Mind · In / Out · In · 文章

Render、Railway、Fly.io 相继完成新一轮融资

后 Heroku 三家:Render 领先一截,Railway 得先修好可靠性,Fly 赌 agent 需要自己的电脑

Render, Railway, Fly.io · 2026-09-18

修好可靠性比等新需求出现更有把握:Railway 的毛病是已知的,Fly 押的需求还没被证明。

Indigo 的结论

Render 靠生产级的可靠拿下了眼前的规模,但估值约是收入的 50–70 倍,买的是增速和前景。Railway 的风险在执行,看它修不修得好可靠性;Fly 的风险在需求,看「给 agent 的电脑」是不是真品类。执行风险比品类风险可控。

怎么读这篇 正文是三家公司自己发的融资公告(Railway 1 月、Render 2 月、Fly 7 月),批注来自 Indigo 9 月做的竞品调研。调研里的收入和估值多是第三方估算和二级市场口径,「12 倍」这类增长数字是公司自己说的:看格局够用,不是尽调。

需要记住的几件事

  1. 三家都不做 GPU、不碰训练,押的是 agent 的执行和落地层:人和 agent 写的海量小服务跑在哪。
  2. 自建和脆弱是同一枚硬币:Railway 自建硬件换来价格和速度,5 月 19 日被 Google 停号,全站黑了 8 小时。
  3. Render 吃的是生产级可靠的溢价,但估值已是收入的 50–70 倍,沙盒底层还落后 Fly 一代。
  4. 别把 DigitalOcean 混进来:它卖要自己运维的虚拟机,Railway 卖把服务器藏起来的平台,只有一小块重叠。

拆解 · 3 步

  1. 01

    Render:拿下规模,押 AI 运行时

    C 轮扩展融资 1 亿美元,估值 15 亿,450 万开发者。下一步做 agent 的一体化运行时:工作流、存储、托管沙盒、AI 网关。 读这一段原文 →

  2. 02

    Railway:B 轮 1 亿美元,要做智能云

    让人和 agent 即时部署、自动修复问题;软硬件都自己造,200 万开发者。通篇讲愿景,几乎不给数字。 读这一段原文 →

  3. 03

    Fly.io:换帅,加码给 agent 的电脑

    D 轮 2500 万美元,Docker 前 CEO 接任;3.7 万家客户里 8000 多家是 agent 原生,最大几家带来的收入一年涨近 12 倍。 读这一段原文 →

什么会让我改口

Railway 下半年把控制面从 GCP 拆成多云、事故明显减少,或者 Fly 的「给 agent 的电脑」成了 agent 采购的标准选项。

怎么读这篇

正文是三家公司自己发的融资公告(Railway 1 月、Render 2 月、Fly 7 月),批注来自 Indigo 9 月做的竞品调研。调研里的收入和估值多是第三方估算和二级市场口径,「12 倍」这类增长数字是公司自己说的:看格局够用,不是尽调。

拆解 · 3 步
  1. Render:拿下规模,押 AI 运行时
  2. Railway:B 轮 1 亿美元,要做智能云
  3. Fly.io:换帅,加码给 agent 的电脑
01

Render:拿下规模,押 AI 运行时

C 轮扩展融资 1 亿美元,估值 15 亿,450 万开发者。下一步做 agent 的一体化运行时:工作流、存储、托管沙盒、AI 网关。

Render 博客,2026 年 2 月 17 日:Render 以 15 亿美元估值融资 1 亿美元

很高兴宣布:Render 完成 1 亿美元的 C 轮扩展融资,公司估值 15 亿美元,CNBC 今天早上已经报道。

本轮由 Georgian 领投,它也领投了我们的 C 轮;我们所有主要伙伴都积极参与,包括 Addition、Bessemer、General Catalyst 和 01A。至此我们累计融资 2.58 亿美元。

Render 现在有超过 450 万开发者,每月新增超过 25 万,已经成为全球增长最快的开发者平台之一。在这段前所未有的超高速增长期,这笔资金将帮助我们服务客户、扩展 Render 的产品。

