In May of the year before last, Meta Platforms revealed details about its MTIA family of custom-designed chips, intended to accelerate its artificial intelligence systems. According to Reuters, training of Meta’s AI models using these chips has already begun.
Currently, training is being conducted on a limited scale with a small number of these chips. However, if the trials prove successful, Meta plans to move toward mass production. By transitioning to its own components, the company aims to reduce the cost of maintaining its computing infrastructure. This year, Meta’s expenses are expected to range between $114 billion and $119 billion, with $65 billion allocated for infrastructure development.
Efficiency and Production Challenges
Meta’s proprietary chips are designed to lower energy costs when performing AI-related calculations. Compared to general-purpose GPUs, these specialized solutions could offer higher efficiency. The production of these chips will be handled by the Taiwanese company TSMC. Sources suggest that preparation for manufacturing and the release of an initial batch could cost Meta hundreds of millions of dollars. If this first batch fails to meet expectations, additional development could require comparable costs and several months of waiting.
Reports indicate that Meta has previously canceled a reasoning AI chip project at earlier development stages, redirecting funds toward acquiring Nvidia accelerators instead. Despite its push for proprietary hardware, Meta remains one of Nvidia’s largest customers and does not plan to completely phase out Nvidia accelerators.
Future Plans and AI Integration
Last year, Meta started experimenting with its chips for reasoning AI within its social networks, remionds NIX Solutions. The company recently announced that it will begin using its own chips for training language models starting in 2026. Initially, these chips will work alongside recommendation algorithms before being adapted for generative AI applications.
Meta’s leadership acknowledges that progress in AI chip development has been inconsistent. However, the first generation of chips for reasoning AI has been deemed a success. We’ll keep you updated as more developments unfold regarding Meta’s AI hardware strategy.