Meta’s In-House Chips May Conserve Power and Cash

While Nvidia’s chips currently rule the market, Meta is looking at alternatives. 

The company has started testing an in-house chip for training AI systems, according to a report from Reuters. Meta’s chip is a dedicated accelerator, one source told Reuters, meaning it’s designed to handle only AI tasks and can save on power compared with conventional GPUs. 

The move by Meta isn’t surprising: The chip is part of its Training and Inference Accelerator series, an effort to build a family of chips to power the company’s AI workloads. The company has also previously said that it intends to start using its own chips for training by 2026. The effort follows similar endeavors by fellow tech giants, like Amazon’s Trainium and Inferentia chips,  and Google’s Tensor Processing Units.

As AI development threatens to devour a growing amount of power, the moves could be the vanguard of a search for energy-efficient means beyond conventional GPUs, said Justin St-Maurice, technical counselor at Info-Tech Research Group.

• With these chips, Meta has the opportunity to consider AI hardware from a “value per watt perspective,” which “doesn’t seem to be getting better with Nvidia,” said St-Maurice. 
• “That brute force, high-energy use approach, we’re starting to see more and more examples of challenges to that,” he said. 

“There’s a lot of work to be done around how energy intensive and how costly these models need to be to train and deploy,” said St-Maurice. 

Meta’s move might also help get its massive infrastructure costs under control. The company expects to have $65 billion in capital expenditures in 2025 up from $38 billion in 2024, with much of that cash being directed toward AI infrastructure. Plus, the interest in in-house chips could signal a desire for self-reliance – and for more competition, said Rodolfo Rosini, CEO and co-founder of Vaire Computing.

“It’s an issue of injecting competition in one of the key areas,” said Rosini. “Because the moment that they exclusively buy from one vendor, that vendor can increase the prices.” 

If other companies follow in the footsteps of the hyperscalers, it could spell trouble for Nvidia. While the chip giant currently dominates the market, Meta’s move indicates that it doesn’t want to “let Nvidia be the only innovator” and may be searching for efficient alternatives, said St-Maurice. 

“The question is whether or not other companies will find other ways to do the same thing, more efficiently, more cost effectively, and start to create their own chips that can effectively do that without having to buy the premium Nvidia product,” said St-Maurice.

The problem, however, is resources. Most firms don’t have access to the same cash or capabilities as tech giants like Meta, Amazon and Google to design and build their own chips. Without the proper resources, Rosini said, high quality and performance are difficult to achieve. “To match the performance of Nvidia, you have to put in Nvidia money,” said Rosini.