Google TPU 8t and 8i: Inside the AI Chip That's Actually Worrying Nvidia in 2026

# Google TPU 8t and 8i: Inside the AI Chip That's Actually Worrying Nvidia in 2026

> **Quick answer:** Google has launched two new AI chips — the TPU 8t for model training and the TPU 8i for inference — and is now selling them directly to AI labs including OpenAI, Meta, and Anthropic. Independent benchmarks suggest Google's chips can deliver roughly 62% lower cost per effective training FLOP than Nvidia's best. Nvidia still controls 81% of the AI chip market, but 2026 marks the first year that's meaningfully in question.

Google TPU 8t and 8i represent the most significant challenge to Nvidia's AI chip dominance since the GPU era began. At Google Cloud Next 2026, Alphabet did something it had never done before: announced it would sell its custom Tensor Processing Units directly to select customers who want to run them in their own data centers — not just rent capacity from Google's clouds. The customers willing to try? OpenAI. Meta. Anthropic. The very companies that built the AI era on Nvidia silicon.

## What Google Just Announced: TPU 8t and TPU 8i

Google's eighth-generation TPU lineup is built around a deliberate split. Previous generations were general-purpose; the new ones are specialized:

**TPU 8t** is purpose-built for AI model training. It handles the computationally intense process of teaching a model from scratch — the weeks-long, power-hungry phase where gradient calculations happen billions of times per second. Google claims the TPU 8t delivers up to 3x faster training than its predecessor, with 80% better performance per dollar.