Replit 的 13 个月：从差点死掉到 2.53 亿，创业底层逻辑
Replit's 13 Months: From Near-Death to $253M ARR — The Real Story Behind the 'Lucky' AI Boom
2024 年 5 月，差点死掉
那时候的 Replit，账面上有两千多万用户，但年收入只有 280 万美元。
用户多，但几乎没人付钱。公司造了一个免费的浏览器编程环境，全球几千万学生、爱好者、入门开发者拿去用，却根本养不活团队。这是 SaaS 创业里最经典的陷阱：用户规模和商业价值完全脱节。
那一年，硅谷有很多对话在讨论 Replit 还能撑多久。
四个月后，2024 年 9 月，他们上线了一个叫 Agent 的东西。
你用大白话描述一句想做什么，它自己配环境、装依赖、写代码，最后直接给你一个能打开的网址。
接下来的数字有点不讲道理：收入从 1000 万，13 个月干到 2.53 亿美元，同比增长 2352%。2026 年 3 月，公司估值 90 亿美元。
很多人看到这会说：运气好，赶上 AI 风口了呗。
这个解释不是错的。但它遗漏了最关键的部分。
从约旦到浏览器——一个关于”入口”的执念
创始人 Amjad Masad 1987 年生于约旦安曼，从小家里没有电脑。
他是靠着学校机房的电脑接触编程的。那种感觉他后来反复提起：编程是一种权力，但这种权力被”你得先有台电脑、先装好环境”这道门槛挡在外面。
2011 年，他在 Codecademy 做工程师时，做了一个东西叫 JSRepl——一个完全在浏览器里运行 JavaScript 的工具，给 Codecademy 和 Udacity 的在线课程用。不需要本地安装任何东西，打开网页就能跑代码。
这个”零配置、浏览器即环境”的执念，从此没有变过。
2016 年，他和妻子 Haya Odeh、兄弟 Faris Masad 一起正式创立 Replit。使命只有一句话：让任何地方的任何人都能编程。
8 年建了 5 块积木
Replit 从 2016 年到 2024 年的历史，如果只看用户增长和融资记录，看起来像一家”慢热型”公司。但换个角度看，你会发现它每一年都在往同一个方向添砖。
2016：浏览器 IDE
最核心的那块积木。不是编辑器里加了个浏览器插件，而是把整个开发环境搬进了浏览器——多语言支持、即时运行、代码分享链接。年底拿到 75 万用户，主要是学生和教师。
2018：进 YC（第四次申请）
前三次申请 Y Combinator 全被拒绝。第四次终于进去了。那一年，他们做的最重要的事是把多人实时协作写进了 IDE——最多 4 个人同时编辑同一份代码，能看到彼此的光标，像 Google Docs 一样。
2020：COVID 让用户破千万
疫情让线上教育爆炸，Replit 的用户从百万级跳到千万级。但更重要的是这段时间做的基础建设：定制容器化沙盒，让每个用户的代码在安全隔离的环境里运行——这是后来 Agent 能自动配环境的前提。
2022：Ghostwriter，第一次把 AI 塞进编辑器
2022 年 4 月，Replit 上线 Ghostwriter——AI 代码补全、生成、解释、调试，集成在 IDE 里。当时市场上没太多人注意，大家觉得这只是 GitHub Copilot 的平替版。
但 Ghostwriter 做了一件别人没做的事：它的 AI 助手不是一个独立工具，而是 跑在已经有沙盒、有部署能力的平台里。AI 写的代码，当场就能运行。
2023：彻底重写部署系统
2023 年，Replit 拿到 9740 万美元的 B 轮融资，估值 11.6 亿美元。这笔钱没有用来扩招，主要用来重写了上线系统（Deployments）——一键从代码到可访问的线上地址，包含 SSL、域名、CDN、自动扩容。
这块积木很关键，但当时看不出来。
2024 年 9 月：所有积木一夜咬合
Agent 的核心功能是：
- 接受自然语言 brief
- 理解你想做什么
- 自动配置运行环境
- 写出全套代码
- 部署，给你一个可以打开的网址
翻译一下，这五步对应的是什么：
| Agent 的能力 | Replit 早已建好的系统 |
|---|---|
| 自动配置环境 | 2016 年就有的浏览器 IDE + 2020 年的容器化沙盒 |
| 写代码 | 2022 年的 Ghostwriter AI 引擎 |
| 实时协作修改 | 2018 年的多人编辑机制 |
| 一键部署上线 | 2023 年重写的 Deployments 系统 |
| 给你一个可打开的网址 | 2023 年的域名 + SSL + CDN 全套 |
别人 2024 年从零做 Agent，得先把这五个系统搭出来，才能跑出第一句话。Replit 只是给已经在跑着的系统，装了一个嘴。
Agent 在第一天就能跑通，不是因为运气，是因为五个前置条件全部提前 8 年建好了。
变现模型的切换
Agent 上线同时，Replit 把定价模型从”订阅制”改成了按使用量付费：
- Replit Core 月费 20 美元，包含一定的 Agent 使用额度
- 超出额度按计算量收费
- 企业版自定义定价
这个转变非常关键。之前的问题是：学生用免费账户，不付钱。付费用户愿意付的金额有限。
Agent 改变了价值感知：你给了我一句话，它帮我建了一个能用的 App，这值 20 美元，可能值更多。同时，对企业用户来说，用 Agent 自动化内部工具开发的价值，远不止月费。
收入从 280 万 → 1 亿 → 2.53 亿，用了大约 18 个月。
提炼：这件事真正反直觉的地方
