Google unveils two new TPU systems, one of which cuts latency in half
Key Facts
Google unveiled two new AI computing systems on November 9, one focused on confidential computing and inference, and the other on expanding compute and memory for workloads with large contexts. The company said TPU 8tx cuts latency by up to half compared with TPU 8t, with near-zero performance penalty. The 8tx includes fifth-generation AMD EPYC processors as host CPUs, while the AI accelerator itself remains designed by Google. TPU 8i Max, meanwhile, aims to increase compute and memory density by combining two TPU 8i chips in one package. The announcement comes as Google Cloud and its TPU systems business expand; the company says its customer base is approaching 100 organizations.
Google links TPU 8tx to workloads that require sensitive data to remain within a single computing environment. The company said the system enables teams to tune models using their own data, distill large models into smaller ones, and train agents with reinforcement learning without moving sensitive data between separate compute clusters. The design places data processing and model weights among the confidential-computing use cases the company is targeting. Google says lower latency is important for running agents that execute multiple steps. Compared with TPU 8t, 8tx cuts latency by up to 2 times, according to the company, with near-zero performance penalty. These specifications position the system differently from 8i Max, which focuses on increasing compute and memory resources.
For its other system, Google combined two TPU 8i chips in one package under the name TPU 8i Max. The company said this configuration doubles compute, SRAM memory and HBM bandwidth compared with TPU 8i. The design targets workloads that need to handle large contexts, according to the company’s description. Combining the two chips increases processing, memory and bandwidth resources together, rather than positioning the product primarily around latency reduction. Google introduced TPU 8t and TPU 8i in April 2026, saying at the time that 8t delivered up to 2.7 times better training performance than the previous generation, while 8i offered an 80% improvement in price performance for inference. These figures place the two new systems within a broader development path for the company’s chips, with different stated objectives for each system.
The unveiling coincided with strong growth at Google Cloud, whose revenue more than doubled in the September 2026 quarter to $30.71 billion, according to Reuters. The agency quoted Alphabet as saying TPU sales contributed meaningfully to the quarter’s results. The customer base offers another indication of the systems’ expanding use: Google said it has nearly 100 customers, including advanced labs, trading and retail companies, and universities. Alphabet shares rose more than 3% in early trading on November 9 after the announcement, Reuters reported. The move came amid expanding demand for Google Cloud services, as the company links TPU performance to growth in its cloud business. The announcement did not provide pricing for either system or a specific date for broader availability.
On the software front, Google said TorchTPU, which enables PyTorch and Hugging Face models to run on TPU units, became open source in October 2026. It also said the TPU Agent Suite was in preview to simplify setup and diagnose performance bottlenecks. The company links these tools to making its systems easier for teams developing models and agents to use. According to Google, TPU 8tx enables teams to tune models with private data, distill models and train with reinforcement learning without moving sensitive data between compute clusters. 8i Max, meanwhile, delivers a stated increase in compute, SRAM memory and HBM bandwidth compared with 8i. Available sources did not give a specific next release date for the two systems or the software suite beyond its preview stage.