Meta has introduced four new in-house artificial intelligence chips as part of its MTIA (Meta Training and Inference Accelerator) family. The first, dubbed MTIA 300, was recently rolled out to train relatively small AI models used in recommendation systems and content ranking on Facebook and Instagram, notably for selecting posts and advertisements.

Three other processors, MTIA 400, MTIA 450 and MTIA 500, are in the pipeline and will primarily target inference tasks tied to generative AI, such as creating images or videos from text. The MTIA 400 chip has already completed its testing phase and is expected to be installed soon in Meta's data centres, with racks able to host up to 72 processors. The subsequent models are expected to enter service by 2027.

Manufactured by Taiwanese foundry TSMC, these chips will integrate more HBM memory to support AI-related workloads. Meta plans to launch a new generation roughly every six months, a rapid pace for the industry. The strategy aims to improve the cost-performance ratio of its infrastructure while diversifying its semiconductor supply, as the group invests heavily in new data centers and continues, in parallel, to buy Nvidia GPUs and AMD processors.