On 9 November 2026 Google added TPU 8tx and TPU 8i Max to its eighth-generation Tensor Processing Unit line, alongside TPU 8t (training) and TPU 8i (inference).
TPU 8tx uses TPU 8t silicon with hardware-level confidential computing for secure inference and small-scale training, on fifth-generation AMD EPYC hosts.
TPU 8i Max targets memory-heavy work such as very long context windows, with twice the compute and memory bandwidth of TPU 8i.
It is the first year Google has introduced four TPU systems; the TPU business serves nearly 100 customers.
| Chip | Announced | Built for |
|---|---|---|
| TPU 8t | 22 April 2026 | Large-scale model training |
| TPU 8i | 22 April 2026 | Low-latency inference and reasoning |
| TPU 8tx | 9 November 2026 | Confidential inference and small-scale training |
| TPU 8i Max | 9 November 2026 | Memory-heavy inference: long context, large agent workloads |
First TPU in use inside Google's data centres
TPUs power AlphaGo in its match against Lee Sedol
TPU publicly unveiled at Google I/O
Seventh-generation TPU, Ironwood, announced at Google Cloud Next
Eighth generation split into TPU 8t (training) and TPU 8i (inference), the first split of the line
TorchTPU open-sourced
TPU 8tx and TPU 8i Max added
An application-specific integrated circuit (ASIC) Google designed for neural-network machine learning. Unlike a GPU, it has no graphics hardware and favours high-volume, low-precision arithmetic, using a systolic array of multiply units. The second generation added floating point and the bfloat16 format.
Simple Analogy: A GPU is a multi-tool; a TPU is a machine built for one job.
GS Paper III > Science and Technology > AI, semiconductors
General Awareness > Science and Technology
Application-specific integrated circuit: a chip built for one task rather than general computing
Hardware-level protection of data and models while they are being processed
Running a trained AI model to produce outputs, as opposed to training it