Meta plans to start mass production of its self-designed AI chip, codenamed 'Iris,' in September 2026, aiming to double its overall AI computing capacity to 14 GW by 2027. This move reduces Meta’s reliance on external suppliers like Nvidia and AMD, potentially reshaping the AI hardware market and lowering costs for large-scale AI training and inference. The chip, part of the MTIA (Meta Training and Inference Accelerator) fourth-generation project, was developed with Broadcom and manufactured by TSMC, passing tests in only six weeks without significant issues.
Background
Meta has been investing heavily in AI infrastructure, with plans to deploy 7 GW of computing power this year and $145 billion in AI infrastructure spending. The MTIA chip series is a custom ASIC designed to optimize Meta's own AI workloads, such as recommendation algorithms and content ranking, offering better efficiency than general-purpose GPUs. By developing its own chips, Meta aims to reduce dependency on dominant vendors like Nvidia and control costs.