Replit Agent Turned AI Coding Into a $250 Million Run-Rate Business

Amjad Masad said Replit went from $2.5 million to $250 million in revenue run-rate in one year, between 2024 and 2025. When Shaan Puri suggested the company might be close to $500 million in annual revenue, Masad corrected the premise upward: “We’re on our way to a billion this year. I’ll just say that.”

That claim landed because Replit had not been compounding smoothly toward the number. Masad said revenue had been “two or three million bucks” for years while the company was still trying to figure out what it should become. Sam Parr asked whether the numbers had been audited. Masad said Replit had passed a PwC audit a couple of months earlier, was gross-margin positive “by a good amount,” and was careful in how it calculated run-rate.

On screen, the discussion was paired with press coverage of Replit’s financing: a Reuters headline reading “AI software developer Replit raises $250 million at $3 billion valuation,” and another article snippet saying the company had secured $250 million and that its valuation was now $3 billion following rapid revenue growth.

The more important point in Masad’s explanation was not that Replit grew quickly, but that the quick growth followed a long stretch of not having the thing. He described product-market fit as something he had heard other founders compare to “stepping on a landmine.” For years, he said, Replit did not have that feeling. That was why the company kept pivoting.

When Replit Agent launched publicly in September 2024, the feeling changed. Masad said the first day produced roughly $1 million of ARR, and the second day $2 million. Parr summarized the implication: in two days, Replit had beaten eight years of prior growth. Masad agreed.

Puri distinguished between companies that push and companies that are pulled. In his phrasing, before product-market fit a founder is pushing a boulder up a hill; afterward, the boulder has started rolling downhill and the founder is sprinting to catch it. Parr cautioned that not every great company gets a sudden “landmine” moment. He used Mars Candy as an example of a business that could have compounded through long execution rather than a discrete explosion.

Masad accepted the distinction and framed Replit’s case as a specific kind of event: a new capability created a new market. He invoked Clay Christensen’s distinction between sustaining and disruptive technology. A sustaining business may innovate incrementally inside an existing market and win through execution. A disruptive or market-creating technology produces a moment where something becomes possible for the first time. Masad placed Replit Agent in that latter category.

His examples of market-creation businesses were internet-era companies such as PayPal, Facebook, and Google: when a capability did not previously exist and then suddenly does, demand can rush in because there is no established substitute.

That is also how he explained what Replit Agent did. AI was already “incredible at writing code,” but code alone did not create software. Someone still had to set up the development environment, install packages, configure a database, debug the code, and deploy the application. Masad said Replit Agent showed that an AI agent could handle the end-to-end path: write code, debug it, create a database, and deploy to the cloud. Even if the product was only half reliable at launch, he believed the 50% of sessions that produced “amazing results” would be enough to show the world something new.

Amjad Masad did not describe the period before Replit Agent as a near-death story in the narrow financial sense. He said he had always been financially responsible enough that the company was not weeks away from missing payroll. The layoff was about extending runway, not avoiding immediate collapse.

The darker part, he said, was watching employees lose belief in him, the mission, and the company. Replit reduced headcount by about a third at first, from roughly 120 to 90, but within about three months it was down to around 60. Most of the additional departures were voluntary. Masad said the layoff burst a bubble. For years he had stood in front of the company and talked about the mission; afterward, the same words began to sound, even to him, as if they might be lies or evidence that he did not know what he was doing.

The setting compounded it. Replit had moved from San Francisco to a large office in Foster City because Masad believed the company was about to scale. He said he was only off by a few months, but in the interim the office was “empty and cold and kind of dark.” Each day he expected that when he got to his desk someone would come over and resign. Each night, whatever sleep he got was marked by wondering who would quit the next day.

Shaan Puri stopped on that point because it matched an experience he said many founders do not articulate. The hardest part was not an abstract metric falling; it was seeing a team stop believing, receiving the “can we meet tomorrow?” message, and knowing a resignation was coming while not having the energy or argument to persuade the person to stay.

Inside the company, however, one room had a different mood. Masad called it the War Room. The group working on Replit Agent believed they were close to a breakthrough. They were dogfooding the product, playing with it, and feeling the possibility of what the AI system could do. Masad described the experience as almost schizophrenic: the company outside that room felt depressed and doubtful; inside it, the team was “super pumped.”

