The new systems extend the eighth-generation TPU lineup announced in April with the 8t for large-scale training and the 8i for inference and reasoning. TPUs have become a fast-growing business, with Google claiming nearly 100 customers including Meta. The more modular approach can deliver specialized systems “in a matter of months rather than years.”
The TPU 8i Max is designed for long-context agentic work. It combines two 8i chips on a single package, doubling compute, SRAM and HBM bandwidth within the same rack footprint. Google claims more compute per square foot and per watt. The chip targets workloads where larger models are targeting one million tokens. Google also introduced TPU Sync, an open-source model-state layer that moves KV caches off the chip’s HBM to host DRAM and, experimentally, shared storage. Google said this can cut KV transfer latency by “more than 90%.” The Agent Compute Engine pauses idle agent sessions and moves their working memory aside.
The TPU 8tx is built on the same silicon as the 8t but optimized for secure confidential inference and small-scale training. It replaces the Arm Axion hosts with 5th Gen AMD EPYC CPUs and adopts the Boardfly network, which Google claims can deliver “up to 2x lower latency compared to TPU 8t.” The hardware isolation is designed to prevent snooping on sensitive data, including model weights, making it suitable for regulated enterprises and governments using frontier models. Google’s Private AI Compute platform, described as a “hardened TPU platform” with protections “comparable to a typical TEE,” already runs confidential serving on TPUs; the 8tx appears to offer that capability as a product.
Google credits its vLLM-TorchTPU stack with “state-of-the-art price-performance marks” on InferenceX and AgentX benchmarks, but the only published third-party numbers so far are for the previous-generation Ironwood TPU. In SemiAnalysis tests in September, Ironwood delivered 50.4% more tokens per dollar than Nvidia’s B200 and 96% more than the B300 at 20 tokens per second per user. Newer November results showed Ironwood’s FP8 beating Nvidia’s FP4 GB300 NVL72 on performance per dollar.
The TPU 8tx and 8i Max are listed as “coming soon” on Google Cloud’s TPU page. Availability dates and pricing have not been announced.