但这个里程碑不只是把现有的东西做大,更是为下一片前沿提供燃料:打造 AI 应用和 agent 的云端运行时。

过去几个月,软件的创造方式发生了根本变化。代码变得充裕,精干的团队可以即时构思、构建、迭代功能。

但写代码已经跃入未来,发布和扩展代码的基础设施却还停在过去。瓶颈不再是人的想象力或打字速度,而是部署、扩展和可靠性的迷宫。

这个问题在 AI 应用和 agent 上尤其突出。

AI 开发者的天然归宿

传统 Web 应用依赖短暂、无状态的请求—响应循环。AI agent 正相反:它们长时间运行、有状态、分布式。它们需要不设上限的执行时间、复杂的内存管理、持久的文件系统,以及可持久执行的工作流。

在超大规模云上搭这些系统,复杂又昂贵;而面向前端的 serverless 平台,对这种新的执行模式限制又太多。

Render 的架构就是为这个新世界打造的。原生支持长时间运行的进程、私有网络、WebSocket、企业级 Postgres 和 Redis,以及基础设施即代码,Render 已经为 AI 提供了顺滑无阻的底座。

这就是为什么已有数千家 AI 公司迁到 Render。一个典型例子是 Base44,全球最受欢迎的 AI 编程平台之一。他们是这么说的:

「我们用一支很精干的工程团队,就能快得多地交付 AI 功能;Render 的灵活和可靠,也跟上了我们快速变化的需求。用了他们的平台一年多,我确信 Render 就是云的未来。他们最新这一轮让我有机会投资他们的前景,这个机会好到不能错过。」

一体化的 AI 运行时

Base44 和其他公司都面临基础设施四处蔓延的难题。今天,做生产级 LLM 应用的工程师,为了让一个 agent 跑起来,不得不把沙盒、向量库、工作流、存储、可观测性等一堆零散的供应商拼在一起。

Render 要终结这种碎片化。

我们正在打造一个完全一体化的平台,让每个 AI 开发者都能快速、有把握、成规模地把应用推向市场。接下来几个月,我们会为平台加上一批专为这个新时代打造的应用层能力:

Workflows:可持久执行的计算层,编排从 agent 循环到数据管道的一切,为异步计算立下标准。

原生对象存储:与我们的全球 CDN 和运行时深度集成。

托管沙盒:安全、按策略执行代码,覆盖开发和生产。

共享文件系统:为复杂的 agent 编排提供持久的上下文。

AI 网关:可观测性、韧性、成本管理和模型路由。

统一的可观测性:覆盖整个技术栈的链路追踪、指标和日志。

……还有更多,我们会继续向客户学习。

和我们一起打造未来

AI 已经彻底改变了软件开发的生命周期,各家公司正借助它,以前所未有的速度和精简程度做产品。无论是人还是机器,开发者都需要跟得上自己节奏的基础设施;下一百万个好点子,会在为这一代专门打造的云上发布。

这就是我们在 Render 做的事。

未来来得比任何人预想的都快。

02

Railway:B 轮 1 亿美元,要做智能云

让人和 agent 即时部署、自动修复问题;软硬件都自己造,200 万开发者。通篇讲愿景,几乎不给数字。

Railway 博客,2026 年 1 月 22 日:Railway 完成 1 亿美元 B 轮融资,为开发者卸下重担

我们融了 1 亿美元,要卸下构建的重担,从你的软件应用开始。

我们的目标很简单:把一切从你手上拿走,让你只需专注两件事:做美好的东西,或者去体验生活的丰富。

我们的做法是打造全球第一家智能云服务商。

它让你或你的 agent 能即时部署、无限扩展、自动修复问题,以思考的速度发布。

开发、部署、诊断,循环往复。

我们相信平和而强大的基础设施:顺滑地发布应用,有友好的向导帮你扩展,需要介入时也能保持从容。我们相信这是每个开发者应得的。

因为上一代云对你要求太多。它们把你拴在桌前,用 DevOps 压着你去扩展应用,用 FinOps 压着你别破产,用 SecOps 压着你别泄露数据。一旦出事,警报大作,一片混乱,恐慌像病毒一样蔓延。