The decisive internal test was not whether engineers could use the product, but whether non-engineers could. Masad watched Jeff Burk, Replit’s head of partnerships, as a proxy for the intended customer: smart, from a consulting background, but unable to configure Python. Burk failed on the first day. He failed again. Then one day he posted that he had been able to make the app. Masad took that as the signal.

The team did not think the product was ready. Masad pushed to launch anyway. During the dark period, he had returned to gaming after years away from it. He had been a serious gamer in Jordan, then abandoned hobbies after coming to the United States to focus on being a Silicon Valley entrepreneur. He bought a Steam Deck as a way to take his mind off the company’s problems and noticed how games framed rough releases: “early preview,” a beta-like state where users expect bugs and give feedback. That became the launch posture. Replit would tell users not to subscribe if they needed polished software, but invite them in if they already subscribed or were comfortable with something buggy.

On September 5, 2024, Masad posted an iPhone-shot office video announcing Replit Agent. The post’s visible text said: “AI is incredible at writing code. But that’s not enough to create software. You need to set up a dev environment, install packages, configure DB, and, if lucky, deploy. Introducing Replit Agent. An automated software developer that places you in creative control.”

A visible quote-post from Andrej Karpathy called it “Very cool” and put it “well under ‘feel the AGI’ category,” adding that making actual apps requires more than code and that automating the surrounding infrastructure would let anyone quickly build and deploy whole web apps. Masad said Karpathy’s reaction was one of the moments that made the post go viral. He also said people inside research firms including OpenAI and Anthropic reached out to say they had not known their models were capable of what the demo showed.

The launch changed the company’s emotional state, market position, and operating problem. Demand had finally arrived; the question became how quickly Replit could catch up to it.

Shaan Puri pointed to Replit’s original 2015 deck as evidence that Amjad Masad had “called your shot Babe Ruth style.” The slide shown on screen was titled “Master Plan” and listed three points: growth by building tools for and signing up teachers and students; building a simple, network- and AI-assisted interface that blurs the distinction between learning and building; and evolving into a platform where people come to learn, build, explore, and host applications.

Masad read the plan as a surprisingly accurate description of Replit’s current shape. The first phase was education: building tools for people learning to code. The second was an AI-assisted interface that blurred learning and building. Masad said that is what Replit had become: users do not need to sit down and learn to code in the old way; they can “vibe code,” build as they go, and pick up skills in the process.

The third point — hosting applications — became more significant than the deck stated. Masad said Replit is not only a development environment, but also a place where applications are hosted. That, in turn, enables monetization and scaling. He said users can already ask Replit to “monetize it,” and the system will integrate Stripe. Soon, he said, users should be able to say “market it,” with the implication that Replit will help with distribution work around the application as well.

That platform ambition matters because Masad framed Replit not as a tool for professional developers alone, but as a way to expand who can create software. The revenue breakout came when the product finally embodied the old deck’s promise. AI was not a feature pasted onto an IDE; it was the mechanism that made the original “learn, build, explore, and host” loop accessible to non-engineers.

He also argued that the platform had matured enough that companies no longer necessarily needed to graduate off of Replit. In earlier years, he said, successful companies that started on Replit often had to migrate away because the platform was incomplete. Now, he said, a full business can run on it.

Masad then offered customer examples, all as his examples and characterizations. He said Medv, which he described as a one-person GLP-1 business covered by The New York Times and as a billion-dollar business, runs a significant part of its stack on Replit. He acknowledged the company was controversial because people objected to its marketing practices, while saying he believed the entrepreneur had fixed many of the issues people reacted to. In Masad’s description, Replit mattered not only for the core application but for day-to-day business automation: spinning up websites, interfaces, agents, and internal tools for vendors and back-office workflows. He described the founder as unusually fast at moving from idea to prompt.

He also cited Spellbook, which he described as a multi-hundred-million-dollar company that started on Replit, and Magic School, which he described as a $500 million business that started on Replit. More recently, he pointed to Try Nearby, a company building influencer marketing for local restaurants and shops. A restaurant can hire an influencer through the platform to physically visit, eat, and post to TikTok or similar channels. Masad said Try Nearby was already over $100,000 ARR within a few weeks.

