Google has introduced TPU 8tx and TPU 8i Max, two AI computing systems aimed at problems that grow as companies deploy autonomous agents: protecting sensitive information during processing and handling much longer context windows.
The systems build on existing eighth-generation TPU chips rather than replacing them. They join TPU 8t for training and TPU 8i for reasoning, announced in April with up to 2.7× better training performance and 80% better inference price-performance than the previous generation. That makes four TPU systems in one year, a first for Google.

A secure path for private AI
TPU 8tx targets regulated enterprises and government organizations that want to run proprietary models with private data. It combines TPU 8t silicon with AMD host processors and adds hardware-level protections designed to isolate data and model weights even while they are actively in use.
Google says its Boardfly networking design delivers up to 2× lower latency than TPU 8t with almost no performance penalty. That could allow organizations to fine-tune models, train agents and run confidential inference on one system rather than move sensitive workloads across different clusters.
Two chips for longer context
TPU 8i Max addresses a different constraint: memory. AI agents working across entire codebases, lengthy video or many concurrent tasks can run into bandwidth limits before they exhaust compute capacity.
Google has put two TPU 8i chips in a single package, doubling compute, on-chip SRAM and high-bandwidth memory bandwidth compared with a single TPU 8i. It also fits twice as much silicon into the same rack footprint, a meaningful consideration for data centers where power and space are expensive.
Why Nvidia should pay attention
TPUs are no longer just an in-house tool. Google says nearly 100 customers now use them, from frontier labs and trading firms to retailers and universities. Some run TPUs inside their own data centers.

The clearest signal came in April, when Anthropic signed for multiple gigawatts of next-generation TPU capacity from 2027, its largest compute commitment yet. Anthropic still runs Claude on Nvidia GPUs and AWS Trainium too, so the real fight is for a bigger slice of each lab's chip budget, not a clean switch.

Google wants developers, not just chip buyers
The hardware comes alongside an effort to make TPUs more accessible to mainstream AI developers. Google's open-source TorchTPU support allows PyTorch code and Hugging Face models to run natively on TPUs. A new TPU agent toolkit, available in preview, works with coding tools including Claude Code and Cursor to help set up and optimize workloads.
Google also introduced TPU Sync, software that moves model state across different memory and storage layers. It claims the approach can cut cache-transfer latency and time to first token by more than 90%.
Google is also using its own AI to design its next chips, with Gemini producing chip layouts in hours instead of weeks. The broader bet is that custom AI infrastructure will be sold as workload-specific systems, not merely faster chips. Google has not disclosed pricing or a firm availability date. The key test will be whether customers see better economics in production, not just on Google's benchmarks.
Get stories like this in your inbox every week. Subscribe to The Week in Perspective.

