According to people familiar with the matter, Google is negotiating with Samsung and plans to introduce the South Korean technology giant as a new foundry partner in the production of next-generation Tensor artificial intelligence chips. Google's current Tensor Processing Unit (TPU) is mainly produced by Broadcom and TSMC. However, as AI chip production capacity tightens in the future, Google is seeking to expand its foundry lineup on the existing basis.
According to "The Information", Google has launched negotiations with Samsung, hoping to use Samsung's wafer manufacturing capacity to produce a new generation of Tensor AI chips. According to reports, this new generation of TPU, internally codenamed "Icefish", will be manufactured using Samsung's 2nm process technology. However, Samsung is expected to only undertake part of the orders, and the main production tasks will still be handled by TSMC. Currently, this new generation of Tensor chips is still in the design stage and is scheduled to be officially put into mass production in 2028.
TSMC has long been regarded as the most reliable high-end chip manufacturing partner in the industry, and technology giants such as Apple and Google are highly dependent on its advanced process capabilities. However, as global demand for AI chips surges, and TSMC prioritizes securing orders for large-scale AI data centers, related production capacity becomes tight. This is considered to be one of the key reasons why Google is considering introducing new foundry partners such as Samsung.
In addition to Samsung and TSMC, Intel is also expected to get some orders. "The Information" reported earlier this week that Google has launched negotiations with Intel to evaluate the possibility of Intel contracting up to 3 million TPUs in 2028. In fact, this is not the first time that Google and Samsung have cooperated in the field of chip manufacturing. Samsung has previously manufactured Tensor chips for the Pixel series of smartphones.
Google's Tensor processing units are currently mainly deployed in cloud data centers for AI model training and reasoning. They directly compete with Nvidia's AI chip product line in the market, and the latter still dominates the field. By increasing the use of self-developed TPU, Google can not only reduce its dependence on third-party chip suppliers, but also create new revenue channels around its own chips. Google has begun to export this kind of computing power to the outside world, and has provided TPU to the AI company Anthropic for training its next-generation cutting-edge large model.