Google today announced 8tx and 8i Max, two TPU systems the company says respond to shifting demands from AI customers. Google says businesses want accelerators that can handle reasoning across large AI model context windows, while some customers in highly regulated sectors want specialized encryption for particular workloads. These are new systems rather than newly designed chips: Google distinguishes the surrounding computing, networking and storage configuration from the silicon itself.
Google says its modular approach lets engineers change memory, packaging, optical connections and security to assemble specialized supercomputers within months instead of years. Designing a chip, by comparison, requires a process spanning multiple years, including transistor layout and manufacturing. The company says it made modularity a priority when it developed its first TPU over a decade ago, using TPUs as components within its AI Hypercomputer.
Google has long worked with Google DeepMind on TPU design and says outside organizations are increasingly contributing feedback about their requirements. Alongside its recent PyTorch work, Google says its hardware and software approach is intended to broaden TPU access for developers and Cloud customers.
The hardware news follows Google's earlier work on software for AI agents: in June, Google, Microsoft and other companies published Agentic Resource Discovery, a draft specification for finding tools, skills and agents during operation. In April, MSD announced a multiyear Google Cloud partnership worth up to $1 billion to bring Gemini Enterprise and specialized agents into research, manufacturing and other business functions.
