right? So I'm not the typical Bloomberg user. My main use case was data, so I'm very happy to have left Bloomberg.

So, I think it's really on a case-by-case basis. Now, back to what the companies were saying: we're on the cusp. We're recording this on April 17. SaaS earnings are going to start next week. We'll see what they say. My guess is that they will still signal optimism, because apparently that's what the sell-side has been saying, right? The brokers and analysts are all saying, "Hey, we're talking to the channel people," and so on. You see a lot of anecdotes from the channel.

The other thing you hear is that large companies—again, picking on PepsiCo, VW, or inserting your big name there—typically look to their existing vendors first. "Hey, I already have this relationship with you. I already pay you millions of dollars a year. Can you agentify this thing for me and make it more useful before I start exploring this other vendor that I don't know, I don't trust, isn't approved, hasn't passed my audits, and all that?"

You hear that over and over again. You hear that from SAP, you hear that from ServiceNow, and so on. It's a competitive world. These companies, I think, are responding and have been responding for a while, and I think the pace is now accelerating.

By the way, what ServiceNow released in the last couple of days with this headless thing is very similar to Openaw. Have you tried Openaw?

I have tried Openaw, and I am tech-incompetent. I'm just completely incompetent, and I had to fall back to Claude Co-Work, which I think is the greatest product ever built. But if you send me a primer on how to use OpenClaw without risking deleting all my files, I would love to get up to speed and do better at it.

Yeah, I'm with you. I tried Openaw many times. The problem I had was that I'm running a local, cheap model—a free model, actually, Gemma 4 31B, which is the most capable of the free Gemma 4 models that Google released a couple of weeks ago.

I'm running it because every time I save something in Excel, it automatically gets picked up by the AI. The AI reads my Excel file and writes a narrative of what changed. It's actually super interesting and useful. But I tried running OpenClaw with this thing, and it does not work as a general driver for Openaw because it hallucinates a lot. I was having these conversations and thinking, "Are you crazy? You just told me this, but it's not this," and so on.

Then I realized this model is very good for certain, narrow tasks, but it's not good in a general way. So what you have to do then is pay the API costs to Anthropic or OpenAI to run an expensive model like Sonnet or Opus to direct Openaw. And that becomes very expensive.

I linked up Sonnet, and within 2 hellos, I was already spending $2 on the API dashboard. I'm like, "This is weird." What I've heard recently is that people are spending $200 a day. The problem is that every time Openaw says hi—every time you say hi—Openaw sends 1 million tokens over, because it sends all your context.

And this gets back to what I was saying about Salesforce. If you look at what Salesforce announced, this context—all it is is just a text file. It's just a bunch of text files. Those text files say things like, "My name is Andrew Walker. I'm an investment analyst, and I like these types of companies. I don't like these types of companies. Whenever you edit my Excel files, never do this. Do not ever touch my system files. Do not touch this folder. Do not delete anything without my approval."

So it has all these guardrails. And that's what Salesforce actually open-sourced, because again, it's just a bunch of files. I think they call it Agent Script. So, yeah, this thing is very interesting because they essentially built Openaw themselves, which is predictable. I think everybody's going to do this.

In fact, I think very soon you're going to see both Windows and Mac ship operating systems with Openaw built in. You're going to have a model running locally, built in. Everything is going to be agentic on your computer and in your operating system. That stuff will talk to Claude Code and to different MCP servers to use all these different tools. It's going to be a wild, interesting future.

But again, going back to our SaaS conversation, I do think it has the opportunity to further entrench the best companies that make themselves so good and improve their products so much that they just serve their customers in a way that makes them ask, "Why am I going to develop my own software?"

And by the way, Andrew, you develop your own software, and then you have to test it, and it's full of bugs. My stuff—I have about 7,000 lines of code—it's nothing, but I find bugs every day. I told Opus 4.6, with high effort, to make no mistakes, right? Did you tell it to make no mistakes and it still had bugs in it?

Many times. And I've told it to do a code review many times. Literally: "Do a full code review. Make sure you have no bugs." Then it runs and says, "Oh, yeah, I found these 3 bugs." I'm like, "Literally the same thing." I do a full code review 3 or 4 times, and it keeps finding bugs.