我们认为这种模式带来焦虑、痛苦和愤怒。我们相信生活不必如此,更好的未来是可能的。

我们从零开始打造了软件和硬件,为开发者卸下重担、解开束缚、释放能量。200 万开发者,还在增加;还有数万家公司,从夫妻店到财富 500 强。

我们希望提供顺滑无阻的体验,让你把创作的念头变成现实。我们相信可以一直处在心流里。

我们向往一个充满丰富滋味、色彩和体验的世界。我们认为,困在糟糕工具的单调里苦熬、被拴在显示器前、和所有人一起消费同一种灰扑扑的粗制内容,很难做到这一点。

在我们看来,一个由灰扑扑的粗制内容构成的未来,既不吸引人,也对人类无益。

我们希望,如果我们能撑起你的应用,你就有更多时间享受生活的丰富体验,而这些体验会化作丰富的创作,你可以分享给世界,最终激励他人。

借助我们正在打造的工具,我们希望让你提高抱负,做出比你想象中人力所能及更大的东西。

这就是我们正在奔赴的未来。

向前,奔向一个滋味丰富的世界。

03

Fly.io:换帅,加码给 agent 的电脑

D 轮 2500 万美元,Docker 前 CEO 接任;3.7 万家客户里 8000 多家是 agent 原生,最大几家带来的收入一年涨近 12 倍。

Fly.io 新闻,2026 年 7 月 24 日:Fly.io 融资 2500 万美元,加码「给 agent 的电脑」

融资、客户采用和新任 CEO Scott Johnston,加速推进 Fly.io 的愿景:专为 AI agent 及其构建的应用打造的基础设施。

旧金山,2026 年 7 月 24 日:AI agent 正在改变软件的构建和部署方式。近十年来,Fly.io 一直帮助开发者构建和运行软件。如今,随着 agent 越来越多地和人一起构建、运维、部署软件,Fly.io 认识到,AI agent 需要的基础设施,和定义了第一波 AI 的那种用完即弃的执行环境有根本的不同。

Fly.io 正在加码「给 agent 的电脑」:为 AI agent 及其构建的应用设计的互联基础设施。Fly.io 给 agent 真正的电脑:有持久的磁盘、能安全连接其他系统,还能扩展到数百万个实例,支撑生产级 AI 应用。

Fly.io 刚刚度过公司史上最强的一个季度,几乎全部由 agent 负载驱动。超过 3.7 万家客户在 Fly.io 上构建,其中超过 8000 家是 agent 原生公司。过去 12 个月,公司最大的几家 agent 原生客户带来的收入增长了近 12 倍。在公司整体最大的客户里,agent 原生公司现在约占收入的三分之二。Firecrawl、Kilocode、Plastic Labs 等 agent 优先的公司,用 Fly.io 构建持续运行、做实事的生产级 AI 应用。

为抓住这个机会,Fly.io 同时宣布完成 2500 万美元 D 轮融资,由 Dell Technologies Capital 和 Intel Capital 联合领投,Andreessen Horowitz、EQT、Geodesic 和 YC 参投。Andreessen Horowitz 普通合伙人 Martin Casado 和 Dell Technologies Capital 董事总经理 Daniel Docter 将加入公司董事会。随着越来越多机构构建和部署生产级 agent 应用,这笔投资将巩固 Fly.io 在「给 agent 的电脑」上的领先。

Fly.io 还宣布,Scott Johnston 出任 CEO 并加入董事会。Johnston 曾任 Docker CEO,也在 Puppet、Loudcloud 和 Netscape 担任过领导职务,在软件开发的几次重大转变中帮助基础设施平台做大。创始人 Kurt Mackey 将留任 Fly.io 董事会,转任顾问,继续帮助塑造公司的长期技术愿景。

Phonic 联合创始人 Moin Nadeem 说:「我们评估过好几个跑 agent 的平台,只有 Fly 把长时间运行、有状态的计算当作一等原语,而不是事后拼上去的东西。能让 agent 环境靠近数据和推理运行,正是我们产品赖以成立的底座。Fly 在打造 agent 生态真正需要的执行层,这就是我们押注他们的原因。」

Andreessen Horowitz 普通合伙人 Martin Casado 说:「每一波基础设施,都由它的主要用户来定义;而这是第一次,这个用户不一定是人。这改变了一台电脑该长什么样。Fly 为这场转变打造基础设施已经好几年了,它给 agent 的是真正的电脑,而不是用完即弃的沙盒。我们很早就相信这条路,也认为这件事将在 Fly 这个平台上建成。」

Fly.io CEO Scott Johnston 说:「每一次平台转变,事后看都显而易见,当下却都有争议;我近距离见过几次,这一次也是同样的形状。agent 是人的新界面,它们需要一台属于自己的真正的电脑。谁把这层基础设施做对,谁就定义下一个十年的计算。Fly 多年来一直朝这个方向建设,我加入是为了帮忙把这个品类做起来,把推动它的公司做大。」

