Google adds TPU 8tx and TPU 8i Max to eighth-generation AI chip lineup

Google has added two new systems, the TPU 8tx and the TPU 8i Max, to its eighth-generation AI accelerator lineup. The company says it has shipped four TPU systems within a single year for the first time, with both new variants built on existing silicon rather than a new chip generation. The additions target confidential computing and memory-intensive workloads as Google markets its in-house AI chips more broadly to external customers.

Confidential computing with TPU 8tx

The TPU 8tx is aimed at enterprises and government agencies that work with confidential data. Google says data and model weights remain protected and isolated even during active processing, closing the usual gap in which data must be decrypted in memory while a processor works with it. Nvidia already offers a comparable feature for GPUs through its confidential computing architecture. The system uses Google's low-latency Boardfly network and is supposed to deliver up to twice the latency reduction compared with the TPU 8t. Google says enterprises can use one system to fine-tune models, distill large models into smaller ones, and train AI agents with reinforcement learning without moving sensitive data between clusters.

Memory-intensive workloads with TPU 8i Max

The TPU 8i Max addresses memory bandwidth limits that can arise when AI agents work across large codebases or hours of video. Google placed two TPU 8i chips on a single package, doubling compute power, memory, and memory bandwidth in the same rack format. The company says this brings significantly more performance per square meter of data center space.

Software and external adoption

Google released TorchTPU as open source so PyTorch code can run on TPUs with minimal changes, and it supports inference frameworks such as vLLM and SGLang. Google now counts nearly 100 external TPU customers, reflecting a shift from treating TPUs as an internal tool to marketing them as a competitive platform. The company also frames its modular system approach as a deliberate design choice that can deliver new hardware variants in months rather than years.

Market context

As AI applications spread, the focus is shifting from expensive training runs to inference, meaning the day-to-day operation of models. Specialized chips such as TPUs can play to their strengths in that area. Amazon, Microsoft, and OpenAI are also developing their own accelerators, while Nvidia has bolstered its inference offering through an acqui-hire and licensing deal with Groq.

Sources: The Decoder