Google plans specialized TPU chips to challenge Nvidia dominance
Google is developing a new generation of Tensor Processing Units with specialized chips for government and scientific workloads. The strategy aims to improve efficiency and reduce reliance on general-purpose GPUs from competitors like Nvidia.
Bloomberg reported that Google is planning a new generation of Tensor Processing Units, specifically the 8th generation, which will include four distinct chips. One chip is designed for confidential workloads for government agencies and regulated industries, while another, the 8i Max, is optimized for large context windows to support scientific research, mathematics, and coding. This strategy focuses on hardware specialization to improve efficiency over software solutions. Google’s chip revenue is projected to grow from $12 billion this year to over $100 billion by 2030. The company partners with Broadcom for production and sells chips to external customers like Oracle, amortizing development costs through internal use and external sales. This move intensifies competition with Nvidia, which has seen its stock decline partly due to these developments and the CEO scandal. The shift highlights a trend where hyperscalers are building custom ASICs to reduce reliance on general-purpose GPUs, aiming to capture market share through tailored hardware for specific AI workloads.