Fly.io 创始人 Kurt Mackey 说:「我们创办 Fly.io 时觉得,开发者不应该在强大的基础设施和好用的配置之间二选一。所以我们做的软件,用起来像真正的电脑。AI agent 的兴起只意味着,我们比以往更需要强大又好用的基础设施;要做到这一点,我们需要高度的专注和速度。Scott 是实现这一点的合适人选。」

Fly.io 打造给 agent 的电脑:为 AI agent 及其构建的应用设计的互联基础设施。不同于用完即弃的执行环境,Fly.io 给 agent 真正的电脑:有持久的磁盘、能安全连接其他系统,还能扩展到数百万个实例,支撑生产级 AI 应用。超过 3.7 万家客户里的数十万开发者,用 Fly.io 支撑生产应用。

判断收口延伸

Indigo 的结论

Render 靠生产级的可靠拿下了眼前的规模,但估值约是收入的 50–70 倍,买的是增速和前景。Railway 的风险在执行,看它修不修得好可靠性;Fly 的风险在需求,看「给 agent 的电脑」是不是真品类。执行风险比品类风险可控。

需要记住的几件事

  1. 三家都不做 GPU、不碰训练,押的是 agent 的执行和落地层:人和 agent 写的海量小服务跑在哪。
  2. 自建和脆弱是同一枚硬币:Railway 自建硬件换来价格和速度,5 月 19 日被 Google 停号,全站黑了 8 小时。
  3. Render 吃的是生产级可靠的溢价,但估值已是收入的 50–70 倍,沙盒底层还落后 Fly 一代。
  4. 别把 DigitalOcean 混进来:它卖要自己运维的虚拟机,Railway 卖把服务器藏起来的平台,只有一小块重叠。

放回主线

证实

约束正从算法层迁到物理层(电力·内存·原子) Railway 自建硬件、Fly 自有服务器:价值从软件抽象往「谁拥有硬件和互连」沉,高毛利和单点脆弱一起来。

补充

archerhume 逆向 Jev/TypeSafe 同是 agent 工程栈里的收费关口:Jev 卡在决策层,Fly 的 Sprites、E2B 卡在执行层那台隔离的电脑。

补充

Ivan Zhao(Notion CEO)钢铁蒸汽与无限心智 Ivan 讲应用层的卡点(上下文碎片化),这三家讲基础设施层的卡点(agent 部署在哪)。

什么会让我改口

Railway 下半年把控制面从 GCP 拆成多云、事故明显减少,或者 Fly 的「给 agent 的电脑」成了 agent 采购的标准选项。

读完了。Indigo 对这篇的判断在这两处:

Mind · In / Out · In · Essay

Render, Railway and Fly.io raise new rounds

After Heroku: Render leads, Railway has to fix its reliability, Fly is betting agents need computers of their own

Render, Railway, Fly.io · 2026-09-18

Fixing reliability is a surer thing than waiting for new demand to appear: Railway's problems are known, while the demand Fly is betting on is still unproven.

Indigo's conclusion

Render won today's scale with production-grade reliability, but at roughly 50–70 times revenue it's priced for growth and prospects. Railway's risk is execution: whether it can fix its reliability. Fly's risk is demand: whether “computers for agents” is a real category. Execution risk is more controllable than category risk.

How to read this The text is the three companies' own funding announcements (Railway in January, Render in February, Fly in July); the notes come from Indigo's competitive survey in September. The survey's revenue and valuation figures are mostly third-party estimates and secondary-market marks, and growth figures like “12 times” are the companies' own claims: good enough for the lay of the land, not due diligence.

What to remember

  1. None of the three does GPUs or training; they're betting on the agent execution and deployment layer, where the flood of small services written by people and agents will run.
  2. Building your own and fragility are the same coin: Railway's own hardware bought price and speed, and on May 19 a Google account suspension took it down for eight hours.
  3. Render earns a premium for production-grade reliability, but at 50–70 times revenue, and its sandbox foundations are a generation behind Fly's.
  4. Don't lump DigitalOcean in: it sells virtual machines you run yourself, Railway sells a platform that hides the servers, and they overlap only in a small corner.

