We're approaching a certain plateau in how good coding models could get, and so we could not have a more relevant guest than Amjad Masad, co-founder and CEO at Replit, joining us in the hot seat.
Cost question is secondary to the performance question. When you focus on cost is when you reach a certain asymptotic plateau in the S-curve.
The man is reshaping what it means to be a great product manager, product leader, and Replit is one of the leading vibe coding tools alongside the likes of Lovable. The name of the game is just staying 1, 2, 3, 4, 10 steps ahead.
This was an incredible discussion with Amjad, and I'm really excited for you to hear it.
For all intents and purposes, IDEs are dead.
Ready to go.
Amjad, dude, I'm so excited for this. I've wanted to make this one happen for a while. So, thank you so much for joining me once again today.
My pleasure. Great to be here.
We were just chatting about kind of the desert that Replit's been through and like the amazing position today. Did it take a long time for the world to see what Replit was, or did technology take a long time to catch up to your vision of what it could be?
It's more the latter, right? So the core insight that I had even before we started the company, maybe at this point 20 years ago, was that software is much more transformative than we actually think it is. You know, Marc wrote 'software is eating the world' and all of that, but I thought it's transformative for people's lives and for the prospect of wealth creation, wealth distribution, entrepreneurship, all of that. Well, it's sort of my story, you know, the impact it had on my life, you know, as a kid growing up in Jordan, coming to the US and now running a multi-billion dollar company. And I've seen that even earlier than that. When I was 15, I built a little business. I was making hundreds of dollars. And that was an amazing amount of money for me. I took my entire class to McDonald's when it opened up in Amman. And so that was the kind of the core insight. And initially when I came to the US, I worked at Codecademy and the goal was like, let's teach everyone how to code, let's make it as simple as possible. And I started seeing the stories back then, even like if a fitness guy learns a little bit of coding, launches an app on the app store, makes a million bucks, I was like, wow, this is amazing. And so when we started the company Replit in 2016, the goal was how do you make programming more accessible? How to get to a point where there's a billion developers, not just, you know, 20 million developers at the time. And so we started solving one problem at a time, right? Solving the development environment, Replit was the first in-browser IDE, solving the hosting environment, solving the package management, the maintenance, the iteration, the version management, the multiplayer. We solved all these problems, but there's one bottleneck, and that is people don't want to learn how to code. And it took me a long time to really accept that because I became the 'learn to code' guy. And in March 2025, when I was in an interview and I said I no longer think you should learn how to code, it went super viral. People were pissed. How could you say that? And it's just the realization that there are people that are now successful. You know, we were talking about Jason Lemkin earlier from SaaStr. People that are building multi-million dollar businesses solo with no developers, they don't need to learn how to code. They need to learn how to create. They need to learn how to build. And the real unlock wasn't just AI. It was AI that could do actions over long horizon. Right? So we've had AI since 2022 with GPT-3, but the unlock in 2024 was agentic AI, and that was the first glimpse of it. And we had to build a lot of infrastructure around it to make it work.
Quick comment: how much of the magic is what you build versus the performance of models beneath you?
It is a dance, I would say. Think about, you know, Elon Musk when he did self-driving 1.0, I know, before version 13 or whatever where it's end-to-end, they had to write a lot of software, right? They had to write a lot of classical, you know, computer vision type software in order to make it work. But then they went to end-to-end learning. So there's this thing that happens in AI, and as a founder it's really important to understand this: at any given point you have to plug in a lot of holes in order to build the most advanced thing that you can build. So September 2024, we had to build a lot of infrastructure, a lot of guardrails in order to make agents work. March 2025, when we released Agent v2, we had to delete a lot of that code because the models got a lot better at staying consistent. But then you upgrade your vision, right? And so September 2025, we released Agent 3, and that was the most autonomous agent on the market. It was the first to run for hours on end, and again we had to write a lot of software in order to, you know, make it stay on track. And now since OpenAI 4.6, autonomy is built into the model. And so you have to understand what is the model capability and how much infrastructure you have to build in order to make it perform. And you have to stay ahead of that in order for your product to stay the most innovative on the market.
When you look at your model usage today, I've heard you say before when I was obviously listening to your prior shows that you have a preference for Anthropic. What does the model usage look like across different providers today?