Then on the seventh day, I get an alert on my phone: something hit a 52-week low. I'm like, "This is wrong." So I go back into Claude Code. It's like, "Oh, yeah, there was a bug."

Let me hard-pivot, though. I want to ask you a different question. We've been talking about software companies. A lot of the software companies that I've looked at or reviewed are on the lower end, as I like to call it. They're point solutions. I've come to view a lot of them as absolutely ripe for displacement by Claude or whatever.

You invest in a lot of better companies, but I'd love to know: in the past 12 or 18 months, is there a portfolio company—or hopefully a former portfolio company—on which you've had a change of heart? You could imagine saying at the end of 2024, for a bunch of companies, "Hey, I don't think AI is coming for this." Then in 2025, saying, "Oh, shoot, AI's coming for this guy."

Is there a company or 2 you can throw out as examples of, "Hey, I think AI is coming for these guys"?

Fortunately, nothing that I own. The ones I like—and again, this is not a recommendation; do your own due diligence—I think ServiceNow is very interesting. I really like the management team there. I really like the product. It has very high retention and is very entrenched.

Salesforce is similar. Believe it or not, I think Atlassian is a good company with very poorly managed finances. I think Adobe is also interesting, and it's extremely cheap. I think most of these are priced for perpetual decline, with the exception of ServiceNow.

Fortunately, I have not owned anything where I said, "I have to sell this now because I think it's a zero, or I think it's in perpetual decline," or whatever.

Are there any that you think are in perpetual decline?

I try not to traffic in them, because I've been burned so many times with value traps and dead money and that sort of thing. So I honestly cannot come up with a name off the top of my head.

But again, thinking about this sort of power-law dynamic, there's such a long tail. A company like Domo—I don't even know what they do, but I thought that they were a SaaS loser 5 years ago. It has a tiny business. I honestly don't even remember what they do.

It is very funny you say Domo because I've always been—and increasingly have been—into the corporate-governance dark arts. I used Lionsgate as an example for a post today, where they've never given their CEO a stock-price-based RSU, PSU, or whatever. Then on Wednesday, they gave him a stock-price-based RSU or PSU for the first time. The CEO has been there for 25 years.

It's Lionsgate, Warner Bros., Paramount, Netflix—I think that's a very interesting signal. Any price-based signal is interesting, but after 25 years, Domo popped up in some of my AI-generated scripts. They said, "Hey, buried in the 10-K, they gave their CFO a change-of-control agreement for the first time."

Their CFO has been CFO since 2024, and he's actually been at the company since 2015. Why did they do that? Well, they're running a strategic process. In early March, they agreed to let all of their directors and employees settle their bonuses in stock instead of cash. Why are they doing that?

So it's just funny you say that, because literally it happened today. The 10-K dropped last night, I believe. Let me go to a different one.

Look, you mentioned that you go to the sell-side and follow these companies closely. You've been to some of these conferences. One thing that I do worry about, because I invested in the busted biotechs before, is stock compensation. I want to talk about stock comp for a second, but we can talk about the valuation component in a second.

I also worry about the turnover and the turmoil that stock comp causes. You give stock compensation to a bunch of very intelligent software engineers, top executives, salespeople—whatever it is...

Your stock is at $100. You give them a bunch of options struck at $100 and some PSUs. And then Monday.com’s stock goes from $300 to $60 as you and I are talking about it. All those options will never vest. All those PSUs are borderline worthless.

Now you’ve got a really interesting complication, right? We can talk about the valuation component of stock comp in a second, but all your engineers feel like they got robbed. A lot of them are going to want to be made whole. And by the way, if you go from a $15 billion company to a $3 billion company, if $300 million in stock comp used to mean that you were diluting yourself like 2% per year—if I’m doing that quick math in my head right—now you’re diluting yourself 10% per year, right?

So all of a sudden, our engineers are way underwater, and we need to go tell them, “Hey, all of you need to take 80% cuts in your stock comp.” That’s a recipe for a lot of turmoil. So I just want to ask you: Are you hearing any turmoil at these companies? Or could you say, “Hey, this is the best of the best we’re dealing with. This is where the engineers really want to be working”?

