Meta Begins Testing In-House AI Training Chips

Meta has begun testing its own chips designed for training artificial intelligence.

According to Reuters, the company has initiated a small-scale deployment and is working with TSMC to explore potential mass production if the experiment proves successful.

The testing phase started after completing the tape-out stage—the final step of development before production begins. This process takes between three to six months and requires investments in the tens of millions of dollars, with no guarantee of success. If the project fails, Meta will have to go through the entire design process again.

Developing in-house components will allow the company to reduce costs on computing infrastructure. In 2025, Meta's expenses are estimated to be between $114 billion and $119 billion, with $65 billion allocated to AI infrastructure development.

According to Meta’s plans, the company intends to implement its custom chips by 2026. Initially, they will be used in recommendation algorithms and later in generative AI products, including the Meta AI chatbot.