Yeah. So Anthropic has been the sort of workhorse for over a year right now. It's like the core agent loop because it can run for a long time coherently. But a few things have changed. Google's Gemini models are the best at price performance, for example, you know, given their price, where do they sit on the Pareto frontier, right? And so for tasks, for example, like tasks like search, we might create a sub-agent that is cheaper and has good enough performance, and we offload that from the main core loop. Right? So now we use, and I wrote this thesis back in 2022, I called it the 'society of models.' Now we use models from every provider. Actually at some point we were sending more tokens to Google than we were sending Anthropic, despite Anthropic being the core workhorse. And so there's this concept of agent labs, right? We talk about AI labs, but there's agent labs, you know, us, Cursor, some of these other companies. Our goal is to start with the user problem. What are we trying to fix? What are we trying to build? And walk back to the technology and use whatever model we need to use. In some cases, we build our own models.
That was going to be my question. So, Amjad, the reason why the show has become more and more successful is because I've done, they ask more and more direct questions. Cursor decided to build their own model. When we look at other players across different verticals, it's always been a mistake to build your own model. Was Cursor's decision to build their own model and your decision to, in some cases, a mistake because the equivalent model performance has just always come alongside you?
It ebs and flows again, like AI is such a changing landscape, the answer will change every 3 to 6 months. So in 2023 when we trained...
But when you're investing your team and your resources then in that, you have to weigh off: is it worth that 3-month advantage?
No. Right. You do, and that's why it makes it so freaking hard and why you have to change your mind all the time, right? Because there's a lead-up time. But the most important thing is optionality. So in 2023 when we're training models, we achieved better coding performance than the state-of-the-art models at the time, GPT-3.5, right? But then since then came out and later Opus, the gap has closed by a lot, and they were spending tens of billions of dollars, if not hundreds of billions of dollars, making agents work, and that would have been a dumb strategy for us to go and try to compete on that. But now I would say the opportunity opened up again for other reasons. The open-source models are getting really good, and we're approaching a certain plateau in how good coding models could get. And so you can use your data to fine-tune a model specifically for your use case. I don't know if you saw, but Intercom yesterday talked about their new model that is better at customer support than the frontier models. And so maybe their model is going to be state-of-the-art for 3 to 6 months, and maybe 6 months from now the models will zoom back ahead.
How do you analyze that? Is it better to be 3 months ahead and spend the money? Like how important is it to be ahead for 3 months?
It's very important. It's, you know, it's a matter of closing a large enterprise deal. Like, you know, we're constantly getting baked off against everything under the sun, right? And we're winning most of our enterprise deals because our product is ahead of the market consistently.
The thing that I always hear is that we're using frontier models to basically set benchmarks. We're seeing where those benchmarks lie and then we're switching to open source to get as close to them as possible with much more efficiency in terms of cost. Is that the future?
It depends, right? Like so cost question I think is secondary to the performance question, right? Especially in a time when it's flush with capital. I think when you focus on cost is when you reach a certain asymptotic plateau in the S-curve and you don't foresee a massive improvement specifically in your domain. Like Intercom might say, you know, we don't predict that models are going to get that much better on customer support for the foreseeable future, therefore, you know, we can focus on building our own, or there's a data flywheel that we can get, or there's a cost advantage that we get. But if you focus on cost at the expense of performance, you're going to lose. And so, you know, it's similar to any era in tech, right? Like it's cloud, it's mobile, it's SaaS, whatever it is, that there are moments of time where the goal is growth and performance and being at the edge, and then when things kind of rationalize, or Uber-Lyft, when things rationalize, you kind of focus on your gross margins.
I think there's this idea, in all transparency, that your Replit, your Base44, your Lovables, and you might kill me for putting you in that bucket, but is it bluntly of $100 that you make, 80 bucks goes to the model providers?
For us, it's not 80. It's way less than that. But it is a significant portion. Like anyone, like even Anthropic, every $100 they make, 60 goes to Nvidia. Like, you know, their margins are public. It's 40%, right? I mean, so and they're also massively subsidizing their service. So really...
Can I be blunt? What are your margins and how have they changed over time?
