Google’s New TPUs Push AI Hardware Into Security and Scale

Google’s New TPUs Push AI Hardware Into Security and Scale
Google began building out its eighth-generation TPU family in April, launching the TPU 8t and 8i for larger training demands and inference work. On Monday, it added two more variants—giving the company four distinct TPU systems in a single generation within a year.

The new TPU 8tx is Google’s security play. Developed with AMD and based on the TPU 8t, it is designed to train and run models on sensitive information without exposing data or model weights during active use. Google says the hardware keeps those assets “protected and isolated even while in active use,” a pitch aimed squarely at government agencies and regulated businesses.

Its companion, TPU 8i Max, addresses a different AI bottleneck: the amount of information a model can hold in view during a query. By effectively combining two inference-focused TPU 8i chips into one package, Google says it doubles memory, processing power and bandwidth, enabling applications built around much larger context windows.

For Google, the significance is not just the two products but the tempo. Chief executive Sundar Pichai highlighted “four distinct TPU systems in a single generation in one year,” arguing that co-design between Google’s hardware engineers and DeepMind lets the company avoid waiting through traditional semiconductor cycles.

That claim lands in an increasingly crowded race. Microsoft, Amazon, OpenAI and Anthropic are all pursuing specialized AI silicon and software, seeking more control over cost, capacity and performance—and, crucially, less dependence on Nvidia. Google’s latest chips suggest its strategy is to compete not with one all-purpose accelerator, but with hardware tuned to the two pressures now shaping enterprise AI: keeping data private and making models remember more.

https://foxvector.com