Replit Went From $2.5M to $250M in One Year

In Brief

The winter of 2024 was a nightmare for Amjad Masad, founder and CEO of Replit. The company had just moved into a large, empty office in Foster City. Half the staff was gone after a round of layoffs, and partners had quietly removed the Replit logo from their websites. Twelve months later, Masad told My First Million that annual revenue had gone from $2.5 to $250 million, audited by PwC. The numbers hide a story about what actually changes when an AI agent can build an entire app from one sentence.

---

The Dark Winter

To understand the numbers, we have to go back to the turn of 2023–2024. Replit was ten years old, had just passed three million dollars in annual revenue, and had been stuck there for years. Masad describes the period as ten years of pushing a boulder up a hill, a metaphor he often borrows from Twitch founder Emmett Shear.

Then came the layoff. Replit went from 120 to 60 employees over a few months. Most left voluntarily, Masad says. He struggled to sleep at night, knowing someone would walk up to his desk the next morning with a resignation. Investors flipped from "cut burn" to "spend more." Partners removed the logo. Invitations to the hot Silicon Valley parties stopped coming.

But in a small room inside that half-empty office, a team was working with something different. Masad called it the war room. Walking in, "the mood was intensely different," he says. The rest of the company felt like a funeral. The war room felt like a wedding. The difference was an AI agent that could write code, set up databases, deploy to the cloud, and build apps that actually worked.

---

What Is Replit, Exactly?

A pedagogical pause is in order. Replit is a development environment that lives in your browser. Instead of downloading Visual Studio, installing Python, setting up a database, and configuring a server before you've written a single line of code, you open a tab on the web and you're ready to go. That was the original idea: lower the barrier to programming as far as possible.

The latest addition is Replit Agent, an AI agent that takes over the whole job. You write: "I want an app that lets users find local restaurants with TikTok recommendations." The agent sketches the database, writes the code, plugs in Stripe for payments, and publishes the app to a URL you can share. You haven't written a single line of code yourself.

This is what Andrej Karpathy popularized as vibe coding: you describe what you want, the AI builds it, you test the result without dealing with the code itself.

---

September 2024: The Day Everything Turned

In late summer 2024, Replit tested the agent internally. Masad paid special attention to non-engineers. Jeff Burke, head of partnerships, became the canary. "He's super sharp, but he can't configure Python to save his life," Masad says. After three days of frustrated feedback, Burke finally posted that he'd built something. That's when the team knew they had it.

The product team wanted more polish. Masad refused. He'd picked up a habit from video games: early access. You ship something half-finished, warn users it may break, and let them help shape the product. Replit Agent launched in September 2024 as an "early preview," with a warning that no one should subscribe if they couldn't tolerate buggy software.

Karpathy saw the launch and tweeted that it belonged in the "feel the AGI" category. That's his shorthand for things that make him feel general-purpose AI is on its way. Soon after, research leads at OpenAI and Anthropic reached out. They had not believed their own models were capable of what Replit had demonstrated.

The numbers followed: $1 million in annual recurring revenue on day one, $2 million on day two. Replit earned more in two days than it had in the entire previous year. This is what product-market fit looks like when it hits.

---

"We Are Already in the Singularity"

Later in the conversation, Shaan Puri asks whether Masad thinks the world is months away from a dramatic AI-driven change. The answer is plain: we are already in the singularity.

The term singularity originally comes from physics. Inside a black hole, the usual equations break down, and you can't reason your way to what's happening on the inside. Mathematician and science fiction author Vernor Vinge borrowed the term for technology in the 1990s to describe a point at which development moves so fast it becomes impossible to predict what comes next.

Masad's claim is that we are there now. In 2019 came GPT-2. In 2020, GPT-3. In 2022, GPT-4. One new model generation every two years. Now new models arrive every week, sometimes every day. Each one unlocks new capabilities (in autonomy, coding, cybersecurity, computer use) that haven't yet had time to be turned into products.

