Google is accelerating the transition of its self-developed Tensor Processing Units (TPUs) from an internal tool to a commercial AI chip available for sale, directly challenging NVIDIA's dominance in the AI hardware market. Through a partnership with Broadcom, Google's TPU business has expanded to provide complete AI infrastructure solutions to external clients such as Anthropic, marking a fundamental shift in its TPU strategy.

Google (3)

Significant cost advantages and accelerated commercialization

Currently, NVIDIA still holds a dominant position in the market with about 86% of revenue from data center chips, but Google's TPU is beginning to challenge this status quo with cost and system advantages. According to reports, the cost of processing AI workloads using Google's self-developed TPU is 30% lower than that of competitors, a significant advantage in large-scale deployment. Google has reached an agreement with Anthropic, which will deploy up to 1 million seventh-generation TPUs for training the Claude model, marking Google's first direct competition with NVIDIA as a hardware supplier.

In terms of products, Google has released eighth-generation TPUs optimized for both training and inference tasks, with plans to launch them later this year. Currently, about 75% to 80% of TPU capacity is still used for internal operations, but analysts predict that production capacity will further expand, with an estimated annual TPU output of 5 million units by 2027.

Considered the most structural threat to NVIDIA

Analysts believe that Google's external sales strategy for TPUs has been seen as the most structural threat to NVIDIA's dominance in the AI chip market. Although Google still faces challenges in software ecosystems, successful case studies with its customers and growing external demand are gradually highlighting the feasibility of Google as an alternative to NVIDIA. The direction of this AI chip competition will profoundly impact the future of the AI computing power market.