Google has officially expanded its eighth-generation Tensor Processing Unit (TPU) lineup with the introduction of two new chips: the TPU 8tx and the TPU 8i Max. This move is a direct play to reduce reliance on third-party hardware, specifically Nvidia, which currently dominates the AI accelerator market. The announcement signals Google’s continued commitment to vertical integration in its AI infrastructure stack.

TPU 8tx Focuses on Secure Training with AMD

The TPU 8tx was designed in collaboration with AMD and is built upon the architecture of the TPU 8t. Its primary function is secure AI model training and operation. Google claims the chip can isolate sensitive data and model-defining information even while models are actively in use. This capability allows for parallel reasoning tasks and training, meaning smaller models can be developed concurrently with larger operations. A key practical application highlighted by Google is the ability to train models using corporate data without transferring that data between devices. This reduces the attack surface for data leaks and streamlines the workflow for enterprises concerned with privacy compliance. The chip effectively acts as a secure enclave for model development, ensuring that proprietary weights and training data remain protected during the entire lifecycle.

TPU 8i Max Doubles Performance Metrics

The second addition, the TPU 8i Max, combines two TPU 8i chips into a single unit. The original TPU 8i was introduced in April 2026, and this new iteration doubles the memory, processing power, and data transfer speed compared to two separate TPU 8i chips. This consolidation aims to help AI models consider more information when answering queries, effectively increasing the context window and throughput for inference tasks.

Industry-Wide Shift Away from Nvidia

Google is not alone in this pursuit. The source notes that other major AI entities, including Microsoft, Amazon, OpenAI, and Anthropic, are actively developing their own custom chips. The collective goal is to break Nvidia’s dominance in the AI hardware space. For developers, this diversification could eventually lead to more competitive pricing and specialized tools, but it also fragments the software ecosystem that currently relies heavily on CUDA.

Key Takeaways

  • The TPU 8tx was designed with AMD to enable secure, parallel training and reasoning without data transfer.
  • The TPU 8i Max combines two TPU 8i units, doubling memory, processing power, and data transfer speeds.
  • Google, Microsoft, Amazon, OpenAI, and Anthropic are all developing custom silicon to reduce Nvidia dependency.

The Bottom Line

Custom silicon is no longer a luxury; it’s a survival strategy. By partnering with AMD and doubling down on their own architecture, Google is betting that vertical integration will win the AI infrastructure war, forcing developers to adapt to a fragmented hardware landscape.