The pattern Masad emphasized is that cheaper software creation changes what kinds of companies are worth building. Silicon Valley usually approaches problems with hyperscale assumptions. But if software becomes cheap enough to make and maintain, a founder can build a multi-million-dollar company without raising venture capital or hiring a large team. He pointed to “local style” software businesses, including a founder in England building ice rink management software who, according to Masad, was already at about $100,000 in run-rate and on his way to $1 million.

Masad’s advice for finding those ideas was deliberately unglamorous: look around your life for things that are not yet computerized. His own first business was software for internet cafes and LAN gaming. Puri connected that to Paul Graham’s idea that startup ideas come from “living in the future” and building what is missing. Replit itself came partly from Masad’s experience in Jordan, moving between computers in internet cafes and wanting a development environment that lived in the cloud.

Masad added another heuristic: be lazy. In programming, he said, laziness is considered a virtue because it drives automation. Now, he argued, everyone is effectively a programmer. People should notice what they hate doing repeatedly and ask whether AI can automate it. That may improve their own productivity, but it may also reveal a product that many others want.

The platform thesis is not only that Replit lets more people build software. It is that the new builders may find markets too small, local, or operationally specific for traditional venture-backed software companies — and still large enough to support real businesses.

Amjad Masad said the first thing that started to break after demand arrived was people capacity. Companies were already using Replit internally and began asking for enterprise deals. Replit had one person who also did three other things trying to close deals as fast as possible. Masad had never done sales and had to learn what selling actually required.

The timing of Replit’s sales hire mattered. Patrick Purvis, previously a VP of sales at ZoomInfo, came to Replit when the company was still in what Masad called its darkest moment, near a layoff. Masad said he asked Purvis why he wanted to join when “the company kind of sucks.” Purvis told him he had spent the prior years reflecting and wanted to do something meaningful; Replit’s mission of empowering people and democratizing software seemed important. Masad’s reaction was to welcome him while warning that there was no one to sell to.

That changed quickly after Agent. The company’s economics were not those of a foundation-model lab. Masad said Replit was not burning nearly as much cash as model companies. At one point the previous year, profitability was in sight and Replit had “something like 30 years of runway.” Since then, he said, the company chose to spend much more, especially on sales and marketing, because many businesses would need help adopting Replit and AI more broadly.

The internal shape of the company changed with that decision. Replit had about four sales reps at the end of the previous year. Masad said sales could be more than half the company by the end of the current year, a surprising outcome for a founder who came from a technical background.

He likes the shift more than he expected. Consumer growth, in his description, is mediated by A/B tests, hype cycles, virality, press, and other forces that can feel like weather. Enterprise sales is more “contact sport”: effort maps more directly to outcomes. If a deal is at risk, he said, he becomes activated. He will call, show up at an office, and do what is needed. The company rarely loses deals, he said, because he and his co-founder and wife, Haya, have built a highly competitive culture.

Masad described Haya as similarly competitive — someone who lifts and boxes — and said winning a deal feels more real than watching a metric move upward. The satisfaction comes from a contest with a visible opponent and a visible outcome.

The relationship between Masad and Haya also became an operating question. Masad said that if they had not been in Replit together, the difficulty of the company might have driven a wedge between them. Startups consume life energy, he said: they take time, vitality, and attention away from spouses, children, friends, and everything else. Because they were both inside the same struggle, he believes the hard periods made the relationship stronger rather than more distant.

But he also listed concrete risks. A founder-couple cannot make major decisions over the weekend and arrive Monday treating them as settled. They have to be careful not to bring the tone of home into the office or make colleagues uncomfortable. Most importantly, they have to work harder to ensure the company is a meritocracy. Masad said founder-spouses need clear performance standards, high expectations for themselves, open communication, and a willingness to consider that someone else may be better suited to run parts of the company.

The sales buildout was therefore not just a go-to-market add-on. It was Replit’s response to the same market-creation dynamic that produced the revenue jump: if customers suddenly wanted the product faster than the company could serve them, the bottleneck moved from invention to organizational throughput.

When Sam Parr asked whether AI was like December 2019 before COVID — a moment when people sensed a distant problem without understanding how quickly daily life might change — Amjad Masad answered directly: “I think we are in the singularity.”

He used the original concept of the singularity as a point beyond which prediction becomes undefined. Borrowing from physics, he described a black hole as a place where what happens beyond the boundary is not knowable in ordinary terms. Early AI thinkers, he said, applied that idea to technological change: once innovation accelerates sufficiently, it becomes very hard to predict what comes next.

