TPUs

Google has expanded its eighth-generation TPU lineup from two systems to four, aiming to meet the demands of AI agents and other increasingly complex workloads. TPU 8t targets large-scale model training, while TPU 8i is designed for real-time reasoning; Google says they offer up to 2.7 times the prior generation’s training performance and 80% better inference price performance, respectively. The new TPU 8tx and TPU 8i Max adapt these designs to two specific constraints: protecting sensitive data during computation and keeping large AI workloads supplied with memory. Google’s broader argument is that modular system design lets it respond to changing AI needs faster than chip manufacturing alone would allow.

TPU 8tx is intended for regulated organizations running proprietary models and agents on sensitive data. Hardware-level confidential computing protects data and model weights while they are in use, and the system combines training-grade compute with a low-latency network. Google says this can reduce latency by up to half compared with TPU 8t, while allowing customers to fine-tune models, distill them, and train agents without moving private data between clusters. TPU 8i Max addresses memory bottlenecks in tasks such as analyzing large codebases or long videos: two TPU 8i chips in one package provide twice the compute, SRAM, and HBM bandwidth, with more computing capacity per rack footprint. Google also presents broader PyTorch and open-source software support as part of making TPUs accessible to developers and customers.