Google 与 Blackstone 正在成立一家新的 AI 云端合资企业，由 Blackstone 提供 5bn 美元资金，旨在扩大 Google 客制化 AI 晶片的触及范围并挑战 Nvidia。双方表示，计划在明年让 500 MW 的资料中心容量上线，之后再进一步扩大规模；Google 将提供硬体，包括其 Tensor Processing Units，用于训练和部署 AI 模型。

这项合资企业将由 Google 资深人士 Benjamin Treynor Sloss 领导，而 Blackstone 将持有多数股权。该项目锁定的部分设施其实早已在施工中，且在加上杠杆后，总投资额最高可达 25bn 美元。Blackstone 的管理资产达 1.3tn 美元，已透过持有 QTS、为 CoreWeave 提供融资，以及投资 Anthropic 和 OpenAI，在 AI 基础设施领域相当活跃。

这笔交易与 Nvidia 的策略相似：Nvidia 透过提供 GPU 和算力合约，支持像 CoreWeave 和 Nebius 这类新兴云端供应商，但同时也凸显 Google 正更积极推广 TPU，尽管它仍是 Nvidia 晶片的重要买家。Google 云端主管 Thomas Kurian 表示，这项合资企业将有助于满足对 TPU 日益成长的需求，并为加速运算提供更多选择；他指出，10 家领先的 AI 实验室中有 9 家使用 TPU。Google 也表示，其 TPU 已协助训练 Gemini 模型，这些模型在基准测试中与 Anthropic 和 OpenAI 相比表现具有竞争力。

Google and Blackstone are forming a new AI cloud venture backed by $5bn from Blackstone, in a move aimed at expanding the reach of Google’s custom AI chips and challenging Nvidia. The partners said they plan to bring 500 MW of data centre capacity online next year and then scale further, with Google supplying hardware including its Tensor Processing Units for training and deploying AI models.

The venture will be led by Google veteran Benjamin Treynor Sloss, while Blackstone will hold the majority stake. Some of the facilities targeted for the project were already under construction, and after leverage the total investment could rise to as much as $25bn. Blackstone, with $1.3tn in assets under management, has already been active across AI infrastructure through ownership of QTS, financing for CoreWeave, and investments in Anthropic and OpenAI.

The deal mirrors Nvidia’s strategy of backing emerging cloud providers such as CoreWeave and Nebius while supplying GPUs and compute contracts, but it also highlights Google’s effort to push TPUs more aggressively even as it remains a major buyer of Nvidia chips. Google cloud chief Thomas Kurian said the venture will help meet growing demand for TPUs and provide more options for accelerated compute, noting that 9 of the 10 leading AI labs use TPUs; Google also said its TPUs have helped train Gemini models, which have benchmarked competitively against Anthropic and OpenAI.