Google Adds TPU 8tx and 8i Max to Its Gen-8 Lineup, Aimed at Agent Workloads

Google has expanded its eighth-generation TPU lineup with two new parts, the TPU 8tx and the TPU 8i Max, according to a November 9 report from NDTV Profit, which frames both chips as aimed at AI agent workloads.

The headline is the workload rather than the silicon. Agent systems behave differently from chatbots: a single user request can trigger a chain of model calls, tool invocations and retries, so the serving system spends more time on latency-sensitive sequential inference and less on throughput-friendly batch work.

That distinction helps explain why hyperscalers have started splitting accelerator families. A chip tuned for training large models and a chip tuned for serving many short, dependent requests strike different balances across memory, bandwidth and interconnect, so vendors increasingly ship separate parts for separate jobs rather than one general-purpose design.

For Google, TPUs have long been an internal advantage, the hardware underneath Gemini training and serving inside Google Cloud. Publishing a differentiated Gen-8 lineup is also a competitive statement aimed at NVIDIA, whose GPUs remain the default for enterprises that are not already inside Google's stack.

What is not yet public matters as much as what is. The report supplies product names and a workload target but no specifications, no pricing, no regional availability and no customer list, so the concrete differences between the 8tx and the 8i Max remain unknown. The suffixes suggest different positions in the lineup, but the reporting does not spell that out.

For buyers, the practical question is cost per agent task, not peak throughput. If agents really do become the dominant inference pattern, the vendor that can show a lower cost for long chains of short calls wins the deployment, and that is the number to check when Google publishes technical details.