The answer is, I don’t know, Andrew. The only thing I can point to is that we saw a similar dynamic in 2022, because in 2022 a lot of stocks also went down a lot. Meta famously went down how much—70%? Something like that.

A lot of these companies repriced the comp at the bottom, or repriced the comp to make their engineers whole to a certain extent. I don’t know, and I’m sure we’re going to hear from people correcting me on this, but I don’t know that dilution exploded. I don’t know how they structured this to make it so that it wasn’t as painful.

Or maybe it was painful and we’re seeing that right now. Atlassian has over 100% of its free cash flow going to SBC buybacks. It’s absurd. It’s completely absurd.

But I do believe that they will change things because they have a new CFO now. They were not happy with the old CFO. What I’m told is that they finally—and this is Mike Cannon-Brookes, the co-founder—finally understand that they need to have a proper glide path in a maturing—quote-unquote, maturing—company, meaning that they’re growing up in terms of revenue, to get that stock comp lower as a percentage of revenue over time.

It’s already meaningfully manageable, I think. At Salesforce, it’s about less than a quarter of free cash flow in stock-based compensation.

A quarter. Yeah. Yep.

And I think for ServiceNow it’s more. It’s almost like 57% of free cash flow is actually free. So the rest is stock comp.

But again, I don’t have a good answer for you. I guess we will see. And if we see short covering—a reversal—so, in the beginning of the year, we saw the jaws of semis and software, and apparently there were levered baskets put out to make these bets. If we see short covering and a reversal, maybe this is moot, because we’ll see the prices recover and it won’t hurt as much.

As much. Yeah. Yeah. Some companies—ServiceNow has come out and said, “We will grow without headcount growth,” which I think makes a lot of sense because they’re all bloated. And now, with the added leverage of Claude Code, they should be able to, instead of doing RIFs—reductions in force—say, “We’re going to be able to grow without any net headcount growth, right? If we need different talent, we can shuffle people around.”

I think RIFs are—I mean, this is one of the things you mentioned. I can’t remember if it was in the thing you sent out earlier in this podcast, but it’s not lost on people. You hear all these people saying, “Hey, a lot of our code is written—all of our code for a lot of these really forward companies is written by AI.”

I hear software engineers who say, “I’m not writing code so much anymore as I am a project manager with 5 AI agents under me writing the code, and then I review.” The 10X engineer becomes a 100X engineer.

But the counter to that would be: If 18 months ago, before all of these coding AI agents really started going, you needed—let’s just use an even number—1,000 software engineers, well, if all of your software engineers are 10X better, you definitely don’t need 1,000 anymore, right? That would suggest that if they’re 10X better, you need 100.

Now, maybe you’re growing so much quicker, and you’re so much more productive, and there’s so much more to do. Maybe the answer is 200. Maybe it’s 300, but you don’t need 1,000. So you’re talking about all these companies—I think they are going to have really interesting margin-expansion opportunities.

The counter to that would be the stock comp. I was saying, “Hey, you’ve got to go to the employees and say, ‘Hey, you need to take an 80% reduction in your stock comp because our stock is down so much.’” A lot of employees 18 months ago might not have liked that, but the counter to that—what I’ve heard from people—is, “Hey, it’s not exactly high times to go get hired.”

These guys might swallow a bitter pill that they would not have swallowed 18 months ago, in 2022. So there’s a very interesting push and pull there.

True. On the point of not needing 1,000 engineers, maybe only needing 100 engineers, there’s this Peter Thiel thing where he says, “I love to argue the opposite sides of the same thing,” to see if the argument fits.

Yeah, that’s Charlie Munger’s thing, right? I want to be the person who can argue both sides the best, because then I understand it the best and I can kind of choose.

Exactly. So let’s argue the opposite side of that, right? Instead of saying, “I have 1,000 engineers, they’re 10X better, now I need 100 engineers instead,” you could say, “I have 1,000 engineers that are 10X better. Let me keep them and give them more tokens so that we can accelerate our roadmap, fix more bugs, and crush our competition because we can run faster than anybody else.”

