Google Unveils TPU 8t and TPU 8i Chips

Google announced two new eighth-generation tensor processing units, the TPU 8t (training) and the TPU 8i (inference), in a Google Cloud Next blog post on April 22, 2026 (cloud.google.com). The blog post also listed expanded infrastructure offerings including A5X bare metal instances powered by NVIDIA Vera Rubin NVL72, Axion N4A VMs, powered by our custom Axion Arm-based CPUs, the Virgo Network data-center fabric, Google Compute Engine 4th-generation VMs, and Google Cloud Managed Lustre (cloud.google.com).

Tech press coverage frames the announcement as a renewed effort to provide an alternative to Nvidia for large-scale AI compute. CNBC reports the company is splitting training and inference across distinct chips and says both TPUs will become available later this year (CNBC). TechCrunch and other outlets report Google-stated performance claims: up to 3x faster model training, 80% better performance per dollar, and the ability to link more than 1 million TPUs in a single cluster (TechCrunch).

Public reporting emphasizes the architectural choice to separate training and inference silicon. Industry coverage notes this mirrors parallel moves by other hyperscalers to build purpose-built processors for specific AI workloads (TechCrunch; CNBC). Reported performance claims-training speedups and cost-per-performance improvements-are presented by media as Google-provided figures rather than independent benchmarks (TechCrunch).

The split-training/inference approach reflects an industry pattern where providers optimise for divergent workload profiles. Training workloads are memory- and compute-intense with long-lived precision needs; inference workloads prioritize latency, memory bandwidth, and cost-efficiency. Companies adopting separate silicon for these roles typically aim to reduce end-to-end cost and energy use for large-scale agentic applications, a trend highlighted across coverage of this announcement (cloud.google.com; TechCrunch; CNBC).

Reporting places Google's move in the broader competitive landscape where hyperscalers and chip firms (notably Nvidia) coexist in customer clouds. Several outlets note Google will continue to offer Nvidia-powered instances, including the Vera Rubin NVL72, alongside its TPUs (cloud.google.com; TechCrunch; Barron's). Coverage cites prior commercial traction for Google TPUs with customers such as Meta and Anthropic, and frames the new chips as an attempt to expand that ecosystem (Bloomberg).

Observers and practitioners should track independently verified benchmarks for the TPU 8t and TPU 8i, cloud pricing and instance availability, and real-world scaling behavior when integrated into multi-accelerator training and agentic inference pipelines. Also watch third-party support across frameworks and model vendors, and whether customers port large frontier models to Google TPUs at scale-these outcomes will determine competitive impact more than vendor claims alone.

Amin Vahdat, Google SVP and chief technologist for AI and infrastructure, is quoted in CNBC: "With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving," (CNBC).