Google Cloud revealed on Wednesday that its eighth generation of custom AI chips, known as tensor processing units (TPUs), will be divided into two distinct versions. The TPU 8t is designed for training AI models, while the TPU 8i focuses on inference—the process of running models after users submit prompts. The company highlights significant performance improvements over earlier generations, including AI model training that is up to three times faster, an 80% boost in performance per dollar, and the capability to link over one million TPUs into a single cluster. This advancement promises substantially greater computing power with lower energy consumption and reduced costs for customers.

EMBED_PLACEHOLDER_0

These chips are called TPUs rather than GPUs because Google’s custom low-power chips were originally named Tensor. However, Google’s approach does not represent a direct challenge to Nvidia’s dominance, at least for now. Similar to other major cloud providers like Microsoft and Amazon, Google is using its custom chips to complement the Nvidia-based systems available in its infrastructure, not to replace them entirely. In fact, Google has committed to offering Nvidia’s latest chip, Vera Rubin, on its cloud platform later this year.

Over time, as enterprises increasingly shift their AI workloads to the cloud and adapt their applications to run on custom chips, hyperscalers such as Amazon, Microsoft, and Google may rely less on Nvidia. Yet, under current conditions, betting against Nvidia remains unwise. As chip market analyst Patrick Moore humorously noted on X, he had predicted back in 2016—when Google launched its first TPU—that it could pose a threat to Nvidia and Intel. Nvidia has since grown to a market capitalization of nearly $5 trillion, demonstrating that such forecasts did not stand the test of time.

If Nvidia’s strategy unfolds as planned, Google’s expansion as an AI cloud provider could actually generate more business for the chipmaker, even as many workloads run on Google’s own chips. Notably, Google has also announced a collaboration with Nvidia to enhance computer networking, aiming to improve the efficiency of Nvidia-based systems within its cloud. Specifically, the two companies are working to strengthen Falcon, a software-based networking technology that Google developed and open-sourced in 2023 through the Open Compute Project, a leading open-source data center hardware organization.