For Masad, the evidence is not just that AI progress is fast, but that the pace itself is accelerating. He pointed to a shift from GPT-2 in 2019, GPT-3 in 2020, and GPT-4 in 2022 — models arriving every couple of years — to a current environment where new models seem to arrive every few weeks, sometimes every day. Each model brings potential shifts in autonomy, cybersecurity, computer use, or other capabilities. Those shifts have downstream applications that entrepreneurs have not yet productized.

His metaphor was potential energy. Models are bundles of capability waiting to be turned into products. That is why he sees the period as unusually explosive for entrepreneurs. The world may change, he said, but not all at once, because there is “capability overhang”: capacities exist before institutions and products fully absorb them.

Parr pressed him to make a bet about employment. Masad declined the strongest displacement view. He said he did not think unemployment numbers would move “a whole lot,” though they might marginally increase. He referenced Anthropic CEO Dario Amodei as having warned about much higher unemployment, while explicitly saying he might be misquoting him. Masad’s own view came from Replit customers: some automate heavily and lay off workers; others automate heavily, make more revenue, and want to hire more people to automate more things and create more products. Net-net, he said, he expects the unemployment question to be roughly neutral.

That view coexisted with his belief that cybersecurity and social engineering risks are rising sharply. AI makes phishing more persuasive, bots better at manipulation, and hacking or privilege escalation easier. Masad cited the Vercel incident as an example, but the operational chain was his account of what happened, and he acknowledged uncertainty about whether it had been planned end to end.

In Masad’s description, an employee at Context.ai downloaded infected Roblox cheat software, after which Context.ai became compromised. Someone at Vercel had installed Context.ai through Google OAuth on their workspace, which Masad described as an entry point into Vercel. He said attackers then escalated privileges and gained database access. When Parr asked what the attackers got, Masad said data, database passwords, and source code, and said he believed the data was being offered for sale rather than used in a ransom demand.

He emphasized that every company has issues and said the point was not to attack Vercel. But he found one detail odd in the incident as he understood it: database secrets, he said, were stored in clear text rather than encrypted at rest, so once an attacker accessed the database, they could reach customer database secrets.

For Replit, Masad said the company is probably a target but “not an easy target.” It had brought in someone from the FBI to brief employees on these kinds of threats. He also noted that even if a threat is not from an adversarial company, it may come from adversarial states, especially where model weights or other high-value AI assets are involved.

Masad’s AI outlook is split: he expects large changes in capability and risk, but not a simple straight line from automation to mass unemployment. In his view, the nearer operating problem for companies is that capabilities are arriving faster than teams can productize, secure, and deploy them.

Asked how the “Game of Thrones” AI war among OpenAI, Google, Elon Musk, and others might play out, Amjad Masad answered through business strategy rather than model benchmarks. He cited Hamilton Helmer’s 7 Powers as a theory of moats and asked whether large language model technology is fundamentally commoditizable.

His working hypothesis is that foundation models have not yet shown the kinds of durable moats that protected earlier technology monopolies. Developers can often switch models easily. In products such as Cursor, he said, changing models can be a one-click choice. That creates a strange competitive condition: labs are always close enough to one another that it is hard to get far ahead. Some of the fighting, in his view, follows from that fact. If a company cannot build an unassailable technical lead, it may seek advantage through courts, government, regulation, or other blocking moves.

By contrast, he said Microsoft’s dominance in the PC era was very hard to unseat because of stronger natural moats such as network effects and economies of scale. He does not yet see equivalent moats around foundation-model companies. The one natural moat may be capital: not just the initial capital to enter, but continuous capital to train the current model, the next model, and the model after that, plus enough revenue and installed base to cover costs.

Even that capital requirement, he argued, may not exclude big companies or governments. He suggested China could act in the foundation-model market somewhat as it did in electric vehicles, subsidizing aggressively in order to compete globally and potentially destroy market pricing.

For entrepreneurs building on top of AI, Masad sees this as favorable. If there were a stable monopoly or tight oligopoly underneath application companies, the platform owners could more easily capture the value and come after downstream businesses. If models remain relatively replaceable, more room remains for application-layer companies.

