Alphabet Unveils 8th-Gen TPUs, Claiming 80% Better AI Performance per Dollar

2 articles · Updated · CNBC · Jun 27

Alphabet’s new eighth-generation TPU lineup splits for the first time into two chips: TPU 8t for AI training and TPU 8i for inference, targeting the industry’s shift toward cheaper large-scale model deployment.
Google said the chips are up to 3 times faster for training, deliver 80% better performance per dollar, and can scale to clusters of more than 1 million TPUs.
That efficiency is central to Alphabet’s push against Nvidia, whose GPUs still dominate AI computing but are costlier, more power-hungry and harder to secure in tight supply chains.
Google is already monetizing the chips through Gemini, cloud rentals and direct hardware sales; Google Cloud backlog reached $472 billion in the first quarter, and analysts see TPU infrastructure revenue rising from about $3 billion in 2026 to $25 billion in 2027.
The TPU push also underpins broader expansion, including Anthropic demand, a multi-billion-dollar Meta deal and a Blackstone venture backed by $5 billion in initial equity and 500 megawatts of planned capacity by 2027.

Sources

Left100%

CNBC18h ago

Alphabet Unveils Eighth-Generation TPUs for Specialized AI Training and Inference

thegputrade.com18h ago

Google readies TPU 8t/8i as hyperscalers diversify AI silicon

Can Google's cheaper custom chips truly break Nvidia's powerful software monopoly?

As nations build 'sovereign AI,' will the global tech market fracture into competing blocs?

Google Challenges Nvidia with TPU 8t/8i: 3x Faster AI Training, 80% Cost Gains, and a New Era of Specialized Chips

Overview

In June 2026, Google Cloud unveiled its eighth-generation Tensor Processing Units (TPUs), marking a major step forward in AI hardware. This new generation introduces two specialized chips: the TPU 8t for training AI models and the TPU 8i for inference, moving away from previous unified designs. This split reflects Google's response to the growing complexity and diversity of modern AI, as the industry enters what Google calls the 'agentic era.' Building on a legacy that evolved from inference-only chips to full training capabilities, these TPUs are now optimized to meet the unique demands of both training and inference workloads.

...