Breakdown · 3 steps

  1. 01

    Render: scale in hand, betting on an AI runtime

    A $100M Series C extension at a $1.5B valuation, with 4.5 million developers. Next comes an integrated runtime for agents: workflows, storage, managed sandboxes, an AI gateway. Read this part →

  2. 02

    Railway: a $100M Series B for an intelligent cloud

    Instant deploys and automatic fixes for people and their agents; software and hardware built in-house, 2 million developers. Almost all vision, almost no numbers. Read this part →

  3. 03

    Fly.io: a new CEO and a bigger bet on computers for agents

    A $25M Series D and Docker's former CEO in charge; more than 8,000 of 37,000 customers are agent-native, and revenue from the largest of them grew nearly 12 times in a year. Read this part →

What would change my mind

Railway moves its control plane off GCP onto multiple clouds in the second half and its incidents drop sharply, or Fly's “computers for agents” becomes the standard option in agent purchasing.

How to read this

The text is the three companies' own funding announcements (Railway in January, Render in February, Fly in July); the notes come from Indigo's competitive survey in September. The survey's revenue and valuation figures are mostly third-party estimates and secondary-market marks, and growth figures like “12 times” are the companies' own claims: good enough for the lay of the land, not due diligence.

Breakdown · 3 steps
  1. Render: scale in hand, betting on an AI runtime
  2. Railway: a $100M Series B for an intelligent cloud
  3. Fly.io: a new CEO and a bigger bet on computers for agents
01

Render: scale in hand, betting on an AI runtime

A $100M Series C extension at a $1.5B valuation, with 4.5 million developers. Next comes an integrated runtime for agents: workflows, storage, managed sandboxes, an AI gateway.

Render blog, February 17, 2026: Render raises $100M at $1.5B valuation

I'm thrilled to announce that Render has raised $100M in an extension of our Series C, valuing the company at $1.5B, as reported this morning by CNBC.

Led by Georgian, who also led our Series C, the round includes strong participation from all our major partners, including Addition, Bessemer, General Catalyst, and 01A, and brings our total funding to $258M.

With 4.5 million+ developers now on Render and more than 250,000 new developers joining every month, Render has evolved into one of the fastest-growing developer platforms in the world, and this capital will help us support our customers and expand Render’s offerings during a period of unprecedented hypergrowth.

But this milestone isn’t just about scaling what exists; it’s about fueling the next frontier: building the cloud runtime for AI applications and agents.

Software creation has fundamentally changed in the last few months. Code is abundant, and lean teams can conceive, build, and iterate on features instantly.

But while writing code has leaped into the future, the infrastructure for shipping and scaling it remains anchored in the past. The bottleneck is no longer human imagination or typing speed; it is the labyrinth of deployment, scaling, and reliability.

This problem is especially acute for AI applications and agents.

The natural home for AI builders

Traditional web apps rely on short-lived, stateless request-response cycles. AI agents are the opposite: they are long-running, stateful, and distributed. They require unbounded execution times, complex memory management, persistent file systems, and durable workflows.

Building these systems on hyperscalers is complex and expensive, while frontend-focused serverless platforms are too limiting for this new execution model.

Render’s architecture is built for this new world. With native support for long-running processes, private networking, WebSockets, enterprise-grade Postgres and Redis, and infrastructure-as-code, Render already provides a frictionless foundation for AI.

This is why thousands of AI companies have already migrated to Render. A prime example is Base44, one of the world’s most popular AI-coding platforms. Here's what they have to say:

“We've been able to deliver AI features much faster with a very lean engineering team, and Render’s flexibility and reliability have scaled to meet our rapidly evolving needs. After using their platform over the last year, I’m convinced Render is the future of the cloud. Their latest round gave me a chance to invest in their trajectory, and the opportunity was too good to pass up.”

A consolidated AI runtime

Base44 and others face complex infrastructure sprawl. Today, engineers building production-grade LLM applications are forced to stitch together a fragmented mess of vendors for sandboxes, vector stores, workflows, storage, and observability, just to get a single agent running.

Render is ending this fragmentation.

We are building a fully integrated platform that enables every AI developer to bring their applications to market quickly, confidently, and at scale. In the coming months, we will augment our platform with application-level capabilities purpose-built for this new era:

Workflows: A durable execution and compute layer to orchestrate everything from agent-based loops to data pipelines, setting the standard for asynchronous computing.

Native object storage: Deeply integrated with our global CDN and runtime.

