Google Pitches TPUs in Challenge to Nvidia’s AI Chip Dominance

Google’s Tensor Processing Units were developed to power its own search, cloud and generative AI workloads, but are increasingly being offered to outside customers through Google Cloud. As companies seek alternatives to Nvidia’s GPUs for training and inference, the economics, availability and performance of TPUs have become central to Google’s effort to monetize AI infrastructure and compete more directly with the world’s largest semiconductor suppliers.

Morgan Stanley estimates Google’s TPU business could generate $108 billion in annual revenue by 2028, a scale that would place it among the world’s leading semiconductor companies. Google is also accelerating its chip-release cadence and tailoring processors separately for large-scale model training and low-latency inference. The moves suggest its in-house computing platform is evolving into a standalone commercial chip business serving external customers.

Google began developing its own TPUs for search and AI workloads in 2013 and later made them available through Google Cloud. It now plans to sell the chips directly, challenging Nvidia, which controls more than 90% of the AI-chip market and has reinforced its advantage through CUDA. Although neo-cloud providers want to diversify their supply, they fear losing their Nvidia GPU allocations if they switch to TPUs, a bind known as “Jensen Jail.”

In May 2026, Google announced a push to sell TPUs directly and provided a $3.2 billion guarantee for Lake Mariner, a project being developed by TeraWulf and FluidStack for use by Anthropic. On May 19, Google also formed a TPU cloud joint venture with Blackstone, which committed $5 billion. The venture is expected to bring 500 MW of capacity online in 2027.

Google is expanding the supply chain for its custom tensor processing units to reduce its reliance on NVIDIA AI chips. Broadcom and MediaTek are involved in chip design, while TSMC handles advanced manufacturing. The strategy could help cloud providers control inference costs and secure supplies while challenging NVIDIA’s dominance of the AI accelerator market.

Google unveiled Ironwood, its seventh-generation TPU, in April 2025 and said it delivered 10 times the performance of its predecessor. Broadcom recently confirmed that Google was diversifying its suppliers, while market reports said MediaTek had secured orders for the TPU v8. Google expects to deploy chips manufactured using TSMC’s 2-nanometer process in 2027. The value of the related partnerships has not been disclosed.

Google has long developed its own Tensor Processing Units, or TPUs, to train and run models for internal AI services and Google Cloud customers, reducing its reliance on NVIDIA GPUs. Its reported talks with networking and custom-chip specialist Marvell could determine whether Google can expand its supply of AI computing capacity and strengthen its competitiveness in the cloud market.

The latest reports say Google is in talks with Marvell to jointly design two AI chips: a TPU memory processing unit, or MPU, and a next-generation TPU. The custom chips are reportedly set to use Intel’s Intel 18A process, with design work potentially completed as early as 2027. The value of the partnership has not been disclosed. The news sent Marvell shares up 6.3%.

Competition in generative AI depends heavily on the coordination of chips, data centers and models. Google has long developed its own Tensor Processing Units, or TPUs, integrating them with Gemini models and Google Cloud services. The strategy reduces its reliance on external AI chips as it competes with major cloud providers for enterprise computing demand.

Google Cloud CEO Thomas Kurian recently said Google now controls about 25% of global AI computing capacity, equivalent to one-quarter of the market. Its integration of in-house TPUs with Gemini has helped narrow the gap with cloud rivals. The report did not disclose the reference date for the estimate or any additional investment amount.