A Better Newspaper

## Overview At Google Cloud Next 2026, Google unveiled two new eighth-generation custom AI chips: the TPU 8t and TPU 8i, designed for training and inference workloads respectively (SiliconANGLE, April 22). ## Details Google said it designed the chip pair to address the next generation of AI workloads by splitting specialized architectures across training (TPU 8t) and inference (TPU 8i) use cases, rather than relying on a single general-purpose design (SiliconANGLE, April 22). This continues Google's long-running Tensor Processing Unit line, its custom silicon alternative to Nvidia GPUs for AI compute. ## Why It Matters Custom AI silicon is a central battleground in the AI infrastructure race, with Google, Amazon (Trainium), Meta (custom accelerators with Broadcom), and others pursuing in-house chip strategies to reduce dependence on Nvidia and control costs at scale. The bifurcation into distinct training and inference architectures reflects growing specialization in AI hardware design as inference workloads scale dramatically with agentic AI deployment. This has direct implications for cloud compute pricing, AI infrastructure investment strategy, and the broader Nvidia/CUDA vs. custom silicon dynamic. ## Developments to Watch - Performance benchmarks against Nvidia's latest GPUs and other custom silicon (Trainium, Meta's Broadcom-designed chips) - Adoption by Google Cloud customers and third parties - Pricing and availability details - Competitive response from Nvidia, AWS, and Microsoft