高通买的不是AI公司,是二十年磨出来的编译器Qualcomm Isn't Buying an AI Company, It's Buying a Compiler Forged Over Twenty Years


AI free token/AI免费积分

 

6 月 24 号,高通宣布收购 Modular。

一家做 AI 软件基础设施的公司。

你可能没听过这个名字。但创建这家公司的人你大概率知道。Chris Lattner。

2000 年他在伊利诺伊大学读博,写了一个叫 LLVM 的编译器框架。今天几乎所有现代编译器都站在它上面。光苹果的 Swift 语言就整个架在 LLVM 上,后来 Lattner 亲自去苹果把 Swift 从立项做成了苹果官方语言,待了十二年。Rust 和谷歌 MLIR 也都站在 LLVM 的肩膀上,但那是别人的故事了。

中间他短暂跳过特斯拉和谷歌,都不太顺。在谷歌做的 Swift for TensorFlow 最后没成。2022 年自己出来创办 Modular,这次要造的东西叫 Mojo,Python 语法,C 的性能,给 AI 开发者用。

二十年来这个人做的事情可以概括成一句话。让代码不绑死在特定硬件上。

现在高通把他连人带公司一起买了。

图片

Modular 在做什么

AI 模型要跑在不同硬件上,底层指令集完全不同。英伟达 GPU 用 CUDA,AMD GPU 用 ROCm,高通 NPU 有自己的指令集。同一个模型换个芯片就得重写底层代码,这件事是英伟达最深的护城河。CUDA 的迁移成本,我之前写 AMD 的时候聊过,不重复了。

Modular 想做的就是把这层成本抹平。一套软件平台,模型写一次,CPU、GPU、NPU、ASIC 都能跑。官方叫「write once, run anywhere」。

Java 当年也是这个承诺。但 AI 时代做这件事的难度和意义都完全不同了。

高通为什么需要它

高通的芯片卖得不少。手机端 Snapdragon 占据安卓旗舰大半江山,笔记本端 Snapdragon X Elite 也在往里挤。

但高通一直有个短板。软件生态。

手机时代这个短板不太致命。安卓系统帮忙兜了一层,开发者不需要直接跟芯片打交道。

AI 时代不一样了。大模型推理对底层性能的要求是手机应用的几十倍。开发者选什么硬件,直接取决于这个硬件上的软件栈好不好用。

英伟达的 CUDA 生态太强了,强到所有 AI 开发者默认就是 CUDA 的用户。高通想让自己的芯片在 AI 推理市场有竞争力,就得让开发者能轻松把模型搬到自己的硬件上来。

自己建软件生态太难了。英伟达花了二十年。高通等不起,那就买。

高通 CEO Cristiano Amon 在收购公告里说了一句话。

The future belongs to developer-friendly, horizontal platforms that can run across diverse compute environments and give customers real choice.

翻译过来就是,未来的 AI 软件平台必须能跨硬件跑,不能绑死在一家芯片上。

这话是在直接挑战 CUDA 的地位。

三种挖墙脚的方式

最有意思的地方来了。

过去半年我一直在看芯片公司在 AI 时代怎么布局。英伟达、AMD、高通,三家在软件生态上的打法完全不同,但解决的是同一个问题。怎么让 AI 代码跑在自己的硬件上。

英伟达是建墙的人。

CUDA 二十年积累,从高校实验室到互联网公司,所有 AI 开发者默认就是 CUDA 用户。现在黄仁勋在推的 DSX 更进一步,把软件、硬件、运营三层全部打包成 AI 工厂的完整蓝图。CUDA 是让开发者离不开你,DSX 是让企业直接买你整套方案。

建墙的人最不怕的就是迁移成本。你跑得好好的,搬一次试试。

AMD 选了填坑。

5 月上海开发者日,苏姿丰往 vLLM 开源社区捐了一套 360 万美元的服务器集群。vLLM 是全球最大的开源推理框架,DeepSeek 和 Kimi 底层都在用。捐给 vLLM 等于把硬件直接塞到了全世界最多推理开发者的手里。同一天放出了 vLLM-ATOM 零代码迁移插件,CUDA 推理服务不用改一行代码就能跑在 AMD 的卡上。先跑起来再说。

高通是买桥的人。

不自建,不捐赠,直接花钱买一座现成的。Modular 的软件平台加上 Snapdragon 的芯片,等于从边缘到云端有了一套完整的 AI 基础设施。

三种打法。建墙、填坑、买桥。

同样的目标,让 AI 开发者用自己的芯片时不用从零开始。

哪种能赢我不知道。但有一件事可以确定。芯片公司光卖硬件的时代结束了。没有软件生态的芯片公司,在 AI 时代活不下去。

这次会不会不一样

大公司收购小公司,最难的从来不是技术。Chris Lattner 在谷歌做 Swift for TensorFlow 的时候,技术很好,组织不支持,项目最后没成。

但 Lattner 二十年来一直在做同一件事。让代码不绑死在特定硬件上。LLVM 是,Swift 是,Mojo 也是。苹果拿走了他十二年的时间做出 Swift,谷歌拿走了他 MLIR 那几年,现在高通拿走了 Modular。

他的技术每次都变成了大公司的基础设施。这一次,大概也不会例外。

但至少这一次,他拿到了规模。每年几亿颗 Snapdragon 芯片的出货量,是 Mojo 语言最缺的那块拼图。故事能不能跑起来,就看高通给不给够时间。

图片



Macaron 🧁 | 软件才是芯片公司的入场券

 

 

On June 24, Qualcomm announced the acquisition of Modular.

