Google Cloud unveiled two new eighth-generation Tensor Processing Units at Google Cloud Next on Wednesday, splitting the chip line into TPU 8t for training and TPU 8i for inference. A single TPU 8t superpod scales to 9,600 chips with two petabytes of shared high-bandwidth memory, delivering 121 ExaFlops and three times the processing power of the prior Ironwood generation. The inference-focused TPU 8i connects 1,152 chips per pod with 3x more on-chip SRAM to run millions of AI agents concurrently at low latency.
Google claims up to 3x faster AI model training, 2.8x better price/performance and 2x better performance-per-watt versus the previous generation, with the ability to coordinate more than one million TPUs in a single cluster. The move aims to cut frontier model development cycles from months to weeks as Alphabet pushes its custom silicon against Nvidia, which now commands a nearly $5 trillion market cap.
Despite the push, Google is not abandoning Nvidia. The company said it will offer Nvidia’s upcoming Vera Rubin chip to Google Cloud customers later this year and continues to co-develop networking optimizations, positioning TPUs as a complement rather than a replacement.
