Meta to Begin Mass Production of New AI Chip Starting in September

To ease the burden of GPU costs caused by parts shortages, Meta will begin mass production of the latest version of its in-house AI-dedicated chip, MTIA (Meta Training and Inference Accelerator), starting in September.

The chip's design will be handled by Broadcom, while manufacturing will be handled by TSMC. Meta also plans to source RAM from Samsung, storage from Sandisk, and optical fiber equipment from Sumitomo Electric. To keep pace with the rapid advancement of AI technology, Meta is adopting a modular chiplets structure as a strategy to shorten design cycles.

Meta plans to use this chip for the following purposes:

• Training models for recommendation algorithms
• Handling a wide range of AI workloads
• Inference for service applications

Meta expects capital expenditures of between $125 billion and $145 billion this year to secure AI infrastructure. This will focus on securing data centers and power for training and deploying new AI models such as the Muse Spark series.