To me, that’s a lot more appealing. And to me, it sounds more right that that’s what happens. Again, in the companies at the tip of that power-law tail—the best companies.

Yes, there might be, if you look at the long tail of businesses in America, people who are firing employees because now AI is doing tasks that that employee did. I do think, in aggregate, though, you will see net job creation as a result of AI.

By the way, this is an interview that Marc Andreessen did recently with Harry Stebbings. Harry is a venture capitalist, and he made the point you made. Then Marc said, “Have you spoken to these software engineers recently?” And Harry’s like, “Yeah, yeah, I have.”

Marc says, “And are they working more or less?” And Harry’s like, “Oh, no, they’re working way more.” You have this AI psychosis, right? You’re probably working way more.

I said it—I recorded a random rambling where I just talked—and I said AI, over the past 60 days or so, makes me feel like I’m taking the Limitless drug. I’m investigating so many more companies, and I’m getting a little bit more comfortable. Maybe you and I should just do a podcast on AI.

I’m getting more comfortable, but I’m investigating so many more companies, and I don’t have to spend all my time doing screening. I don’t do screening like, “Hey, sort the whole market on a price-to-earnings basis.” But it’s handed me the flags, and I don’t have to spend all my time doing stuff.

And you mentioned Excel. I say, “Hey, I want to look at BDCs today. I don’t have to go price all of them by price-to-book myself. AI, send me all the companies on price-to-book. And by the way, I want companies that have insider buying.” Then I can just go spend my time reading that.

It makes me feel like the Limitless drug. And yes, I feel seen—that the software engineers feel like I do, where the tools are so much better. I’m having more fun working. I’m getting a lot more done. And I’m having some of my stuff say, “Dude, dial it back. You need to go do stuff other than sit at the computer and work.”

Yeah, you hit the nail on the head there. I feel exactly the same way. Actually, the other day I woke up and I said, “Oh, is it time to go to work already?” because I’m so excited to go to work. I have this project I’ve been working on.

All right, let’s get up. It’s kind of dark. So I go to the bathroom, leave my phone in the bathroom on the charger, tap the screen, and it’s 2:00 in the morning. I’m like, “Oh no, it’s not time to wake up yet.”

You and I are having the same thing. I wake up and I’m like, “I’ve got all these ideas.” Marcelo, this has been really fun. I’m going to let you go because we’re starting to run long.

But I want one last thing. There were several things in all your write-ups and stuff that I hadn’t thought about, but there was one really interesting thing that I thought beat back a lot of my “Hey, will Claude launch this all?” Can you quickly go through the warranty argument and the maintenance argument for why these companies might have a moat that isn’t so easy?

My vision is, “Hey, we can just vibe-code it, and Marcelo and I can hire a software engineer and try to price out Salesforce at 5% of the cost because we’re a skinny team and we’ll focus on one company.”

Yeah, I think you’re referring to that one “throat to choke” type of argument. When you’re paying—let’s pick on SAP for a second.

When you're paying the SAP invoice, I guess, first of all, you're the chief technology officer of PepsiCo, to pick on them again, and you have this huge SAP installation that spans all of your factories, all of your invoicing systems, all your accounting systems, et cetera. The hurdle for you to replace that is very, very high because it's mission-critical for your business. You're going to have a huge interruption if there's a screw-up.

By the way, I don't know if you remember the cyberattacks that happened, I guess, with Jaguar last year, et cetera. One or 2 hours of downtime for these companies, that's tens of millions of dollars of lost profit.

Jaguar. I mean, for some companies—you think airlines—1 or 2 hours is not just about the revenue. I mean, it's safety. It's really easy to forget about cable companies: 1 or 2 hours of downtime, cool. The hospital's offline.

A lot of these companies you don't even think about, but even companies you might not think of are life or death. I don't want to be too hyperbolic, but if you have downtime because you outsource to someone, it can be a really big deal.

Yeah. So imagine being the CTO of one of these companies that decided not to spend $20 million with Okta to get proper authentication so that people do or do not have permission to access their platform, and now they're facing hundreds of millions of dollars of losses because of downtime and lawsuits and all that because of a cyberattack.