1. 坚持的前提是方向对
Replit 撑过了 YC 三次拒绝、用户增长但不赚钱的几年、硅谷对”在线 IDE”这个方向的长期质疑。但它坚持的方向从 2011 年 JSRepl 到 2024 年 Agent，本质没变过：让编程的门槛消失。
盲目坚持是固执。在正确方向上坚持，是护城河在慢慢变深。
2. 基础设施的价值在危机前看不见
浏览器 IDE、容器沙盒、多人协作、一键部署——这些东西在 2016、2018、2020、2023 年分别完成时，每次都有人问”这个有什么用？谁会在浏览器里写代码？”。
答案是：AI Agent 需要的时候，全都有用。
基础设施的回报不是线性的，是阶跃式的。当某个临界条件被触发（这里是 LLM 能力达到阈值），所有积累同时变现。
3. 使命是撑过”看起来没用”阶段的唯一燃料
Amjad 在约旦没有电脑的童年，塑造了他对”编程门槛”的真实愤怒。这种愤怒让他能在用户多但不赚钱的年代继续建基础设施，而不是转型做更容易变现的东西。
使命感不是 PPT 上的句子，是创始人在最难的季度里，解释为什么继续干的理由。
4. 市场”风口”只给那些早就准备好的人
AI 编程工具的爆发是 2023-2024 年的事。但 Replit 的基础设施是 2016-2023 年建好的。当风来的时候，能飞起来的不是当场开始扎筝的人，是那个筝已经放在风口上等了 8 年的人。
5. 最危险的时刻往往最接近转折
2024 年 5 月的 Replit，是一个”用户数几千万、年收入 280 万”的公司——这是一种非常难受的状态：太大了不好关，太小了撑不住。但就是在这个节点，Agent 上线，三个月后年收入过亿。
创业里有个残酷的规律：很多公司死在离答案最近的地方。坚持到 2024 年 9 月，Replit 就是那个没有死的。
Amjad Masad 在 2026 年 3 月的一篇文章里写道：“未来其实非常人性化（The Future is Actually Very Human）“。他说的是 AI 时代人类的位置，但我觉得这句话也在描述 Replit 自己的故事——不是技术的胜利，是一个约旦少年二十年执念的胜利。
相关链接：
- Replit 官网：replit.com
- 创始人博客：replit.com/blog/author/amjad-masad
- Replit $250M 融资公告：replit.com/news/funding-announcement
May 2024: Almost Dead
Replit had over 20 million users. Annual revenue: $2.8 million.
The users were there. The money wasn’t. The company had built a free browser-based coding environment used by millions of students, hobbyists, and beginners around the world — and almost none of them were paying. This is the classic SaaS trap: user scale and commercial value completely decoupled.
There were conversations in Silicon Valley about how long Replit could last.
Four months later, in September 2024, they launched something called Agent.
You describe what you want in plain English. It configures the environment, installs dependencies, writes the code, and hands you a live URL you can open in your browser.
What followed is hard to believe: revenue went from $10M to $253M in 13 months — 2,352% growth. By March 2026, the company was valued at $9 billion.
Most people hear this and think: lucky timing, caught the AI wave.
That explanation isn’t wrong. But it misses the most important part.
From Jordan to the Browser — An Obsession with the Entry Point
Founder Amjad Masad was born in 1987 in Amman, Jordan, in a family without a computer.
He learned to code using school computers. He’s talked about that feeling repeatedly: programming is power, but that power is locked behind a gate called “first buy a computer, first install an environment.”