They changed quite a bit, and there are moments of time, like I said last year, like we were close to profitability. And then we're like, Agent 4 is going to be this massively parallel agent. Agent 4 is the most powerful agent in the world right now. You can create 20 agents and have them all coordinate and run and build your product 10x faster. But then, you know, that's when we work on... So in programming there's this old saying, 'premature optimization is the root of all evil,' because as a programmer, if you think about performance and optimization early on, you're going to make the wrong decisions. So the reason margins keep moving around is like we're trying to build the best product, and then after we build the best product possible, then we look around, we're like, okay, there's so many ways we can optimize, then we go into an optimization period.
When we think about price sensitivity, how do you think about essentially routing different behaviors where some models are actually okay for certain things and some need the frontier? How do you think about intelligent model selection for different functions?
I would say that is the core competency IP of an agent lab, right? If you think of yourself as an agent lab, you have this tacit knowledge first of evaluating models. Like I think about our AI engineers, it's kind of psychologist in many ways. When a model comes out, the first thing they do, they sit down with it for a day or two, right? They play around with it. They plug it in. They're like, 'Okay, what are the limits? What can it do?' And then we plug it in, and that's more like the tacit aspect of it, right? Which is very easy to underestimate. But the reason when a new model comes out, we're able to build state-of-the-art performance, even better than the lab itself. Like we're big partners with Google, and Gemini is one of the best models at design. I would say our products are better at design using Gemini than Google's products. It's because we know how to evaluate these models, how to get the best performance out of them. And then we have a bunch of proprietary benchmarks that we use. And finally, we do a lot of A/B testing as well. So that's really what you're doing as a company that is building on top of foundation models.
How concerned should we be about the proliferation of Chinese models in Silicon Valley companies?
I'm not sure about this question. Like, you know, yeah, I guess, you know, it's a little bit above my pay grade to tell you what are the intentions of the Chinese Communist Party with regards to AI? Is their intention to actually innovate and play in the market fairly? Is their intention to destabilize? Is there an intention to destroy the value of US AI labs? I don't know. I think there are reasons to be careful and there are reasons to be optimistic about the progress. Like what I see from...
Would you have a moral issue with using a Chinese model?
I don't think there's a moral issue per se. Like I don't think they're using slave labor or anything like that to do that. I would have a security issue, especially since we have enterprise customers that depend on us for sensitive data and things like that. So we haven't taken the step yet. I wouldn't preclude it from taking in the future. I would love to see a US corporation investing in open source. It looks like Nvidia is making moves in that regard. But open source is going to be very important for us to actually have a free market around AI, because if we're going to end up in an oligopoly of AI companies, then there's actually an economic theory of how they'll naturally collude on price and prices will not go down as fast as possible. They'll also control how we use these AIs. They'll also not provide everything through the API and keep some of the models for themselves. And so I think that it would be bad if we're in a situation where AGI or AI is only controlled by a few corporations. So open source is going to be very, very important. I would venture to say like maybe the US government should start a consortium of companies that are creating the best open-source, national open-source model so that the market stays competitive.
We mentioned Jason Lemkin earlier. He said something brilliant to me, which is he said inference is the new sales and marketing. Obviously free is a large part of your business and many others in the space. Do you agree with Jason on inference is the new sales and marketing, and how do we think about that?
Yeah, if you think about the hype period that we got with like Claude Code and Codex towards the end of 2025, early '26, a lot of it was driven by how much free tokens they were giving out, how much, you know, every other day they're like, 'Oh, it's 50% more tokens or less rate limits' and things like that. So, it's clear that companies are using inference as a way to loop people in. And I think agentic development is addictive. It's, I think, the better kind of addiction than like social media or passive consumption. It's like creative addiction, which is kind of good. But I think companies are realizing that a way to kind of rope people in is through a lot of free tokens. Now there's a question about retention, but I don't disagree fundamentally with the statement that you can use free tokens to acquire users.
Enterprises will kind of try things in certain respects. I was chatting to Jason Lemkin before, and he said your number one ICP is product teams. As product teams kind of code or vibe more, what do they do okay in 2 to 3 years? How do they think about those functions?
