Google began its eighth-generation TPU rollout in April with the TPU 8t and 8i, aimed respectively at training-heavy and inference-focused AI workloads. This week, it filled out that generation with two more specialised designs — four systems in a single year.¹

The first newcomer, TPU 8tx, is Google’s security play. Developed with AMD and based on the TPU 8t, it is designed for confidential training and inference, targeting government agencies, regulated industries and companies reluctant to expose proprietary data during model development. Google says the chip “builds security directly into the hardware layer,” protecting sensitive information and model weights even while they are being used.²

The second, TPU 8i Max, tackles a different bottleneck: scale. By effectively combining two inference-oriented 8i chips into one package, Google says it doubles memory capacity, compute power and bandwidth, allowing models to work across far larger context windows. The company is framing that density as crucial for AI applications that must process broad swaths of information in a single query.¹

For Google, the announcement is also an argument about speed. Sundar Pichai said the company’s hardware engineers are co-designing systems with Google DeepMind rather than waiting through “years for a traditional semiconductor cycle,” a model he says lets Google keep innovating quickly for customers.Sundar Pichai (sundarpichai)³

That is the competitive subtext. Microsoft, Amazon, OpenAI and Anthropic are also pursuing custom AI silicon and software, seeking more leverage against Nvidia’s dominant position. Google’s bet is that specialised chips — one built around confidentiality, another around giant workloads — can make its cloud harder to ignore and its AI supply chain less dependent on anyone else.²