THE CRUNCH
Google has added two systems to its eighth-generation TPU lineup: the TPU 8tx, built for confidential computing, and the TPU 8i Max, aimed at memory-intensive workloads. With the additions, Google says it has shipped four TPU systems within a single year for the first time, part of its shift from keeping the chips in-house to selling them to outside customers, where it now counts nearly 100 external TPU customers.
Neither variant is a new chip generation. Both are new system configurations around existing silicon, with different packaging, networking and memory setups. Google frames that modularity as deliberate, letting it deliver new hardware variants in months rather than years.
The TPU 8tx targets enterprises and government agencies handling sensitive data. Data is normally encrypted in storage and transit but must be decrypted in memory while a processor works on it; Google says the 8tx keeps data and model weights protected and isolated even during active use. Nvidia already offers a comparable confidential computing feature for its GPUs. The system uses Google's low-latency Boardfly network and is claimed to deliver up to twice the latency reduction of the TPU 8t, with Google suggesting one system can handle fine-tuning, model distillation and agent training without moving sensitive data between clusters.
The TPU 8i Max addresses the memory bottleneck that emerges when AI agents work across large codebases or hours of video simultaneously. Google put two TPU 8i chips on a single package, doubling compute, memory and memory bandwidth in the same rack format. On the software side, Google has released TorchTPU as open source, so PyTorch code should run on TPUs with minimal changes, and supports inference frameworks including vLLM and SGLang.