OpenAI seeks alternative chip suppliers

OpenAI, a major client of NVIDIA, is reportedly turning to Google to supply chips for its AI operations. According to reports, Google has begun offering its Tensor Processing Units (TPUs) to OpenAI, marking a significant development in the AI chip market.

OpenAI's growing computing needs

OpenAI has been seeking additional computing power to support its demanding AI models and has been exploring multiple avenues to achieve this. In June, the company unveiled its first custom chip designed with Broadcom, expected to be in service by the end of the year.

Google's TPUs gain external interest

According to The Information, OpenAI is preparing to purchase chips designed by Google, specifically TPUs, which are tailored for artificial intelligence (AI). Unlike NVIDIA's GPUs, which were originally designed for graphics, TPUs have been developed specifically for AI tasks.

Historical context of Google's TPUs

Google has been developing TPUs for over ten years, initially to accelerate its search engine and later for AI calculations. For a long time reserved for internal use or rented via Google's cloud, TPUs have been sold directly to certain clients since April, who can install them in their own data centers.

Strategic ties between OpenAI and Google

This would be a real turning point for OpenAI. If it has already rented Google chips in 2025, buying them is another matter. Especially since the ties between the two companies are closer than they seem: the team responsible for OpenAI's chips is led by Richard Ho, a former Google employee who worked on the TPU.

NVIDIA faces competition from Google

While this development may be a setback for NVIDIA, its GPUs remain the absolute reference in a market that it dominates comfortably. However, it is very good news for Google: its chips are now attracting the biggest names in AI, including its direct competitors.

Growing demand for AI chips

In February, Meta would have signed a contract worth several billion dollars to rent them and is discussing purchasing them as early as 2027. Oracle has also recently announced that it would offer these TPUs to its own cloud clients.

Google's strategic challenge

The trend is such that Google researchers must wait to access the processors. It now needs to capitalize on this opportunity. Alphabet expects to account for only a small portion of the revenue related to TPU by the end of the year, with the majority expected to arrive in 2027.

Expanding AI infrastructure spending

However, there is clearly a card to play: the cloud giants plan to spend up to 600 billion dollars this year on AI infrastructure.