Google ships four TPUs in one year: 8tx guards secrets, 8i Max takes on million-token context
As agents multiply, the chip pitch changes from how fast it computes to whether you dare put confidential data inside it, and whether it can read a million tokens in one gulp.
Three key facts
Google announced TPU 8tx and TPU 8i Max on November 9, bringing its eighth-generation lineup to four systems: 8t for large-scale training, 8i for real-time reasoning, plus the two new ones. Google says this is the first time it has introduced four TPU systems in a single year. The TPU business now has nearly 100 customers, including frontier labs, trading firms, retailers and universities.
TPU 8tx tackles the problem that data has to be unlocked while it is being processed. Normally encryption covers data at rest and in transit, but memory is exposed during computation. 8tx uses the same silicon as 8t, paired with fifth-generation AMD EPYC host CPUs, and pushes encryption into the hardware layer so sensitive data and model weights stay isolated even in active use. It also uses the low-latency Boardfly network built for real-time reasoning, delivering up to 2x lower latency than 8t. Teams can fine-tune on private data, distill large models into smaller ones and train agents with reinforcement learning on one system, without moving sensitive data between clusters.
TPU 8i Max attacks the memory wall by putting two low-latency 8i chips in a single package, doubling compute, SRAM and HBM bandwidth within the same rack footprint. Google also open-sourced TPU Sync, a model-state layer that tiers KV caches off-chip and streams context across prefill and decode nodes; Google says it cuts KV transfer latency and time to first token by more than 90 percent.
Wang Dou's Take
Look at the names. One is called tx, one is called Max. Neither is called stronger.
The 8tx customer profile is spelled out without embarrassment: regulated enterprises and governments running frontier or proprietary models next to their most sensitive data. Translation: banks, hospitals, defense, government agencies. These are exactly the buyers who stayed off public clouds, not because they did not want AI but because they did not dare. Google is not selling compute here. It is selling the sentence: we cannot even see your data.
8i Max is blunter. When one chip is not enough, glue two into one package. What does that tell you? That the bottleneck in inference has moved off compute and onto memory bandwidth. Pull context to a million tokens, have dozens of agents read an entire codebase at once, and the road between memory and the compute unit gets jammed before the math does. Doubling SRAM and HBM bandwidth is widening that road.
The 90 percent cut in time to first token is the number users will actually feel, because nobody cares about flops per second. They care about how long they wait after hitting enter.
One more thing
Google is also lowering the on-ramp. TorchTPU is open source, so PyTorch code and Hugging Face models run on TPUs without rewrites. An open-source TPU agent suite plugs into coding tools like Claude Code and Cursor to handle migration and tuning in plain language. And Cloud Pathways, the runtime behind Gemini, is now generally available on GKE for any cloud customer.
Source: Google Official Blog
