# Modal 年化收入突破 10 亿美元并完成 7.5 亿美元 D 轮融资
> 原标题:Modal crosses $1B in Annualized Run Rate and raises $750M Series D
- 来源:Modal 官方工程博客
- 发布时间:2026-10-14T00:00:00.000Z
- AIHOT:https://aihot.news/items/x5umoaa0xhd8n76uysjppz2fa
- 原文:https://modal.com/blog/1-billion-annualized-run-rate-series-d
## 摘要
Modal 宣布年化收入突破 10 亿美元,并完成 7.5 亿美元 D 轮融资,投后估值 157.5 亿美元,由 Accel 领投,Kleiner Perkins 及现有投资方跟投。
## 推荐理由
作者作为当事方给出融资规模、估值和业务数据,并说明自研底层技术栈与新产品线,读者可据此了解 Modal 的增长轨迹。
## 正文 · 中文译文
就在五个月前,当我们宣布 C 轮融资时,我写道[需求正在加速](https://modal.com/blog/modal-series-c)。自那以后,使用量增长了两倍,我们的同比增速超过 15 倍。我们最近突破了 10 亿美元的年化运行率,且丝毫没有放缓的迹象。
## 软件能做到几年前还不可能做到的事
很明显,AI 开启了一整套全新的软件能力。如今存在着许多以前根本无法构建的惊人应用。旧的技术栈并不太适合它们。它们需要不同的构建模块,以及一种全新的基础设施思维方式。
我们正在为大量这类应用提供支持。其中包括 TypeSafe AI 的 Jev,它是有史以来增长最快的产品之一。它在头三天里就在 Modal 上处理了 [1T tokens](https://modal.com/blog/typesafeai-case-study),而 TypeSafe 团队则专注于交付,而不是管理基础设施。
## 而且这正发生在许多不同的垂直领域
DoorDash 的 AI 平台团队使用 Modal 将公司各处的应用从实验推进到生产。Cognition、Ramp、Linear 和 Hebbia 正在构建各自领域内的后台智能体专家。Runway 正在世界各个地区生成逼真的视频。Physical Intelligence、Generalist AI 和 1X 正在训练机器人。Chai Discovery 和 Achira 在 Modal 上运行药物发现,通过 [Claude Science](https://modal.com/blog/modal-integration-brings-scalable-compute-to-claude-science) 开展工作的科学家们也是如此。
## 每一次转变都需要新的基础设施
软件的每一次转变都会改变人们能构建什么,以及构建它所需的基础设施。我从九十年代初就开始写软件,这并不是第一次阶跃式变化:我们经历过 Web、云、大数据、移动,以及其他几次转变。
当云刚起步时,你突然不得不应对分布式系统、分片和配置管理。我们原有的工具都不是为此而构建的,所以开发者创造了新的工具:Docker、Kubernetes、Terraform 等等。
AI 带来的摩擦如今已经相当明显。GPU 稀缺,模型相当大且加载缓慢。LLM 现在会写代码了!而且它必须在某个地方运行,与其他一切隔离开来。当然,你可以在旧技术栈上为所有这些考量进行构建,很多人也确实这么做了,但最终你会陷入 YAML 地狱和 GPU 采购炼狱。
所以我们深入了下去。我们抛弃了大部分现有技术栈,在一个如今横跨 20 多个云的容量池上,从零开始构建了自己的文件系统、容器运行时、调度器和网络。
## 开发者需要的是更好的构建模块
软件之美在于可组合性。如果你拥有合适的构建模块——通用、久经考验且可扩展——你就能构建出无限组合的东西。
函数是我们的第一个构建模块:你编写代码,Modal 来运行它。事实证明这非常有用,尤其是对生成式 AI:用扩散模型生成图像、Suno 的音乐、用 vLLM 和 SGLang 进行 LLM 推理,以及模型训练。
我们在 2023 年推出了 Sandboxes,当时我们注意到人们正在使用 LLM 生成代码。Sandboxes 爆发式增长,三年后,它仅凭自身就成为一个年运行率 3 亿美元的产品。用例范围从后台智能体,到强化学习、vibecoding 平台,等等。我们已经启动了近 30 亿个 sandbox,它们可以扩展到[百万级并发](https://modal.com/blog/scaling-to-1-million-concurrent-sandboxes-in-seconds)。
## 我们还将推出更多
一个全新的软件能力世界需要一套新的构建模块,而我们才刚刚开始。
我们最近推出了 [多节点训练](https://modal.com/blog/modal-clusters-generally-available) —— 一个装饰器就能为你提供一个完全 RDMA 连接的集群,无需容量承诺。我们正在向上层推进,并推出了一系列 RL 和后训练方面的产品。
在推理方面,我们一直在推出 [Modal 服务器](https://modal.com/blog/serverless-servers),这是更适合 HTTP 服务器的解决方案,以及一些围绕 LLM 服务的非常酷的功能,例如 [端点候选](https://modal.com/blog/endpoint-candidates)。
借助 Sandboxes,[虚拟机](https://modal.com/blog/vm-sandboxes-agent-computers)现已正式可用。
这些只是未来众多进展中的一小部分。
## 我们筹集了更多资金
五年前我们创立 Modal 时,我无法想象今天在上面运行的绝大多数应用。同样难以想象两年后人们将创造的软件,但我确信的是,它将需要一套全新的构建模块。我们希望其中很多都能在 Modal 上运行。
为了继续构建未来,我们完成了 7.5 亿美元的 D 轮融资,投后估值 157.5 亿美元。本轮由 Accel 领投。Kleiner Perkins 加入,我们所有现有投资者均参与其中,包括 Menlo Ventures、General Catalyst、Redpoint、Lux Capital 和 Bain Capital Ventures。
我显然有偏见,但我想不出比现在更好的构建软件的时机。
_如果你有兴趣加入我们位于纽约、旧金山和斯德哥尔摩的 200 多人团队,请查看我们的[开放职位](https://modal.com/careers)。_
## 正文 · 原文
Just five months ago, when we announced our Series C, I wrote that [demand was accelerating](https://modal.com/blog/modal-series-c). Since then, usage has tripled and we’ve grown more than 15x year-over-year. We recently crossed $1B in annualized run rate with no sign of slowing down.
