Yicai 2026-11-10 07:54:22 Listen to the news
Author: Liu Xiaojie Editor: Ning Jiayan
Google is accelerating the iteration of its AI chips. On November 9 local time, Google announced the launch of two new TPUs (Tensor Processing Units): the TPU 8tx and TPU 8i Max, designed for secure and confidential inference and large-context processing, respectively.

This is Google's second update to its TPU lineup this year and the first time it has launched four TPUs in a single year. Unlike in the past, when the focus was primarily on improving computational performance, Google is now emphasizing the demand generated by AI agents, with the blog describing the new TPUs as “built for the agent era.”
Among them, the TPU 8tx emphasizes hardware-level data security, enabling the processing and training of AI models while handling sensitive data; the TPU 8i Max, through a dual-chip design, doubles computational power, memory capacity, and memory bandwidth compared to the TPU 8i, to support context processing at the million-token scale.
At the opening of U.S. equities on November 9, Google (GOOGL.US) surged approximately 4% before the gains narrowed; by closing, Google rose 1.18% to $288.15 per share, with a total market capitalization of $3.54 trillion.
At the Google Cloud Next conference in April this year, Google launched two eighth-generation TPUs: the TPU 8t for large model training and the TPU 8i optimized for real-time inference. The two new TPUs released this time further optimize for specific challenges in enterprise AI applications.
Typically, enterprise data can be encrypted during storage and network transmission, but it often needs to be decrypted for actual processing on computers, introducing potential security risks. According to Google, the TPU 8tx integrates security capabilities directly at the hardware level, ensuring that sensitive data and model weights remain protected even during actual computation. This capability is especially valuable for customers with high data security requirements, such as financial institutions and government agencies.
The other product, the TPU 8i Max, addresses the memory wall in the AI industry. As AI agents shift from simple question-answering to complex task execution, computational efficiency is increasingly limited by whether memory can deliver data to the processor in a timely manner. To solve this, Google has integrated two TPU 8i chips into a single package, doubling both computational power and HBM bandwidth. Google states that this design primarily serves applications requiring extremely large context windows and high-density computations.
The commercialization of Google's TPU is accelerating. Amin Vahdat, Senior Vice President and Chief Technology Officer at Google, said that TPU has become a rapidly growing business with nearly 100 customers, including leading AI labs, trading firms, retailers, universities, and more.
In the past, Google TPU's business model primarily involved providing computing power services to external customers through Google Cloud. In April of this year, Google announced for the first time that it would directly sell TPU systems to select customers, allowing them to deploy the systems in their own data centers.
In the second quarter, Google stated that it had completed its initial deliveries. In its most recent third-quarter earnings report, Google further indicated that TPU sales have become a significant contributor to the growth of cloud revenue, with expected substantial growth in related revenue by 2027.
Google is also one of the largest customers of its own TPU chips in the large model space. In October 2025, large model company Anthropic announced an expansion of its partnership with Google Cloud, planning to use up to one million TPUs in a collaboration valued at tens of billions of dollars. Then, in April this year, Anthropic further announced an expanded partnership with Google and Broadcom, adding several gigawatts of next-generation TPU computing power.
OpenAI is also considering purchasing Google's TPU. In late October, reports emerged that OpenAI plans to buy Google’s AI chips to reduce its reliance on NVIDIA GPUs. Additionally, OpenAI’s official website is currently hiring TPU compiler engineers to optimize the performance of its cutting-edge models on TPU hardware. However, the procurement plan has not yet been officially confirmed.
At the end of October, Google also announced a partnership with Oracle, planning to offer TPU computing power through Oracle’s cloud services starting in 2027. Oracle will become the first hyperscale cloud provider other than Google to offer TPU.
Although NVIDIA still holds a significant position in the AI computing market, Google is attempting to expand its market share through additional products and sales channels, posing a challenge to NVIDIA.
However, to compete with NVIDIA, Google cannot rely solely on hardware performance. A major advantage of NVIDIA is its long-established CUDA software ecosystem. CUDA is NVIDIA’s parallel computing platform and programming model, forming the foundational basis for a vast number of AI developers using GPUs. Companies switching to other chip platforms often face significant additional costs for software adaptation and migration.
Google is also working to address this issue. In this release, Google highlighted progress in opening up the TPU software ecosystem. Last month, Google open-sourced TorchTPU, enabling PyTorch code and Hugging Face models to run directly on TPUs without requiring code rewrites. Additionally, Google has added TPU support for major open-source inference frameworks, aiming to lower the barrier for developers to use TPUs through multiple avenues.
The competition in AI chips is multidimensional, encompassing software and hardware ecosystems, deployment flexibility, and overall computational costs. Whether Google can secure more orders may still require validation from the market and time.
However, another advantage for Google, widely recognized by outsiders, is its "full-stack AI" capability—it is both a developer of AI chips and large models, as well as a cloud computing service provider, enabling it to integrate the development of chips, software, and models.
In this blog, Google mentioned that the Gemini models trained on TPUs are, in turn, accelerating the development of the next-generation TPUs. "The team is using the AlphaChip reinforcement learning system, along with Gemini models to assist in chip layout design, reducing tasks that previously took weeks to just a few hours." Hardware-software co-design is a differentiated competitive advantage Google aims to establish.
Article author


