In March 2026, Meta announced four new generations of custom AI chips, designed to reduce its reliance on Nvidia. What is the name of these chips?

In March 2026, Meta unveiled its latest advancements in custom artificial intelligence hardware: four new generations of chips known collectively as MTIA, which stands for Meta Training and Inference Accelerator. This family of chips is central to Meta's strategy to enhance its AI infrastructure and reduce its dependency on external suppliers like Nvidia and AMD for the specialized processors critical to powering its vast array of AI-driven services.

The development of these in-house chips allows Meta to tailor hardware specifically for its unique AI workloads, primarily focusing on inference tasks. Inference involves deploying trained AI models to generate recommendations, provide answers, or power generative AI features within applications like Facebook and Instagram. While general-purpose GPUs from companies like Nvidia are powerful, they can be over-specified for these specific inference needs, making custom chips a more cost-effective and efficient solution for Meta's massive scale.

Meta's ambitious roadmap includes deploying these new MTIA generations at a rapid pace, with a new chip expected roughly every six months through 2027. The MTIA 300 is already in production, handling tasks like ranking and recommendations, while subsequent versions such as the MTIA 400, 450, and 500 are designed to tackle more demanding generative AI inference workloads. This iterative approach and focus on custom silicon aim to optimize performance and cost for Meta's expansive AI operations, marking a significant step in the company's long-term technological independence.