Google Cloud Next 2026: TPU 8t and 8i, the eighth generation splits into two chips

2026-04-22 · Publisher : Google Cloud · Domain : Cloud & DevOps

At Google Cloud Next 2026, Google announces the eighth generation of TPUs, for the first time as two chips: TPU 8t for training (121 exaflops per superpod) and TPU 8i for inference, with 80% better performance per dollar.

What changes

  • The eighth TPU generation splits into TPU 8t (training) and TPU 8i (inference).
  • A TPU 8t superpod has 9,600 chips and 121 exaflops; the Virgo network links more than one million TPUs.
  • TPU 8i: 384 MB of SRAM, 288 GB of HBM and 80% better performance per dollar (vendor figure).
  • A5X instances with NVIDIA Vera Rubin NVL72 expected later in 2026.

### Background On 22 April 2026, at Google Cloud Next, Google announces the eighth generation of its TPUs, which for the first time comprises two distinct chips and systems, designed for the era of AI agents. TPU 8t targets training; TPU 8i targets inference and reinforcement learning. Both will be offered to cloud customers soon, according to Google. ### Key facts - TPU 8t: nearly 3x the compute of previous generations; 9,600 chips per superpod, giving 121 exaflops and 2 petabytes of shared memory; up to more than one million chips in a single cluster. - TPU 8i: SRAM tripled to 384 MB, HBM memory of 288 GB, interconnect bandwidth doubled to 19.2 Tbps, and a collectives acceleration engine; 80% better performance per dollar for inference than the previous generation (vendor figures). - Virgo Network: up to 134,000 TPUs in a single fabric at one site, and more than one million across several sites. - A5X instances with NVIDIA Vera Rubin NVL72, expected later in 2026, and Axion N4A VMs for agent workloads. - Storage: Managed Lustre reaches 10 TB/s of bandwidth (10x more than a year earlier) and 80 petabytes of capacity. ### Why it matters Google now separates training and inference chips, which reflects the difference in needs between the two uses. For architecture teams, the choice between TPU and GPU becomes sharper: Google offers both, with scaling to the million-chip level. The announced performance is the vendor's own, and pricing is not yet known. ### Takeaway The eighth TPU generation specializes training and inference: AI silicon is segmenting.

Links

Tags : Google Cloud, Puces, Hyperscaler, IA, Conférence