Google splits TPU 8 into a security chip and a long-context speed chip

Neither chip is new from the ground up. Google took the training-focused TPU 8t and the inference-optimized TPU 8i and rebuilt them for tighter jobs: the TPU 8tx carries a security badge, and the TPU 8i Max is aimed at models that need to keep a lot of information in view at once. The TPU 8tx protects data at the chip level through hardware encryption built with AMD, rather than leaning on software layers stacked on top.

The target buyers for the 8tx are government agencies and regulated industries training on sensitive data. For them, the AMD work gives compliance teams a specific feature to evaluate instead of a broad promise of security. The pairing is notable on its own: two companies usually found on opposite sides of the chip fight cooperated on encryption, a sign confidential computing is becoming table stakes for landing certain contracts.

The TPU 8i Max chases a different problem. Google says it brings increased compute density and improved memory bandwidth for AI applications working with long context windows. That follows the base TPU 8i, whose pods scale to 1,152 chips and deliver 11.6 Exaflops, with 384 MB of on-chip SRAM — a bigger allotment than training chips usually carry, which helps cut the wait that slows real-time answers.

The training side has its own baseline. Google announced the TPU 8t on April 22, 2026, pitching nearly 3x the compute performance per pod versus the previous Ironwood generation. A TPU 8t superpod stretches to 9,600 chips and hits 121 ExaFlops, backed by 2 PB of shared high-bandwidth memory.

The split is the point: training rewards raw throughput across thousands of chips, while inference rewards speed, memory access and cost per answer. Alphabet shares rose more than 2% in premarket trading after the announcement, with the market reading the specialization as a positive for Google's position against third-party accelerators — including Nvidia hardware that enterprise cloud clients have leaned on.