Managed sandboxes: Secure, policy-driven code execution for development and production.

Shared filesystem: Persistent context for complex agent orchestration.

AI gateway: Observability, resilience, cost management, and model routing.

Unified observability: Traces, metrics, and logs across the entire stack.

…and much more as we continue to learn from our customers.

Build the future with us

AI has already revolutionized the software development lifecycle, and companies are leveraging it to build products faster and leaner than ever before. Builders, whether human or machine, will need infrastructure that matches their cadence, and the next million big ideas will be launched on a cloud that's purpose-built for this generation.

This is what we’re building at Render.

The future is arriving sooner than anyone expects.

02

Railway: a $100M Series B for an intelligent cloud

Instant deploys and automatic fixes for people and their agents; software and hardware built in-house, 2 million developers. Almost all vision, almost no numbers.

Railway blog, January 22, 2026: Railway raises $100M Series B to unburden the builders

We’ve raised $100M to remove the burden from building, starting with your software applications.

Our goal is simple: we want to take everything off your plate so that you can focus on only two things; building beautiful things, or spending time experiencing the richness life has to offer.

We’re doing that by building the world’s first intelligent cloud provider.

One that allows you, or your agents, to deploy instantly, scale infinitely, automatically fix issues, and ship at the speed of thought.

Develop, Deploy, Diagnose, repeat.

We believe in peaceful, powerful infrastructure. A smooth way to ship applications, a friendly guide to help you scale them, and calmness should intervention be required. We believe every builder is owed.

Because the last generation of clouds asks too much of you. They chain you to a desk. They weigh you down with DevOps to scale your app, FinOps to make sure you don’t go bankrupt, and SecOps to make sure you don’t leak data. When things break, alarms sound, chaos ensues, and panic spreads like a virus.

We believe this model causes anxiety, suffering, and anger. We believe life doesn’t have to be this way. We believe a better future is possible.

We’ve built software and hardware from scratch to unburden, unlock, and unleash builders. 2 million, and counting, and tens of thousands of companies, from small mom and pop shops to Fortune 500s.

We aim to provide a frictionless experience to will your creative endeavors into existence. We believe a perpetual flow state is possible.

We aspire to live in a world of rich tastes, colors, and experiences. We think it’s very difficult to do this drudging through the monotony of poor tooling, tied to your monitor, consuming the same greyscale slop as everybody else.

It is our opinion that a future of greyscale slop is neither compelling, nor good for humanity.

Our hope is, if we can support your applications, you will have more time to enjoy life’s rich experiences, and those experiences will translate to rich creative endeavors that you can share with the world to ultimately inspire others.

With the tooling we’re building, we hope to allow you to raise your ambitions, to build things larger than you ever thought humanly possible.

This is the future we’re building towards.

Onwards, towards a world of rich tastes.

03

Fly.io: a new CEO and a bigger bet on computers for agents

A $25M Series D and Docker's former CEO in charge; more than 8,000 of 37,000 customers are agent-native, and revenue from the largest of them grew nearly 12 times in a year.

Fly.io news, July 24, 2026: Fly.io doubles down on computers for agents with $25M

Funding, customer adoption, and new CEO Scott Johnston accelerate Fly.io’s vision for infrastructure purpose-built for AI agents and the applications they build.

SAN FRANCISCO — July 24, 2026 — AI agents are transforming how software is built and deployed. For nearly a decade, Fly.io has helped developers build and run software. Now, as agents increasingly build, operate, and deploy software alongside people, Fly.io recognizes that AI agents require a fundamentally different kind of infrastructure than the disposable execution environments that defined the first wave of AI.

Fly.io is doubling down on computers for agents — connected infrastructure designed for AI agents and the applications they build. Fly.io gives agents real computers that have durable disk drives, secure connectivity to other systems, and the ability to scale to millions of instances to power production AI applications.

Fly.io just closed its strongest quarter in company history, driven almost entirely by agent workloads. More than 37,000 customers build on Fly.io, and more than 8,000 of them are agent-native. Over the past 12 months, revenue from the company’s largest agent-native customers has grown nearly 12 times. Among the company’s largest customers overall, agent-native companies now represent approximately two-thirds of revenue. Agent-first companies like Firecrawl, Kilocode, and Plastic Labs use Fly.io to build production AI applications that run continuously and perform real work.

