Two new Google TPU systems, 8tx and 8i Max, target private AI processing and tasks that draw on large amounts of data. Google announced them online on November 9, 2026.
Google says these are new system designs built around existing eighth-generation chips. Thus, changes to memory, networks and security can give a chip new roles without a new chip design.
How the new Google TPU systems differ
TPU stands for Tensor Processing Unit. Google Cloud’s technical guide describes it as a custom chip designed to speed up machine learning, especially calculations on large grids of numbers.
Google assigns different jobs to the two systems below. The table shows their intended roles, rather than a ranking of performance.
| System | Main role | Design focus |
|---|---|---|
| TPU 8tx | Private AI use and smaller training jobs | TPU 8t silicon, AMD EPYC hosts and hardware security |
| TPU 8i Max | Work with large amounts of context | Two TPU 8i chips in one package |
Source: Google
Data as of: 2026-11-09
Google assigns the two systems different roles: protected processing and large-context workloads.
For 8i Max, Google reports twice the compute and memory bandwidth of TPU 8i. It says the design packs more computing power into the same rack space.
Why Google TPU hosts use AMD processors
AMD says fifth-generation EPYC processors power the 8tx host computers. These general-purpose processors prepare data, assign work and pass data to and from the AI chips.
According to AMD, the design protects data in host memory and on its path to and from the AI chips. First, it checks the AI chip’s identity and security setup before granting access to protected memory.
In turn, this role gives context to AMD’s recent data center growth. Its announcement shows how the company can supply host processors to systems built around another company’s AI hardware.
What a larger context window means
Google’s Gemini documentation compares a context window to short-term memory: the information a model can draw on when producing a response. For example, its guide lists tasks such as answering questions about long texts or video.
However, that guide also describes limits. Finding several separate facts in a large input can be less accurate than finding one. Longer requests also tend to delay the first response, according to Google’s guide.
This distinction helps explain the hardware goal. Google says 8i Max addresses the flow of data to processors when agents work across large codebases or hours of video.
Hardware security still needs secure software
France’s cybersecurity agency, ANSSI, describes confidential computing as an extra layer of protection for data while it is being used. However, its technical paper says this does not secure a whole system on its own.
For example, ANSSI says weak passwords and flaws in an app remain risks even when the app runs inside a protected space. The paper covers the technology broadly; it is not a review of TPU 8tx.
AMD also says its hardware security features must be enabled by the server maker or cloud provider. It directs customers to their provider to confirm which features are supported.
Sources: Google announcement, AMD, Google Cloud TPU guide, Google context guide, ANSSI
