Meta wants to use Google's chips to challenge Nvidia.

An agreement to use TPU chips developed by Google to power artificial intelligence (AI) models at technology company Meta could be worth billions of dollars and erode the dominant market share of chipmaker Nvidia.

According to a VNA correspondent in New York, on November 25, the Wall Street Journal (WSJ) reported that Meta Platforms is in negotiations to use chips manufactured by Google to accelerate its AI development, a step aimed at reducing its dependence on Nvidia.

A deal could be worth billions of dollars, but negotiations are not yet complete and may not reach a final conclusion.

According to an anonymous source contacted by the WSJ, it is currently unclear whether Meta will use these chips—known as tensor processors or TPUs—to train AI models or for inference. Inference is the process by which a trained model generates responses to queries and requires less computational power than training.

For years, Google has been working to refine its chips and expand its business in this area. A major deal with Meta would be an opportunity for Google and other chipmakers to challenge Nvidia's market dominance. Nvidia's stock fell 7% on the morning of November 25th following news of talks between Google and Meta.

Google says its Google Cloud service is seeing “accelerated demand” for both custom TPU chips and Nvidia GPUs, and the company is committed to supporting both types of chips, as it has done for many years.

Google began using TPU chips over 10 years ago, initially for internal purposes such as improving the efficiency of its search engine. In 2018, the company began offering its cloud customers access to TPUs for training and inference purposes.

More recently, Google has used these chips to train and run its large Gemini language models, and sells them to customers like Anthropic—a startup that develops the Claude model.

Last month, Anthropic announced it will spend tens of billions of dollars starting in 2026 to purchase 1 million TPU chips from Google, equivalent to about 1 gigawatt of computing power, to support AI research and meet the growing demand for the company's enterprise AI tools.

Currently, most large-scale programming language models are trained using Nvidia GPU chips. Thousands of application developers order chips from Nvidia either directly or through cloud service providers that deploy them in massive data centers.

In contrast, Google's TPU is a specialized integrated circuit (ASIC) chip, designed for a specific computing task, making these centers more energy-efficient.

(VNA/Vietnam+)

Source: https://www.vietnamplus.vn/meta-muon-su-dung-chip-cua-google-de-thach-thuc-nvidia-post1079390.vnp