OpenAI has started using Google’s AI chips, known as tensor processing units (TPUs), to power ChatGPT and other products, according to The Information.

The chips are rented through Google Cloud and are expected to lower costs related to inference computing, where AI models make predictions based on trained data.

This marks a shift from OpenAI’s earlier reliance on Nvidia GPUs and Microsoft-backed data centers.

However, Google is reportedly not providing its most advanced TPUs, as it competes with OpenAI in the AI space.

OpenAI has been one of the biggest buyers of Nvidia GPUs for training and inference.

Neither OpenAI nor Google has commented on the matter.

🔗 Source: Reuters

🧠 Food for thought

1️⃣ Compute costs are driving strategic partnerships in the AI arms race

OpenAI’s move to Google TPUs highlights how computational costs have become a critical competitive factor in AI development.

With computing projected to exceed 80% of OpenAI’s total expenses in 2025, finding cost-efficient alternatives has become imperative for AI companies1.

Google’s TPUs reportedly operate at approximately 20% of the cost of Nvidia’s GPUs, offering significant savings for inference computing (when AI models generate responses)1.

This cost differential is especially important as training and running large language models becomes increasingly expensive. ChatGPT reportedly cost over $100 million to train, and daily operational costs continue to grow as usage expands.

The partnership demonstrates how even fierce competitors in the AI space, like Google and OpenAI, will collaborate when facing shared challenges such as infrastructure costs and compute constraints.

2️⃣ The AI chip landscape is diversifying beyond Nvidia’s dominance

Nvidia has maintained over 70% of the GPU market share, becoming the dominant provider of AI computing hardware since the 2012 ImageNet competition demonstrated GPU effectiveness for deep learning2.

Recent OpenAI developments