The two new configurations
Google is expanding its eighth-generation artificial intelligence chip family with two additional systems, the TPU 8tx and the TPU 8i Max, bringing the number of TPU systems launched within a year to four. AI News reports this. Neither is a new architecture: they are different configurations of the silicon already presented in April at Google Cloud Next, when the company divided a generation of TPUs into two distinct projects for the first time, the TPU 8t for large-scale training and the TPU 8i for inference.
The TPU 8i Max pairs two TPU 8i chips in a single package. According to Google, the result is double the compute power, SRAM memory and HBM memory bandwidth compared with a standard TPU 8i, with twice as much silicon that can be hosted in the same rack and, for the same amount of compute in data centers, less space occupied and lower power consumption. The stated goal is models that must keep large amounts of text or data in memory within a single request.
Security and networking for inference
The TPU 8tx reuses the same silicon as the TPU 8t, dedicated to training, but pairs it with fifth-generation AMD EPYC processors. A secure enclave integrated at the hardware level isolates and protects sensitive data and model weights. The system also adopts the Boardfly network, developed for real-time reasoning on the TPU 8i; according to Google, it halves latency compared with the TPU 8t with almost no performance penalty in exchange for greater security.
The choice responds to a concrete demand from companies: running models on confidential data and handling long contexts without multiplying costs. The material refers to banking, healthcare and public administration, sectors in which data confidentiality affects the possibility of bringing a model into production. In parallel, Google announced a suite of agents for TPUs, open source and in preview, which connects to programming tools such as Claude Code, Cursor and Google Antigravity to automate chip adoption, identify bottlenecks and tune custom kernels.
The comparison with Nvidia remains open
For the two chips presented in April, Google claims up to 2.7 times the training performance and an 80% higher performance/price ratio in inference compared with Ironwood, the seventh generation. However, it has not published specific benchmarks, power consumption data or prices for the TPU 8i Max: a direct comparison with Nvidia's Rubin systems on efficiency and cost is therefore not yet possible. Until those numbers are available, the competition with Nvidia remains on paper, and the first cloud customers will indicate how much weight it actually carries.
The main limitation of the material is precisely this: the performance of the new systems is what the company claims, not measured by third parties, and data on consumption and prices are missing. The available information makes it possible to reconstruct the offering and the stated aims, but not to assess the real impact on the artificial intelligence data center market.
