Two different bottlenecks
Google announced two AI accelerator systems on November 9. Both extend the company's eighth-generation TPU lineup, but they target different constraints. One is designed for organizations that need to protect data and model weights while a workload runs. The other is built to move data through memory faster when a model has to process large inputs.
Google says TPU 8tx combines the accelerator silicon used in TPU 8t with AMD hosts and hardware protections for data in use. The company positions it for inference and smaller training jobs, such as adapting a model using private data. The announcement describes the design, but does not establish compliance with any particular organization's rules or verify the behavior of an application deployed on it.
Google says TPU 8i Max puts two TPU 8i chips in one package. The company reports twice the high-bandwidth memory throughput of a standard TPU 8i. That design is relevant to long inputs, such as large codebases or video, where feeding the processors can become the limiting factor.
What buyers can verify today
The announcement did not give a public ordering date for either system. CGH found no entries for these two models in the Compute Engine location table and no public prices in the TPU pricing material reviewed. For teams comparing infrastructure, those gaps matter as much as the design specifications. A hardware comparison still depends on where capacity can be provisioned, how it is priced, and how the team's own workload performs.