According to sources, OpenAI plans to purchase Google’s AI chips, a significant development for NVIDIA, OpenAI’s investor and rising competitor.
The financial terms and other details of OpenAI’s proposed purchase are currently unknown. However, if the TPU chip procurement deal is finalized, it would be a major win for Google’s chip business. NVIDIA’s GPUs have long dominated the AI computing market, and Google’s TPUs are increasingly challenging this dominance. OpenAI is one of NVIDIA’s most valuable customers, and in return, NVIDIA has invested tens of billions of dollars into the startup.
In August this year, NVIDIA secured over $100 billion in credit support for a large data center built entirely with NVIDIA chips for OpenAI. At the time, NVIDIA’s CEO Jensen Huang wrote in a blog post that OpenAI committed to spending approximately $600 billion on NVIDIA servers by 2030.
While planning to procure Google chips, OpenAI is hiring an engineer whose job description requires the candidate to "lead the optimization of OpenAI’s most advanced models on TPUs."
Meanwhile, Google's chip business also achieved another milestone this week: On Tuesday, it officially announced that Oracle will become the first major cloud provider outside of Google to offer TPU computing power for rent.
Sources familiar with the agreement said that Oracle directly purchases TPU chips, and the joint major clients of both companies will actually use the TPU computing power, although the identities of the clients have not been disclosed.
OpenAI is already a major customer of Oracle Cloud; last week, OpenAI also announced that it will be listing its models on Google Cloud. OpenAI has also been renting NVIDIA chips on Google Cloud for some time.
Google developed TPUs over a decade ago, and the company claims that TPUs offer better energy efficiency than NVIDIA GPUs. For years, Google has rented out TPU computing power through its own cloud, serving clients such as Anthropic (which uses TPUs to train models), Apple, trading firms like Citadel Securities, and numerous startups. Recently, Google has begun supplying TPUs to other data center operators, with emerging cloud providers like Fluidstack being the first partners.
The Information reported last year that OpenAI had begun testing Google’s TPUs to reduce its reliance on NVIDIA chips. After the news broke, NVIDIA CEO Jensen Huang swiftly took action, leading NVIDIA to renegotiate and reach a new cooperation agreement with OpenAI. Subsequently, an OpenAI spokesperson stated that the company “currently has no definite plans to use TPUs.”
The job description from OpenAI indicates that model optimization for TPU is still in its early stages. This role will be fully responsible for this technical area. The job posting states: “Candidates must have real-world workload optimization experience on TPU, with experience in either training or inference.”