There's a big question about how product teams will look in the future. If I were to make a prediction, I would say that we'll still have engineers inside the organizations. Those engineers are responsible for more infrastructure, AI, ML, embedded systems, you know, more low-level engineering. And then you have product organizations. And product organizations will have people that tilt a little more technical and have people tilt a little more design, have people tilt a little more product, but you're not really calling them anything different. They're product builders, and their responsibility is to figure out what to build next. I agree with Jason that a lot of the ICP right now is product. But one thing I'm really excited about is operations teams, and they're kind of underserved. Like operations teams are sitting at the nexus of a lot of data flow, and typically they'll buy a lot of SaaS software. They're typically not happy with it because there's all these SaaS softwares siloing the data. They try a lot of automation software that doesn't work very well. They have a lot of Excel sheets, a lot of manual work. And so, you know, we see a lot of our customers are building, you know, quote configurators for their sales team, automating their deal desk, automating support operations. And when they use Replit, the return on investment is as good if not greater than product teams. With product teams, you're cutting down product development life cycle, which is hugely valuable. With operations teams, you're actually being more efficient. You're selling more, you reduce headcount, or you need less people.
You said ops have a lot of tools that they're not happy with. That has led to the SaaS apocalypse and a lot of people believing kind of the deterioration of value in a lot of these public companies. Is that overexaggerated or is it a just cause for concern that they have lost the market caps they have?
I'll tell you what we're seeing in enterprise. We're not seeing people rip out Salesforce or Workday or like really fundamental, I think the hype term is, system of record type SaaS tools. And instead they're building on their APIs, and we're creating MCPs and hooks into the HubSpots of the world and Salesforce and all of that. There's also another side where we have this great partnership with Databricks, and people are skipping the SaaS tools entirely and building on top of their data warehouse. So, you know, there's another view on this where the system of record is actually your data warehouse, and this is like a way to be bullish on Databricks and companies in that space.
Well, that goes to a statement though that Jason has said though, which is like then it is justified, because what you're seeing is like the maiming of SaaS company growth. If you have 20, 30% of customers or users who are using Databricks as their warehouse, and that maims that audience for the core SaaS public companies, that's enough to cause the growth to decline significantly.
I would agree with that. I would agree with that. Now, a lot of vertical SaaS is in trouble as well, where, you know, maybe it's you don't think of it as like a system of record, where like there's a lot of survey SaaS software, and we see that getting replaced wholesale with Replit. And then the other side of this, there's a lot of micro-entrepreneurs starting on Replit today, and they're undercutting the price of a lot of SaaS companies out there, especially, you know, kind of vertical point solutions. And then there's another pricing pressure there as well.
What about the traditional incumbents in our space who we're thinking of Squarespace, Wix? What happens to them?
I mean, if you look at the behavior of Wix, they look like they're pouring everything into Base44. So it seems like...
Is Wix now Base44 basically?
I don't know much about their company, but it looks like they're spending a lot on ads, definitely behind Base44.
What did you think of the Super Bowl?
You know, I haven't heard that many people talk about it. I think it's good that people are, you know, getting vibe coding out there and more people are talking about it, but I don't think it's still in the public consciousness.
How do you think about maintenance in this case? You have ops teams building tools, you have entrepreneurs building tools. You got to maintain these...
Enough running a business. I got to maintain this now.
This is where Replit shines, and you know, if you talk to Jason or some of our other customers, Replit goes way further than any other vibe coding product on creating more maintainable software. For example, and part of the reason Replit has been slightly more expensive than others is that we do a code review for every code change that we make. So we spend a lot of tokens on maintenance as much as we spend on creating that software. Replit also has a built-in tester. So if you enable all the power features, whenever the agent writes code, it goes into a testing phase, spins up a browser, tests everything in the app, goes into a code review session, reviews that, kicks it back to the coding agent, gives feedback, you know, the test failed here, the code review is not good. And people enjoy looking at the code review agent because it's kind of a dick. It's like, 'This looks like AI-generated slop.' It'll actually say that, and then it goes back. We're also building agents that are sitting in production software. So we already have security agents right now that are sitting in enterprise deployments and are monitoring activity, and they're monitoring packages, monitoring for supply chain attacks. And so the thing about AI, any problem AI creates, there's more AI that you can build to solve that problem. So I think...