I think it's similar with these mission-critical systems. I'm going to try to save some money by adopting a vibe-coded solution, but in the process, I'm going to jeopardize my business and put a lot at risk. And you're paying SAP not only for the product; you're also paying SAP for liability management, in the sense that now you have an SLA. They guarantee you certain nines of uptime every year. They have the liability if there's downtime, if there are bugs, et cetera. So it's a different game when you're talking about mission-critical enterprise software.

By the way, I'm sure it's similar with you, right? The stuff that we are building for us, it's stuff for us. It's not like the software that I built—as amazing as I think it is—it is not enterprise-quality software that I can go and put on a website and earn money by selling it as a service. No way. I mean, this is crazy. So it would need a lot of work to get over that hurdle.

And this stuff is hard, right, Andrew? Sometimes I tell Claude—and I know that Anthropic, by the way, keeps a list of all the bad words you say to Claude, because I guess they want to see if people are dissatisfied or getting angry. I get angry at Claude so much. I've got Terminator in the back of my mind. Every time Claude does a good job, I say, "Thank you so much, Claude." And I never—I don't curse anyway—but I never say anything mean to him.

No, but what you're hitting on is exactly right. I'm worried about vibe coding and cheap solutions. And for these big companies, especially for things that—you know, their technology budget is big, but their Salesforce budget, just to keep using that, isn't going to make or break them. If it goes down for a little bit, the revenue loss is a disaster in something like SAP. It crushes the business.

You vibe-code it—you, me, and our engineer friend whom we hired for 5% of the cost—and it goes down, we're not going to get that thing up for weeks. SAP goes down, they're going to send 100 engineers, and they're going to have that thing up and fixed for you on a temporary basis, probably. It's going to be up and fixed in 30 minutes, 1 hour, 2 hours, right?

And you're kind of, as you said, paying for that insurance, and I think that's a really interesting moat. It won't work if Claude comes and takes over and creates the greatest thing in every industry for every software category, but it's a really interesting moat against vibe coding.

Marcelo, this has been awesome. I'll give you the last word, but we're quite far over an hour at this point. I've really enjoyed this.

Sorry, Andrew, for going over the hour.

Dude, no, no. I only do it when I really enjoy it, and I was really enjoying the conversation. If there's anything else you want to hit on, happy to do it.

Just the last thing: I think it intuitively makes sense to me. And maybe this is wrong, but we'll see. Again, we'll see. But it makes sense to me that Anthropic had this enterprise software event in late February, where they came out and said, "Look, we're not here to destroy SaaS. We're here to partner with SaaS, with software companies, and become the infrastructure layer."

To me, that makes a lot of sense—the same way that AWS is the infrastructure layer that software as a service and other software businesses run on top of, right? And they have all these primitives. They have S3, which is storage. They have EC2, which is compute. They have databases, et cetera.

I think it makes a lot of sense for Anthropic and OpenAI to be that substrate that provides intelligence—intelligent tokens that plug into software. Now, from time to time, yes, Amazon has released competitive products. I don't know if you know MongoDB. Amazon released DocumentDB. MongoDB's stock crashed and then fully recovered, and it's up a lot.

Again, that's like a free alternative—or not free. By the way, the substrate of MongoDB is open source, and they take that open-source code, make it better, and sell it for money. So, to me, we'll see. We'll see what the industry structure looks like, but it makes a lot more sense that these model companies are going to be more at the infrastructure layer rather than trying to take over applications and compete directly with ServiceNow or SAP, et cetera. We'll see.

No, it makes total sense. I've got to wrap it up here. This was great. Marcelo Lima, thank you so much for coming on, and we're going to have to have you back on to either talk about specific company AI tools we're building, whatever you want to do.

Sounds great, Andrew. Thanks so much for having me. It's a pleasure.

A quick disclaimer: Nothing on this podcast should be considered investment advice. Guests or the hosts may have positions in any of the stocks mentioned during this podcast. Please do your own work and consult a financial advisor. Thanks.