OpenAI is reported to be seeking to buy the tensor processing unit (TPU), an artificial intelligence (AI) chip developed in-house by Google. If the two companies' cooperation materializes, Nvidia, which has dominated the AI chip market with its graphics processing units (GPUs), would face a powerful rival, observers say.
U.S. IT outlet The Information reported on the 27th (local time) that OpenAI plans to buy Google TPUs. Foreign media including The Information, Reuters and Bloomberg reported that details of the deal, such as purchase volume and price, have not been disclosed.
OpenAI has also posted job listings for TPU-related engineers. According to the listing as relayed by The Information, OpenAI said, "We are looking for engineers to optimize OpenAI's cutting-edge frontier models on TPUs," and "We welcome talent with experience developing TPUs in training or inference." The outlet interpreted this as "an early sign that OpenAI has begun developing its own custom TPUs."
OpenAI is one of Nvidia's major customers. In August, Nvidia pledged $100 billion in credit support for a data center project in Ohio that OpenAI is to lease. Nvidia CEO Jensen Huang said at the time, "OpenAI plans to invest a total of $600 billion in using Nvidia servers through 2030." Analysts say OpenAI, which is preparing an initial public offering (IPO) next year, needs to broaden its chip sources.
Rival Anthropic also said in its IPO filing with the U.S. Securities and Exchange Commission (SEC) on the 26th (local time) that it uses four types of chips: Google TPUs, Nvidia GPUs, Amazon Trainium and AMD GPUs.
Google has so far leased TPUs to multiple customers including Apple, Anthropic and Citadel, and recently has been expanding beyond leasing into the business of selling TPUs directly. Foreign media projected that a deal with OpenAI would be a major boon to Google's semiconductor business. They also added that demand for high-bandwidth memory (HBM) going into Nvidia GPUs as well as Google TPUs is expected to rise further.