Google 开源 TorchTPU，继 Amazon 后挑战 Nvidia 的 AI 芯片主导地位
Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers
- 1Google 宣布 TorchTPU 开源，PyTorch 应用改配置即可迁移到 TPU 和 Google Cloud，无需重写
- 2Amazon 前一日发布同类工具 TorchNeuron 支持迁移到 Trainium 和 AWS，并向 PyTorch 主项目提交解耦 Nvidia 的代码
- 3Google 称已与 Meta、Hugging Face、vLLM 方 Inferact、SGLang 方 RadixArk 合作适配，另推出 TPU agent suite 辅助迁移和排错
- 4PyTorch 据称被约 85% 的 AI 应用使用；文章判断开发者不会因此立刻弃用 Nvidia 硬件
云厂商争夺的实质不是芯片参数，而是 PyTorch 生态的默认绑定；迁移成本降到改配置一级后，Nvidia 的优势被从生态层打开。
- 影响：使用 PyTorch 的 AI 应用切换算力门槛从重写代码降到改配置，云与芯片采购的议价筹码增多；两家迁移方案都绑定自家云和芯片，总拥有成本需要重新核算。
- 参考：可考虑先盘点现有 PyTorch 应用的算力依赖清单，选非关键负载在 TPU 或 Trainium 上小规模试运行并记录迁移耗时；值得把替代算力报价纳入下一轮云预算谈判。
- 依据：迁移免重写、生态适配均为 Google 与 Amazon 官方口径，暂无独立验证；85% 使用率原文标注为据某些统计。
On Tuesday, Google announced that TorchTPU, its long-awaited version of PyTorch, is now available as an open-source project. With this, Google joins Amazon in trying to break Nvidia’s stranglehold on the AI industry, where most AI apps are written by default to run on Nvidia’s chips, or GPUs.
PyTorch is the popular open-source toolkit that AI app developers and startups use to build, train, and run AI models. It is, by some accounts, used by 85% of the AI apps. PyTorch was originally developed by Meta (then Facebook) in 2016, long before the world had an inkling that LLMs would change the world. At the time, Nvidia’s GPUs were already the standard hardware for AI research, so many of PyTorch’s features were built to work specifically with them. In 2022, Meta turned PyTorch over to the Linux Foundation, a nonprofit that manages many important open source projects.
Google’s version, TorchTPU, lets AI apps run on Google’s AI chips, called Tensor Processing Units (TPUs). It also makes it easy to move existing PyTorch apps to TPUs and, therefore, to Google Cloud. Developers can do this with a simple configuration change and no re-writing of their apps.
Amazon made a similar announcement Monday, when it released its own version of PyTorch, called TorchNeuron. TorchNeuron allows AI apps to be ported to Amazon’s homegrown AI chip, Trainium, and to AWS with no rewriting. Amazon also said it going even further in its effort to loosen Nvidia’s hold on PyTorch. It is submitting code to the main PyTorch project that decouples various features from Nvidia’s chips, making it more open to any hardware by default.
Google may be one-upping Amazon in a few ways, however. It says it worked directly with Meta, Hugging Face, Inferact (maker of vLLM) RadixArk (maker of SGLang), and others to ensure that TorchTPU’s features and code libraries work well with these popular AI development tools, just as the original PyTorch does.
Plus the search giant has launched a dedicated tool, the TPU agent suite, to help developers move their apps to Google’s chips and cloud and fix errors along the way. It even updated its popular LiteRT tool, which runs AI on phones, laptops and embedded devices, to work better with PyTorch AI apps.
None of this means AI app developers will automatically ditch Nvidia’s hardware en masse just because they can now easily move to alternative chips and clouds. But Amazon and Google are clearly working to end Nvidia’s advantage.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
Julie Bort is the Startups/Venture Desk editor for TechCrunch.
You can contact or verify outreach from Julie by emailing [email protected] or via @Julie188 on X.
来源：TechCrunch · AI · techcrunch.com