To accelerate this opportunity, Fly.io also announced $25 million in Series D funding co-led by Dell Technologies Capital and Intel Capital, with participation from Andreessen Horowitz, EQT, Geodesic, and YC. Martin Casado, general partner at Andreessen Horowitz, and Daniel Docter, managing director at Dell Technologies Capital, will join the company’s board of directors. The investment will accelerate Fly.io’s leadership in computers for agents as organizations increasingly build and deploy production agentic AI applications.

Fly.io also announced that Scott Johnston has joined as CEO and a member of the board. Johnston previously served as CEO of Docker and held leadership roles at Puppet, Loudcloud, and Netscape, helping scale infrastructure platforms through major shifts in software development. Founder Kurt Mackey will remain on Fly.io’s board and transition to an advisory role, continuing to help shape the company’s long-term technical vision.

“We evaluated a number of platforms for running agents, but Fly was the only one that treated long-running, stateful compute as a first-class primitive rather than something bolted on,” said Moin Nadeem, co-founder of Phonic. “Being able to run agent environments close to data and inference is exactly the foundation our product depends on. Fly is building the execution layer the agent ecosystem actually needs, and that’s why we’re betting on them.”

“Every wave of infrastructure is defined by who its primary user is, and for the first time that user isn’t necessarily a person. That changes what a computer has to look like,” said Martin Casado, general partner at Andreessen Horowitz. “Fly has spent years building infrastructure for that shift and is giving agents real computers instead of disposable sandboxes. We believed in that path early, and we think Fly is the platform on which this gets built.”

“Every platform shift looks obvious in hindsight and contested in the moment — and having seen a few up close, this has the same shape. Agents — the new interface for people — need real computers of their own. Whoever gets that infrastructure right will define the next decade of computing,” said Scott Johnston, CEO of Fly.io. “Fly has been building toward this for years, and I joined to help build the category and scale the company driving it.”

“When we started Fly.io, we felt devs shouldn’t have to choose between infra that was powerful and a setup that was easy to use. That’s the reason we built software that feels like real computers,” said Kurt Mackey, founder of Fly.io. “The rise of AI agents just means we need powerful, easy infra even more than we did before, and we’re going to need serious focus and speed to get there. Scott is the right person to make that happen.”

Fly.io builds computers for agents — connected infrastructure designed for AI agents and the applications they build. Unlike disposable execution environments, Fly.io gives agents real computers that have durable disk drives, secure connectivity to other systems, and the ability to scale to millions of instances to power production AI applications. Hundreds of thousands of developers across more than 37,000 customers use Fly.io to power production applications.

Where Indigo landsFurther

Indigo's conclusion

Render won today's scale with production-grade reliability, but at roughly 50–70 times revenue it's priced for growth and prospects. Railway's risk is execution: whether it can fix its reliability. Fly's risk is demand: whether “computers for agents” is a real category. Execution risk is more controllable than category risk.

What to remember

  1. None of the three does GPUs or training; they're betting on the agent execution and deployment layer, where the flood of small services written by people and agents will run.
  2. Building your own and fragility are the same coin: Railway's own hardware bought price and speed, and on May 19 a Google account suspension took it down for eight hours.
  3. Render earns a premium for production-grade reliability, but at 50–70 times revenue, and its sandbox foundations are a generation behind Fly's.
  4. Don't lump DigitalOcean in: it sells virtual machines you run yourself, Railway sells a platform that hides the servers, and they overlap only in a small corner.

Back on the long-running theses

confirms

Constraints are shifting from algorithms to physics (power, memory, atoms) Railway's own hardware and Fly's own servers: value sinking from software abstractions toward who owns the hardware and interconnect, with high margins and single points of failure arriving together.

adds to

archerhume: Jev’s Architecture Unmasked Both are tollgates in the agent engineering stack: Jev holds the decision layer, while Fly's Sprites and E2B hold the execution layer, the isolated computer itself.

adds to

Ivan Zhao (Notion CEO): Steam, Steel, and Infinite Minds Ivan's chokepoint is at the application layer (fragmented context); these three are fighting over the infrastructure layer (where agents get deployed).

What would change my mind

Railway moves its control plane off GCP onto multiple clouds in the second half and its incidents drop sharply, or Fly's “computers for agents” becomes the standard option in agent purchasing.

Finished. Indigo's take on this piece is in two places: