Meta claims its MTIA (Meta Training and Inference Accelerator) chips deliver a 44% reduction in total cost of ownership and a 40% improvement in power efficiency compared to general-purpose Nvidia GPUs for targeted workloads, including ranking and recommendation systems, ad optimization, and generative AI inference. Meta’s roadmap plans four chip generations (MTIA 300, 400, 450, and 500), with a new generation roughly every six months, compared to the typical one-to-two-year cycle in the semiconductor industry. The MTIA 300 is already in production, and the MTIA 400 (codenamed Iris) is scheduled for September 2026. Meta’s projected capital expenditures for 2026 sit between $115 billion and $135 billion, with a target of 14 gigawatts of computing capacity by 2027. These ASIC chips complement Nvidia and AMD GPUs used for model training, as inference represents the largest volume of compute once models are deployed at scale.
Source: Read the original article