Have you been surprised by how price sensitive people are around security and code reviews?
Not in our segment. I would say in the engineering segment, and we have some engineers use Replit, but 75% are non-engineers. Engineers are more price sensitive because they have a lot of options. They can use a lot of different products on the market. Now, when you're an operations manager using Replit and you just saved $10,000 on a SaaS software, you've saved another, you know, $200,000 on headcount, and you're spending an additional $1,000 to just make sure that the software is more secure. That's like a no-brainer. The ROI has been a hundredfold for companies we work with. On the consumer side, there's more price sensitivity, especially if I'm an entrepreneur just dipping my toes, which is why we reduced the price on our core plan. So, I think there's going to be, and you hinted at that earlier, there's going to be different models for different use cases or different parts of your journey. If you're just starting out, you don't want to be hit with a $1,000 bill. You want to be able to play around with $20, $30 before you commit.
Totally get that. Is core a loss leader? Are you like, 'Hey, I'm willing to lose money on core for the conversion to pay to pro'?
Yeah, it's sort of like become kind of a bit of the new freemium, because you can't totally run an amazing free tier because tokens are still very, very expensive. So, you want people to pay something to recoup some of the losses. But as the percentage of our revenue becomes more enterprise, more pro, then we're willing to, you know, subsidize the cost of core a little more.
Token costs today, how do they change over the next 3 years? Like is it like they get 50% cheaper or like 10 times cheaper?
It's an interesting question because there's different ways to slice it. You can look at the price of intelligence. The price of intelligence has gone down tremendously. It's really hard to quantify it, right? But like everyone's spending way more on OpenAI than they were spending on GPT-4, but they're getting way more intelligence, way more productivity out of that. Pure unit price is not going down as much as people expected it to go. And I think part of it is...
I think part of it is we're living in a world where the true frontier models is actually just between GPT-5 and Claude. And so there's not a lot of pricing pressure now. We're seeing Gemini catch up. We're seeing the open-source models catch up. And so maybe as there becomes more pricing pressure, we'll see token unit token prices go down. Also the other thing is I don't think these companies are all that profitable partly because, you know, the underlying chips are not all that cheap. And there isn't a lot of competition. We have sort of at bottom everyone's running on Nvidia, and there's not much competition there, and Nvidia has amazing margins especially for a hardware company. What is it, 80%?
Yeah, it's pretty phenomenal. What is the world's most valuable company? I'm confused with something. I'm confused with Cursor. Can you help me out? And this is Jason's question, so you can blame him. All the hard questions, just blame Jason. Twitter says Cursor is completely dead. I don't have a single portfolio company that uses Cursor. I don't know anyone that uses Cursor. Sorry. But they just hit two billion in revenue. Is Cursor dead or is this narrative completely...
I'll tell you a few principles. One is the market is so large. The market for software is not just a market for SaaS, although it's part of that. It is a market for... it is an expanding market. There's no existing TAM. You can say this is the TAM we're going after. Some people will say, like I think Sequoia at some point said it's like the labor TAM, right? The knowledge worker labor TAM. That could be that. But it's not like we're displacing labor. We're actually supercharging labor. So companies are becoming more productive. And so it's like an ever-expanding TAM, and it's going to be perhaps as big as the internet, if not more. And that's like I'm talking about just software generation broadly, it's vibe coding, it's AI-powered coding, it's fully agentic coding, coding generation in general. I think it's a hugely valuable market. So I think a lot of... and the world is large, and a lot of different companies will prefer different products. There are people at Replit today that still prefer to review every piece of code, and something like Cursor is good for that because it's still sitting in the IDE, right? So there's the portion of the market that still really likes that. And the other thing is I think Cursor has done well at enterprise sales, and enterprise is a very sticky customer, and once they adopt something, it has to really fall tremendously behind, perhaps GitHub Copilot, in order for people to replace it. And I don't think Cursor fell that far behind, right? They still use the latest agents. They have like a good agent harness. And so I think Twitter is a distortion machine, and Twitter is like the inside of inside baseball. It is the people at the edge of the adoption of AI. So I think if you're constantly on Twitter, which why I think VCs don't have the best information landscape because a lot of them just live on Twitter, but the world is much larger than that.
