CEO Amjad Masad on How Replit Is Changing Who Gets to Build Software | StrictlyVC

Well, it's great to be here. I think it is the truth. It is kind of hard being an independent, smaller AI company that's building on foundation models, and especially if you're burning a ton of cash. I think part of the reporting that came out is that they have like negative 23% margins, and if you're starting to want to invest in training models, you're going to have to burn a lot of money providing a service and burn a lot of money training your models. That makes it incredibly hard to stay independent. For us at Replit, partly because we target a different customer set, we've actually been able to run the business more rationally. We've been gross margin positive for over a year now. We're slightly more expensive, but we provide a lot more. So our audience tends to be mostly non-technical users or non-developers that had previously not been able to create any software. And so we provide a platform end-to-end from the prompt all the way to a deployed application that can scale. We handle the security, which is becoming a big issue. We handle databases, database migration. If you're using tools outside of the kind of Replit ecosystem, you might run into a lot of issues. Recently, just a couple of days ago, there was a startup that lost their entire database because the agent ran commands that dropped the database. We actually ran into all these issues like a year or two ago because we were so early in the field, and we built a lot of these primitives into the platform that makes it so that people are fine paying a premium on Replit in order for us to provide the end-to-end experience. And the other thing is super encouraging as well because there's a lot of doubt in the VC and investor space about, well, are these companies valuable? Is there IP here? Well, turns out companies like us and Cursor and others, in many ways, we know how to use these models better than the foundation labs themselves. There's a lot of benchmarks showing that. For example, there's a benchmark called Design Arena showing that Replit generates the best designs, even better than the foundation labs' design tools themselves. And so there's this tacit knowledge in figuring out how to create products on top of foundation models and how to combine the different models. So if you're on Anthropic, you can only build products with Claude. At Replit, we generate a lot of the code with Claude, but we do code review with OpenAI. We generate the designs with Gemini. We use open-source models for code search. And so this concept of more like an agent lab, a lab that knows how to create agents from different groupings of different models.

That's on the enterprise business, but total is less than that, but it's not too bad.

Close.

So yeah, it's pretty good. And also you just talked about the fact that, I mean, I think what I'm hearing is we're not for sale. I mean, our AI development tools are going to become like a verticalized stack at some point.

I mean like Cursor. I mean, I guess...

Because you can argue Replit is more verticalized in that we provide the cloud hosting, we provide the cloud security, we provide the development environment and the sandboxes, the collaboration tools. And so when an enterprise buys Replit, that's really all they need to buy. They don't need to configure the CI/CD from a different product. And so putting it all together, now vertically integrating down to the model, like I said, has its negatives because now you're incentivized to only just promote that model. And the moment you start using other models, it's seen that your models are not as good, right? So for us, the moment there's a new model on the market that is better at any one thing that we're doing, we're just going to adopt it immediately.

Yeah. I mean, you know, we have amazing partners and they sometimes bring up these topics. I would love for us to stay an independent company. And you know, we've been around for 10 years. You know, before it was even accepted, this idea that you can make apps just from ideas, that's always been our dream. And we were talking about creating a billion developers, a billion software creators back in 2018 when we were at YC, and that was almost like laughable. Like people sometimes actually laughed at that dream. And now that dream is possible. And we kickstarted this entire revolution September 2024 with our agentic coding experience, and it just feels like we can take it much further, and we're going to try.

We are partnered with all three. Our biggest partner, you know, two equal partners are Google and Anthropic. Anthropic is still undefeated on the core agentic loop. Like they have the best tool calling. The agent can stay coherent much longer. GPT is catching up quickly to that, I will say. Google, like the Flash family of models, is just amazing price performance. If you want to get something really fast and cheap, they're actually better than open source. When we benchmark against open source, Google's just with their massive scale, they're just able to achieve amazing results. So we use a lot of that. And also, just the fact that we have our entire stack on Google just makes them a great partner. We think that all three will continue to innovate, and my bet is all three will continue to innovate. But I would also not discount the new labs. You know, Reflection AI is another US company that is coming out with open-source models. We hear great things about them. And honestly, the Chinese models are super impressive, like especially Kimi and now DeepSeek. But Kimi is as good as Anthropic's generation model in January. So it's only three months behind. And at some point, for some use cases, that extra intelligence will start not to matter as much, right? Especially as we get into long-running background tasks. I predict that over the next six months, you're going to be able to run your agent for days and weeks. You're going to be able to, what I'm calling the post-prompting world. You know, we're entering sort of an era where you can give AI goals. You know, the dream has always been for agents is to give them goals. So you can tell the agent, optimize my onboarding and make conversion rate go from 3% to 5%. And the agent can go work for weeks doing A/B tests and then collecting data and reflecting on that. Once you start doing that, it's fine to use perhaps not the latest and greatest because you get more shots on goal. The agent can iterate in the background faster.

Obviously, most of our sales tend to be inbound and organic, very PLG. You know, we've acquired really big customers like Zillow, Meta, and many others purely PLG by just people adopting it and then raising their hand and buying our enterprise plan. Whenever it comes more top-down, they do a bake-off. They bake us off against everything in the field. And we usually win on product. But even in cases where there's one feature or another that might be missing and the deal might look at risk, the moment it hits the CISO and the IT group, it's obvious that Replit is the winner because structurally Replit is built in a more secure way. A lot of the vibe coding tools out there will generate a website and connect it to an external database like Supabase, great company, but it makes the security a lot trickier. So you have to, because the database is open to the public, you need to do something called row-level security, and it's especially harder for vibe coders to configure this kind of stuff. Replit being full-stack, the database built into the project itself, the database not open to the public, and we control the routes and all of that, that just makes the app inherently more secure. And then the infrastructure that we run on itself, we spent 10 years battling crypto scammers, hackers, all sorts of roasters. And so we built a cybersecurity function inside the company that is as good as a cybersecurity startup. And so all our architecture is built on zero trust. We have amazing partnerships with Google that gives us access to internal technology that's very powerful. For example, every time you deploy an app on Replit, we create an entire new isolated project on Google Cloud. So we inherit Google security. And now we started partnering with Databricks and Microsoft. And when you use Databricks with Replit, you build a Databricks app, you deploy it on your Databricks instance. And so this living within your data environment, within your warehouse, for example, Accenture uses that, a big customer of ours, they use it to build client-facing apps inside their Databricks instance. Orban Company, for example, they replaced Tableau, Power BI, and some of these tools with Replit and Databricks. So all these partnerships also give us an advantage.

These consultant groups are actually more important than ever because, you know, getting AI into the enterprise is not that easy. There's a lot of change management that needs to be done that these companies are experts at.