## Software can do things that were impossible just a few years ago
It’s clear that AI has unlocked a whole new set of software capabilities. So many incredible applications exist now that couldn’t have been built before. The old stack doesn’t quite fit them. They need different building blocks and a new way to think about infrastructure.
We’re powering a ton of these applications. Among them is TypeSafe AI’s Jev, one of the fastest-growing products ever. It served [1T tokens on Modal](https://modal.com/blog/typesafeai-case-study) in its first three days, and the TypeSafe team stayed focused on shipping instead of managing infrastructure.
## And it’s happening in many different verticals
DoorDash’s AI platform team uses Modal to get apps across the company from experiment to production. Cognition, Ramp, Linear, and Hebbia are building background agents that are experts in their fields. Runway is generating life-like video, in every region of the world. Physical Intelligence, Generalist AI, and 1X are training robots. Chai Discovery and Achira run drug discovery on Modal, as do scientists working through [Claude Science](https://modal.com/blog/modal-integration-brings-scalable-compute-to-claude-science).
## Every shift needs new infrastructure
Every shift in software changes both what people can build and the infrastructure they need to build it. I've been writing software since the early nineties, and this isn't the first step-change: we've had the web, the cloud, big data, mobile, and several other shifts.
When the cloud started, suddenly you had to deal with distributed systems, sharding, and config management. None of our original tools were built for it, so developers created new ones: Docker, Kubernetes, Terraform, etc.
The friction with AI is pretty clear by now. GPUs are scarce, and models are quite large and slow to load. LLMs write code now! And it has to run somewhere, isolated from everything else. Sure, you can build for all of those considerations on the old stack, and plenty of people have, but you end up in YAML hell and GPU procurement purgatory.
So we went deep. We threw out most of the existing stack and built our own file system, container runtime, scheduler, and networking from the ground up, on a capacity pool that now spans 20+ clouds.
## What developers need is better building blocks
The beauty of software is composability. If you have the right building blocks — general-purpose, battle-tested, and scalable — there’s infinite combinations of things you can build.
Functions were our first building block: you write code, and Modal runs it. This turned out to be incredibly useful, especially for Gen AI: image generation with diffusion models, music at Suno, LLM inference with vLLM and SGLang, and model training.
We launched Sandboxes in 2023 when we noticed that people were using LLMs to generate code. Sandboxes exploded and three years later, it’s a $300M run rate product just on its own. Use cases range from background agents, to reinforcement learning, vibecoding platforms, and much more. We’ve spun up close to 3 billion sandboxes, and they scale to a [million concurrent](https://modal.com/blog/scaling-to-1-million-concurrent-sandboxes-in-seconds).
## We’re launching a lot more
A new world of software capabilities requires a new set of building blocks and we’re just getting started.
We recently launched [multi-node training](https://modal.com/blog/modal-clusters-generally-available) – one decorator gives you a fully RDMA-connected cluster with no capacity commitment. We’re moving up the stack and launching a number of products in RL and post-training as well.
On the inference side, we’ve been launching [Modal servers](https://modal.com/blog/serverless-servers), which is a better solution for HTTP servers, as well as some really cool features around LLM serving, such as [Endpoint Candidates](https://modal.com/blog/endpoint-candidates).
With Sandboxes, [virtual machines](https://modal.com/blog/vm-sandboxes-agent-computers) are now generally available.
These are just a few of the many more developments to come.
## We’ve raised more capital
When we started Modal five years ago, I couldn't have imagined most of the applications running on it today. It's just as difficult to imagine the software people will be creating two years from now, but what I do know is that it's going to require an entirely new set of building blocks. We want a lot of it to run on Modal.
To keep building what's next, we’ve raised a $750M Series D at a $15.75B post-money valuation. Accel led the round. Kleiner Perkins joined, and all of our existing investors participated, including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, and Bain Capital Ventures.
I'm obviously biased, but I can't think of a better time to be building software.
_If you’re interested in joining our team of 200+ across NY, SF, and Stockholm, check out our [open roles](https://modal.com/careers)._