Google announced two additional eighth-generation TPU systems on November 9: TPU 8tx for confidential inference and TPU 8i Max for large-context workloads. Availability dates and pricing remain unannounced, TechTarget reports.
The designs target data protection during processing and capacity for larger contexts or concurrent workloads. For small teams evaluating whether to use customer context in AI-assisted lead response, those capabilities could matter. They aren’t evidence of better customer-service results.
Two distinctions frame the announcement:
The distinction is systems, not entirely new silicon. Google says the configurations change how compute, networking, memory and security surround existing chips.
These are specialized infrastructure designs, not newly available customer-service applications.
Google introduced TPU 8t for large-scale training and TPU 8i for real-time reasoning in April.
Protecting data in use
TPU 8tx is Google’s first TPU with confidential computing capabilities. Google says it combines TPU 8t silicon with fifth-generation AMD EPYC CPU hosts to protect and isolate sensitive data and model weights while they’re in use. Its stated audience is regulated enterprises and governments running proprietary or frontier models alongside sensitive information.
Mark Lohmeyer, Google Cloud’s vice president and general manager for AI and computing infrastructure, told reporters protection covers data moving from the host to TPUs and active inference, Network World reports.
“Then that entire system is attested so that we know it is what it says it is,” Lohmeyer said.
The system uses Google’s Boardfly low-latency network to shorten communication paths between chips. Google claims inference latency up to half that of TPU 8t, with a near-zero performance penalty for security. Teams can also fine-tune with private data, distill models and run reinforcement learning without moving sensitive data between clusters, Google says.
Capacity, not just speed
TPU 8i Max packages two TPU 8i chips together. Google says it doubles compute, on-chip SRAM and high-bandwidth memory bandwidth within an existing rack footprint.
Network World reports 384 GB of high-bandwidth memory, 33% more than TPU 8i, alongside 17.2 TB/s of memory bandwidth and 768 MB of SRAM.
“It’s like adding more lanes to the highway,” Lohmeyer said. “We’re not necessarily making each car travel faster, we’re creating more capacity so more traffic can move through the system at the same time.”
Google also announced Agent Compute Engine software to suspend and resume idle agent sessions. Its open-source TPU Sync coordinates chip memory, host memory, shared storage and networking, moving working memory off-chip during pauses and restoring it when work resumes. Network World reports support for v5, v6 and v7 TPUs as well as v8.
Google claims TPU Sync reduces key-value cache transfer latency and time to first token by more than 90%. That describes moving stored model context and starting a response, not a measured sales outcome. Google supplies no small-business customer-service savings figure.
For a small-team owner considering customer-facing agents, the implication is evaluation rather than an infrastructure purchase: cloud availability, pricing and evidence from relevant workloads are still needed to judge whether these systems make using customer information practical.
Sources
- https://blog.google/innovation-and-ai/infrastructure-and-cloud/global-network/tpu-8i-max-and-8tx/
- https://live.euronext.com/en/financial-news/google-unveils-new-ai-chip-variants-cloud-demand-surges
- https://www.networkworld.com/article/4245897/google-targets-ais-next-infrastructure-bottlenecks-with-two-new-tpus.html
- https://www.techtarget.com/news/google-specializes-tpus-for-agentic-confidential-computing/874084/
- https://biz.chosun.com/en/en-it/2026/11/09/ZLZV3H4TYNC5LM5XL57F2EPZ7Y/