This argument also explains why Replit’s position is not simply “AI coding tool.” Masad is betting that value can accrue in the workflow around model capabilities: environment setup, deployment, hosting, databases, monetization, enterprise adoption, and the trust required to let non-engineers build real software.

Amjad Masad tied Replit’s story to the cultural setting that made the company possible. He said America, especially liberal America’s vision of itself, is built around tolerance and inclusivity, and he argued there is real truth to that. In New York or San Francisco, he said, an immigrant can feel like anyone else and believe they have as much chance as someone born there.

The more specific advantage of Silicon Valley is idea amplification. Masad said that in many places outside the United States, when someone shares an entrepreneurial idea, the likely response from friends is ridicule or a warning not to stand out. He used to guard ideas not because he feared they would be copied, but because they were fragile. In Silicon Valley, he said, ideas grow when people talk about them. Sometimes they grow too much and become absurd — “let’s start a fucking ship island country” — but he sees that as a better problem than a culture that crushes ideas early.

Shaan Puri added that San Francisco is extreme even within the United States. In a coffee shop, he said, founders pitch investors all day, and the abnormal thing is to have a conventional job. He recalled meeting someone from JP Morgan who introduced himself almost apologetically, immediately adding that he planned to leave.

Masad also emphasized the unusual accessibility of highly successful people in the ecosystem. He said figures such as Marc Andreessen and Paul Graham will engage with obscure, intellectually interesting people, appear on tiny podcasts, or meet young founders at local events. The status reward is less about displaying wealth than about being intellectually stimulated.

That idea of proximity appeared in his reading habits. Masad named Ben Horowitz’s The Hard Thing About Hard Things as an important book, especially for a sales-driven business, because of its stories about deals, losing deals, and hiring salespeople. He also cited Hard Drive, a 1990s biography of Bill Gates, and said reading about Gates, Steve Jobs, and other Silicon Valley figures made him want to be part of that world.

He described ambition as a source of meaning. Looking at friends nearing 40 who had not found the thing they wanted to pursue, Masad said he felt fortunate to have found something to get after. The cost was real — Replit had consumed years, friendships, vitality, and family attention — but the alternative he described was a different kind of crisis: not knowing what to aim at.

For Replit, that culture matters because the company required a founder to keep a long-range thesis alive through years when the numbers did not justify it. The same environment that can inflate ideas also made room for an odd one: a cloud development platform for students that eventually became an AI agent, hosting platform, and sales-led enterprise company.

Sam Parr asked whether Amjad Masad’s life had changed now that he was a billionaire, including whether he feared for his security. Masad said yes, unfortunately: he has protection. But he also said he does not feel like a billionaire. He feels “like a guy” who wakes up in the morning. After some lifestyle upgrades, he can feel like a millionaire; he does not feel like a realized billionaire, in part because he has not adopted a billionaire-level lifestyle and does not like to upgrade too quickly.

The spending that has mattered most, he said, is spending that improves health and saves time. A chef was his clearest example: being able to eat exactly what he wants, eat healthily, maintain weight, and preserve discipline without effort. DoorDash, even when filtered for healthy options, still felt inadequate. He also recommended coaching, though he said it may take a few attempts to find the right person, and emphasized the value of someone who shows up at the door.

But he warned that services can be overdone. Too many people in the house can make home feel like an office or hotel. It also increases exposure to people who may want to leech, scam, or otherwise take advantage. He wants his children to have a normal life.

Parr looked up a billionaire statistic during the discussion, and the screen showed a highlighted Forbes passage saying: “Forbes found 71 billionaires aged 39 or younger who built fortunes (rather than inheriting them) as of December 2023.” Masad reacted with surprise, then said he had mixed feelings. Parr told him that among people under 40, he was probably among the richest thousand ever. Masad said that showed he had not reflected on it much; his impulse was still to keep going and build more.

He clarified that “more” did not mean more money in itself. It meant more success, more impact, and enough reserves to do things he believes are positive. Then he added a more straightforward indulgence: “getting a really kick-ass car” as soon as one has money is, in his view, important. His own purchase was an AMG GT Black Series.

Even here, the operating pattern is consistent with the company story: Masad described wealth less as arrival than as another change in constraints. It buys time, health, security, and optionality, but it also creates exposure, management overhead, and another reason to decide deliberately what the work is for.