Alphabet Launches Eighth-Generation TPUs to Rival Nvidia AI Hardware

At the Cloud Next 2026 conference, Alphabet Inc. announced its eighth-generation Tensor Processing Units (TPUs), introducing two specialized chips designed to challenge Nvidia's dominance in the AI hardware market. Developed in collaboration with Broadcom Inc, the TPU 8t focuses on AI model training, offering three times the compute performance of its predecessor to reduce development cycles from months to weeks. The TPU 8i is optimized for inference and reasoning, providing 80% better performance-per-dollar to support low-latency requirements for collaborative AI agents.

Both chips utilize liquid cooling and pair with Google's Axion ARM-based CPU host. To support massive scaling, Alphabet introduced the Virgo Network, which enables up to one million TPUs to operate within a single logical cluster. The company uses this hardware to train its Gemini foundational model and power its Search business more cost-effectively than competitors using general-purpose GPUs. Through a partnership with Broadcom Inc, select customers can purchase these chips for use in Google Cloud or external data centers.

Despite the push for internal hardware, Alphabet maintains a strategic relationship with Nvidia. The company announced an engineering partnership to improve networking via the open-source Falcon technology and committed to integrating Nvidia's Vera Rubin chip into its cloud services later in 2026. The eighth-generation TPUs are expected to be generally available later this year.