Meta’s custom silicon poses challenge to Nvidia’s AI dominance

Meta isn’t just buying Nvidia’s chips anymore. It’s building its own. The company unveiled four new generations of its Meta Training and Inference Accelerator (MTIA) chip family on March 11, 2026: the MTIA 300, 400, 450, and 500. Each is designed for specific AI workloads that Nvidia’s general-purpose GPUs handle today, but at lower cost and with tighter integration into Meta’s sprawling infrastructure. The silicon strategy takes shape The MTIA lineup uses a modular chiplet design that lets Meta iterate roughly every six months, far faster than the traditional GPU development cycle. The earlier MTIA generations, the 100 and 200 series, have already been deployed in production. Hundreds of thousands of those chips are running inside Meta’s data centers, handling ranking and recommendation workloads that power everything from your Instagram feed to Facebook’s ad-targeting engine. Some have even been tested with Meta’s Llama large language models. The newer MTIA 450 and 500 variants will push further into generative AI inference, featuring improved high-bandwidth memory (HBM) to handle the data throughput that transformer-based models demand. An internal memo surfaced in July 2026 rev...