In 2011, while working as an engineer at Codecademy, he built JSRepl — a tool that ran JavaScript entirely in the browser, powering exercises on Codecademy and Udacity. No local installation. Open a webpage and write code.
That obsession — zero configuration, the browser is the environment — never changed.
In 2016, he co-founded Replit with his wife Haya Odeh and brother Faris. One-sentence mission: make programming accessible to anyone, anywhere.
8 Years, 5 Building Blocks
Replit’s history from 2016 to 2024, viewed through funding rounds and user growth, looks like a slow-burn company. Viewed differently, you see an organization adding a brick to the same wall every single year.
2016 — Browser IDE: Not a plugin in an existing editor. The entire development environment moved into the browser — multi-language support, instant execution, shareable links. 750,000 users by year-end, mostly students and teachers.
2018 — YC (4th application): After three rejections. That year’s key work: real-time multiplayer editing — up to 4 people writing the same code simultaneously, seeing each other’s cursors, like Google Docs for code.
2020 — COVID pushes users past 10 million: Remote learning exploded. More important than the user growth: the custom containerization sandbox built during this period. Every user’s code running in a secure, isolated environment. The prerequisite for Agent to configure environments automatically.
2022 — Ghostwriter: AI code completion, generation, explanation, debugging, integrated into the IDE. The market didn’t pay much attention. But Ghostwriter did something others hadn’t: the AI assistant ran inside a platform that already had a sandbox and deployment capability. Code the AI wrote could be run immediately, on the same platform.
2023 — Deployments rewrite: $97.4M Series B, $1.16B valuation. The money went primarily into rewriting the deployment system — one click from code to a live URL, with SSL, domain, CDN, and auto-scaling. This brick didn’t look important at the time.
September 2024: All Five Bricks Snap Together
Agent’s core loop:
- Accept natural language input
- Understand what you want to build
- Automatically configure the runtime environment
- Write the full codebase
- Deploy — hand you a URL you can open
Translated:
| What Agent does | What Replit had already built |
|---|---|
| Auto-configure environment | 2016 browser IDE + 2020 containerized sandbox |
| Write the code | 2022 Ghostwriter AI engine |
| Iterate collaboratively | 2018 multiplayer editing mechanism |
| Deploy in one step | 2023 Deployments system |
| Hand you a live URL | 2023 SSL + domain + CDN stack |
Anyone building an Agent from scratch in 2024 had to construct all five systems before they could run the first sentence. Replit just gave a mouth to something that was already running.
Agent worked on day one not because of luck, but because five prerequisites had been built, one per year, for eight years.
The Five Real Reasons
1. Persistence only works when the direction is right. Replit survived three YC rejections, years of user growth without revenue, and persistent skepticism about “browser IDEs.” But from 2011’s JSRepl to 2024’s Agent, the core direction never changed: make the barrier to programming disappear. Stubborn persistence is stubbornness. Persistence in the right direction is a moat getting deeper.
2. Infrastructure value is invisible until a threshold is crossed. Browser IDE, containerized sandbox, multiplayer editing, one-click deployment — each one prompted the question “what’s this useful for?” at the time. The answer: when AI Agent needed them, all of them were useful, simultaneously. Infrastructure returns aren’t linear. They’re step-function jumps triggered by a critical condition.
3. Mission is the only fuel for the “apparently useless” phase. Amjad’s childhood in Jordan without a computer gave him a genuine, personal fury about programming’s gatekeeping. That fury let him keep building infrastructure through the years when users were growing but revenue wasn’t — instead of pivoting to something more immediately monetizable.
4. “Wind” only lifts those who were already ready. The AI coding explosion happened in 2023-2024. Replit’s infrastructure was built 2016-2023. When the wind came, the kite that flew was the one that had been standing in the wind for eight years.
5. The most dangerous moment is often the closest to the turn. May 2024 Replit: 20+ million users, $2.8M ARR — too large to shut down easily, too small to sustain. A brutal position. Three months later: Agent launched, and within a year, revenue crossed $100M. Many companies die closest to the answer. Replit was the one that didn’t.
In March 2026, Amjad Masad wrote a post titled “The Future is Actually Very Human.” He was writing about humanity’s place in the AI era. But the sentence also describes Replit’s own story: not a technology victory, but the vindication of a Jordanian kid’s twenty-year obsession.
Links:
- Replit: replit.com
- Amjad Masad’s blog: replit.com/blog/author/amjad-masad
- $250M funding announcement: replit.com/news/funding-announcement
💬 评论与讨论
使用 GitHub 账号登录后发表评论