He compares models to bundles of potential energy that entrepreneurs have to convert into finished products. The gap between a new model capability and a finished product he calls capability overhang. The world will change, he says, but it's hard to say exactly when. Here he diverges from Anthropic CEO Dario Amodei, who has predicted AI will push unemployment toward 20 percent. Masad expects employment to stay roughly flat. The companies that automate most, he says, tend to grow, and then hire more.

---

Why LLMs Don't Have a Moat

One of the conversation's most interesting moments is when Masad applies Hamilton Helmer's 7 Powers framework to today's AI market. Helmer is a strategy economist and the author of the book of the same name. It's about what gives a company durable competitive advantages: network effects, switching costs, economies of scale, and so on.

Masad's observation is sharp: large language models are effectively a commodity. A good that's easy to swap. Inside a developer tool like Cursor, switching from GPT to Claude to Gemini is literally one click. No model has managed to break away from the others. None have built network effects that lock customers in.

The only natural moat Masad sees is capital. It costs billions to train the next generation of model, and you need enough revenue to cover operations while you write those checks. That means every big tech conglomerate has a shot, every major state can join in, and China can choose to subsidize the market until competitors fall away. The bad news for mid-tier model companies is clear. For entrepreneurs building products on top of the models, the opposite is true: when the platform is a commodity, there's room for many.

---

When Silicon Valley Is No Longer the Center

This leads us to perhaps the most practical observation in the conversation. When software becomes cheap to build, it opens the door to the kinds of businesses Silicon Valley has historically ignored.

"You can actually build a multi-million dollar business without raising venture capital and without growing a large team," Masad says. He cites concrete customers:

• A British developer is building software for ice rink management. He's on his way to $100,000 in annual revenue, and then to a million.
• TryNearby lets local restaurants hire TikTok creators who actually walk in, eat, and create content. Already past $100,000 in annual revenue after a few weeks.
• Spellbook: contract help for lawyers, built on Replit, now used by over 4,400 legal teams in 80 countries.
• MagicSchool: AI tools for K-12 teachers, estimated to be a $500 million business.
• Medvi: Matthew Gallagher's GLP-1 startup, which booked $401 million in its first year on $20,000 in starting capital and zero employees. The New York Times wrote about him. (The FDA also sent a warning letter because much of the marketing turned out to be AI-generated. That part Masad doesn't mention.)

The point is that the most interesting opportunities now lie where traditional software development never quite reached: in a local type of business, an industry niche, a daily problem nobody previously thought big enough to build software around. Masad's own first startup, before Replit, built management software for internet cafés. Local problems he knew well from growing up in Jordan.

---

The Vercel Hack: When Agents Become Targets

Toward the end of the conversation, Masad turns to the Vercel security incident from April 2026. The sequence is instructive for anyone thinking about giving an AI agent broad access:

1. An employee at Context.ai (an AI company that builds an "intelligent contact book") downloaded a Roblox cheat program. The program was infected with an infostealer called Lumma Stealer.
2. Lumma Stealer harvested the employee's Google Workspace password along with access keys for Supabase, Datadog, and Authkit.
3. An employee at Vercel had installed Context.ai via Google OAuth with the "Allow All" permission, granting full access to their work account.
4. With the Context.ai account in hand, the attacker pivoted into Vercel's infrastructure.
5. Inside, they discovered that database secrets were stored in plaintext, unencrypted. Get into one database, and you had access to all of them.
6. The stolen data is now being offered for $2 million on BreachForums by the group calling itself ShinyHunters.

Masad uses the story to argue that both social engineering and cyber attacks have escalated sharply in the AI era. The bots are persuasive. Phishing emails have fewer typos than they used to. And every time you grant an AI agent broad OAuth access to your work account, you open a door an attacker can walk through, without needing your password.

---

What We Take Away

Replit's hundredfold growth is a concrete data point that AI coding agents are no longer theory. They've gone from demo to revenue, and faster than investors could turn around. Masad's philosophical frame (we're already in the singularity) is useful because it reminds us that the next few years are hard to predict. Local apps are a blind spot for most investors. Security around agents is a new attack surface. And every company that hasn't planned for AI agents yet is doing so on the clock.

---

Glossary

---

Sources and resources