Google Unveils AI Hypercomputer Combining TPUs, GPUs, CPUs

Google announced the AI Hypercomputer at Cloud Next 26, a unified datacenter architecture that pairs its new 8th Gen TPU family with Axion Cloud CPUs and NVIDIA Rubin GPUs to target the Agentic AI era. The flagship training silicon, TPU 8t, is advertised to scale to 9,600 chips and 2 petabytes of shared high-bandwidth memory per superpod, delivering 121 ExaFlops of compute and native FP4 arithmetic.

The TPU 8t emphasizes throughput, interchip bandwidth, and memory capacity to reduce frontier-model training timelines from months to weeks. Key features called out include:

• •Massive scale: single superpod of 9,600 chips with 2 petabytes of shared HBM and double the interchip bandwidth versus prior gen
• •Data-path optimization: TPUDirect and 10x faster storage access to reduce IO stalls and improve utilization
• •Software scaling: near-linear scaling claims with JAX and Pathways across very large logical clusters

The announcement also introduces a second SKU, TPU 8i, and positions Axion Cloud CPUs for host orchestration and mixed workloads while allowing customers to augment capacity with NVIDIA Rubin GPUs, creating heterogeneous clusters.

This is an infrastructure-first response to agentic, stateful, and large-context models that demand huge memory pools and low-latency interconnects. By combining in-house TPUs with third-party GPUs and custom CPUs, Google is offering both a vertically integrated stack and a hybrid path for customers who need GPU-based tooling. The FP4 support and emphasis on system-level IO are direct optimizations for massive parameter models and retrieval/agent workflows.

Actual availability, pricing, and filesystem/network topology details will determine enterprise adoption. Verify real-world scaling beyond vendor benchmarks and how orchestration integrates heterogeneous TPUv8 and Rubin resources into single-job scheduling.