Google announced two new AI accelerator systems this week. The TPU 8tx is built for enterprises and government agencies that need to keep data encrypted while it is being processed. The TPU 8i Max stacks two eighth‑generation TPU 8i chips in a single package to double compute, memory and bandwidth within the same rack. Both variants use the same underlying silicon as Google’s existing eighth‑generation TPUs, but they come in different packaging, networking and memory configurations. Google says the modular approach lets it deliver new hardware in months rather than years.
The TPU 8tx closes a security gap: data and model weights stay encrypted even during active use. Google compares it to Nvidia’s confidential computing features for GPUs. The system runs on Google’s low‑latency Boardfly network and claims up to twice the latency reduction versus the older TPU 8t. Google says a single 8tx system can fine‑tune models, distill large models, and train reinforcement‑learning agents without moving sensitive data between clusters.
The TPU 8i Max targets the memory bottleneck that appears when AI agents handle large codebases or hours of video. By placing two TPU 8i chips together, the new system doubles memory capacity and bandwidth, delivering roughly double the compute power in the same rack footprint. Google says this translates into more performance per square metre of data‑center space.
Beyond the hardware, Google is easing TPU adoption. It released TorchTPU as open‑source so PyTorch code can run on TPUs with minimal changes, and it added support for popular inference frameworks like vLLM and SGLang. The company now counts nearly 100 external customers using its TPUs, a shift from its earlier internal‑only strategy.
Why the new TPUs matter
Enterprises that handle confidential data can now run AI workloads without decrypting information, reducing the risk of leaks. Developers working on large‑scale inference tasks, such as video analysis or code‑base search, get more memory and speed without expanding their data‑center footprint.
Why it matters
Companies that need to protect sensitive data can run AI models without exposing raw information, lowering security overhead. Teams that run large inference jobs gain more performance per rack, which can cut hardware costs and space requirements.
