Meta and Broadcom Co-Developing Four Generations of MTIA Custom AI Inference Chips

Tom's Hardware·low signal

Meta revealed four new MTIA chip generations (300, 400, 450, 500) being co-developed with Broadcom on a six-month cadence for AI inference workloads. MTIA 300 is already in production for ranking/recommendations training, with 400 in lab testing. Hundreds of thousands of MTIA chips are deployed across Meta's apps. The parallel custom silicon strategy complements — not replaces — Meta's multi-billion Nvidia partnership for frontier training on Blackwell and Vera Rubin architectures.