Google is accelerating the transformation of its self-developed Tensor Processing Units (TPUs) from internal tools to commercially sold AI chips, directly challenging NVIDIA's dominance in the AI hardware market. The tech giant has positioned TPUs as core components of its AI supercomputers. Through its partnership with Broadcom, Google's TPU business has expanded to offer complete AI infrastructure solutions to external clients such as Anthropic.
In the current AI chip market landscape, NVIDIA still holds a dominant position with approximately 86% of data center chip revenue. However, Google's TPUs are leveraging cost and system advantages to disrupt this landscape. Reports indicate that Google's in-house TPUs can process AI workloads at costs up to 30% lower than competitor processors, an advantage that becomes particularly significant in large-scale deployments.
On a strategic level, Google has completed the transition from purely internal use to full commercialization in recent years. Previously, Google reached an agreement with Anthropic, which will deploy up to 1 million of Google's seventh-generation TPUs to train its Claude model. This deal is reportedly Google's first competition with NVIDIA as a direct hardware supplier, marking a fundamental shift in its TPU strategy.
Simultaneously, Google has released its eighth-generation TPUs, specifically optimized for training and inference tasks respectively, with plans to bring them to market later this year. Currently, about 75% to 80% of the company's TPU production capacity is still used for internal operations, but analysts predict further expansion of its TPU capacity, with annual production potentially reaching 5 million units by 2027.
Analysis suggests that Google's external sales strategy for TPUs is viewed as the most structurally significant threat to NVIDIA GPU dominance in the AI chip market. Although Google still faces challenges in software ecosystems, the success stories of its clients and growing external demand are increasingly highlighting its viability as an alternative to NVIDIA.
Comments