One of NVIDIA’s most important customers is placing orders with its competitors.

On October 27th, according to The Information, which cited informed sources, OpenAI plans to purchase Google's tensor processor (TPU) chips.

This is a significant breakthrough for Google's chip business. Google has long been seen as a potential competitor to NVIDIA GPU in the AI computing power market, and OpenAI is one of NVIDIA's most valuable customers.

Meanwhile, OpenAI has released a job advertisement seeking an engineer to "lead" the optimization of its most advanced models on TPU – this indicates that the related efforts have entered a substantial advancement phase.

NVIDIA and OpenAI: Tens of billions of dollars in deep ties

OpenAI has an extremely close relationship with NVIDIA.

In August this year, NVIDIA agreed to provide over $100 billion in credit support for a large data center project planned by OpenAI in Ohio, which will fully utilize NVIDIA's chip architecture.

NVIDIA CEO Jensen Huang wrote in a blog at that time that OpenAI had committed to spending approximately $600 billion on NVIDIA servers by 2030.

However, this has not prevented OpenAI from seeking diversified supplies.

It's not my first attempt, but this time I'm going one step further.

According to a report by The Information last year, OpenAI began using Google TPU to reduce its dependence on NVIDIA chips. As soon as this news emerged, Jensen Huang took immediate action, and NVIDIA subsequently reached a new agreement with OpenAI. A spokesperson for OpenAI stated at the time that the company "currently has no plans to use TPU exclusively."

Today, this statement has become a thing of the past.

The job advertisement released by OpenAI this time indicates that the related work is still in its early stages. The job description states that the position will have "extensive technical leadership" over the entire optimization process, and it mentions that "experience accumulated in training or inference on TPU as well as a solid foundation in optimizing real workloads on TPU is welcome."

Google TPU: From Domestic Use to International Expansion

Google has been developing TPU for over a decade and claims that its efficiency is superior to that of NVIDIA GPU.

For many years, Google has rented out TPU to external customers through its own cloud infrastructure, with users including Anthropic (which once used TPU to train models), Apple, as well as trading companies like Citadel Securities and numerous startups. Recently, Google has begun to sell TPU directly to other data center operators, initially targeting emerging cloud service providers such as Fluidstack.

On the same day, Google announced that Oracle will become the first major cloud service provider to rent TPU to customers. According to people familiar with the matter, Oracle is the one that directly purchases these chips, and the end-users will be common large customers of both companies, although their specific identities are not yet clear.

It is worth noting that OpenAI itself is an important cloud customer of Oracle. Last week, it announced that it would provide its model services through Google Cloud, and it has also been renting NVIDIA chips through Google Cloud prior to that.