Google expands AI hardware with two new TPUs focused on confidential inference and larger memory

1–2 minutes

TPU 8tx brings hardware‑level security for private AI training

Google announced the chips on Monday, unveiling the TPU 8tx, a device built in partnership with AMD and derived from the earlier TPU 8t. The chip is engineered for confidential AI workloads, embedding security directly into the silicon so that sensitive data and model parameters remain isolated even while the system is active. In addition, a custom inferencing network lets the chip perform reasoning alongside training, enabling on‑the‑fly distillation and training of small models using proprietary data without moving that data between devices.

The chip “builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use,” the company wrote.

TPU 8i Max doubles memory and performance in a single package

The second offering, TPU 8i Max, stacks two of the inference‑focused TPU 8i chips that debuted in April, effectively doubling memory capacity, processing power and bandwidth within one module. This consolidation allows AI models to process a broader set of information during queries, expanding the scope of tasks the hardware can handle.

Strategic push as rivals race to reduce Nvidia dependence

Both new TPUs belong to Google’s eighth‑generation silicon family introduced in April. The launch comes as major cloud providers—Microsoft, Amazon—and AI research labs such as OpenAI and Anthropic develop their own AI‑specific semiconductors and software stacks, aiming to lessen reliance on Nvidia’s dominant GPU portfolio.

The timing also coincides with a period of heightened vulnerability for Nvidia. Over the weekend, Nvidia CEO Jensen Huang disclosed an affair with an individual alleged to be a Chinese spy, adding another layer of uncertainty to the competitive landscape.