A company building AI software infrastructure.

You may not have heard of this name. But you most likely know the person who founded it. Chris Lattner.

In 2000, while pursuing his Ph.D. at the University of Illinois, he wrote a compiler framework called LLVM. Today, almost every modern compiler is built on top of it. Just Apple's Swift language is entirely built on LLVM; later, Lattner personally went to Apple to take Swift from inception to becoming an official Apple language, staying there for twelve years. Rust and Google's MLIR also stand on LLVM's shoulders, but that's someone else's story.

In between, he briefly worked at Tesla and Google, neither going very smoothly. The Swift for TensorFlow project he worked on at Google ultimately didn't succeed. In 2022, he left to found Modular himself; this time, the thing they are building is called Mojo—Python syntax, C performance, for AI developers.

What this person has done over the past twenty years can be summarized in one sentence: making code not tied to specific hardware.

Now Qualcomm has bought both the person and the company.

图片

What Modular is doing

AI models need to run on different hardware, but the underlying instruction sets are completely different. Nvidia GPUs use CUDA, AMD GPUs use ROCm, and Qualcomm NPUs have their own instruction sets. Switching a model to a different chip requires rewriting the underlying code; this is Nvidia's deepest moat. The migration cost of CUDA was discussed when I wrote about AMD previously, so I won't repeat it.

What Modular wants to do is level this cost. A single software platform where a model is written once and can run on CPU, GPU, NPU, and ASIC. The official term is "write once, run anywhere."

Java made this promise back in the day. But the difficulty and significance of doing this in the AI era are entirely different.

Why Qualcomm needs it

Qualcomm sells quite a few chips. On the mobile side, Snapdragon dominates more than half of the Android flagship market; on the laptop side, Snapdragon X Elite is also making its way in.

But Qualcomm has always had a shortcoming: software ecosystem.

In the mobile era, this shortcoming wasn't too fatal. The Android system provided a layer of protection, so developers didn't need to deal directly with the chip.

The AI era is different. Large model inference demands tens of times the underlying performance of mobile apps. What hardware developers choose depends directly on how good the software stack on that hardware is.

Nvidia's CUDA ecosystem is too strong, so strong that all AI developers are by default CUDA users. For Qualcomm to make its chips competitive in the AI inference market, it must enable developers to easily port models to its hardware.

Building a software ecosystem from scratch is too hard. Nvidia spent twenty years on it. Qualcomm can't wait, so it buys.

Qualcomm CEO Cristiano Amon said one sentence in the acquisition announcement.

The future belongs to developer-friendly, horizontal platforms that can run across diverse compute environments and give customers real choice.

Translated, it means future AI software platforms must be able to run across hardware and cannot be tied to a single chipmaker.

This is a direct challenge to CUDA's status.

Three ways to poach

Here comes the most interesting part.

Over the past six months, I've been watching how chip companies are positioning themselves in the AI era. Nvidia, AMD, and Qualcomm all have completely different strategies for software ecosystems, but they are solving the same problem: how to make AI code run on their own hardware.

Nvidia is the wall-builder.

Twenty years of CUDA accumulation, from university labs to internet companies, all AI developers are by default CUDA users. Now, what Jensen Huang is pushing with DSX goes further, packaging software, hardware, and operations into a complete blueprint for an AI factory. CUDA makes developers unable to leave you; DSX makes enterprises buy your entire solution directly.

The wall-builder fears migration costs the least. You're running fine, try moving it.

AMD chose to fill the pits.

At the May Shanghai Developer Day, Lisa Su donated a $3.6 million server cluster to the vLLM open-source community. vLLM is the world's largest open-source inference framework, used at the core of DeepSeek and Kimi. Donating to vLLM is equivalent to putting the hardware directly into the hands of the world's largest number of inference developers. On the same day, they released the vLLM-ATOM zero-code migration plugin, allowing CUDA inference services to run on AMD cards without changing a single line of code. Just get it running first.

Qualcomm is the bridge-buyer.

Not building its own, not donating, but directly spending money to buy a ready-made one. Modular's software platform plus Snapdragon chips equals a complete AI infrastructure from edge to cloud.

Three approaches: building walls, filling pits, buying bridges.

The same goal: letting AI developers use their chips without starting from zero.

I don't know which will win. But one thing is certain: the era of chip companies just selling hardware is over. Chip companies without a software ecosystem cannot survive in the AI era.

Will this time be different

When a big company acquires a small one, the hardest part has never been technology. When Chris Lattner worked on Swift for TensorFlow at Google, the technology was good, but the organization didn't support it, and the project ultimately failed.

But for twenty years, Lattner has been doing the same thing: making code not tied to specific hardware. LLVM was, Swift was, and Mojo is too. Apple took his twelve years to create Swift, Google took his MLIR years, and now Qualcomm has taken Modular.

His technology has each time become the infrastructure of big companies. This time, it probably won't be an exception either.

But at least this time, he has gained scale. The annual shipment of hundreds of millions of Snapdragon chips is the missing piece Mojo language needed most. Whether the story can take off depends on whether Qualcomm gives enough time.

图片

Macaron 🧁 | Software is the entry ticket for chip companies

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry