Google's eighth-generation TPU for large-scale AI pre-training, scaling to 9,600 chips per superpod.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
TPU 8t is the training-focused system in Google's eighth-generation Tensor Processing Unit family, introduced at Google Cloud Next 2026 alongside TPU 8i for inference and reasoning. Google positions it for large-scale pre-training and embedding-heavy workloads.
A single superpod scales to 9,600 chips and, per Google, delivers nearly 3x the compute performance per pod compared with the previous generation (Ironwood). A YouTube creator summary cites 2 PB of shared memory, 121 exaflops per superpod and a goodput target above 97%; these figures were not confirmed on the scraped official page. Google also states TPUs can be clustered up to 1 million chips.
TPU 8t is accessed through Google Cloud as part of the AI Hypercomputer stack. The software stack includes JAX, PyTorch via TorchTPU, OpenXLA, MaxText, Tunix and vLLM, with GKE for orchestration. The official page marks TPU 8t as "Coming soon", so it is not yet generally available. No TPU 8t pricing is published; the pricing page lists only earlier generations (Ironwood, Trillium, v5p, v5e and older).
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Eighth-generation TPU built for training. A superpod scales to 9,600 chips with 2 PB shared memory and 121 exaflops, with 3x compute per pod versus the previous generation and a goodput target above 97%.
Closest neighbours by meaning, not just by category.
Hardware · Google
Google's eighth-generation Cloud TPU for post-training and low-latency inference of large MoE models.
Hardware · Google
Google's custom AI accelerators, offered through Google Cloud for training and inference of machine learning models.
Hardware · Google
Google research initiative to put TPU-equipped satellites in orbit as a prototype space-based AI data center.
Hardware · Engine AI
Full-size T-800 humanoid robot from Engine AI, shown in a teleoperated humanoid MMA event in Shenzhen.
Hardware · OpenAI
Custom AI inference chip announced by OpenAI and Broadcom, built to serve ChatGPT-style workloads faster and cheaper.
Model
Open model family (9B dense, 35B MoE, 397B MoE) trained with a self-improvement loop for reasoning, coding and agentic tasks.
