Meta will deploy four new generations of custom AI chips by the end of 2027, accelerating its in-house silicon development to a six-month release cycle that challenges industry norms. The social media giant announced the MTIA 300, 400, 450, and 500 processors this week as part of a push to diversify hardware sources and reduce reliance on external chipmakers like Nvidia and AMD.
While typical AI chip development cycles span one to two years, Meta has developed capacity to release new silicon every six months or less using modular, reusable designs.
MTIA 300 already handles content ranking and recommendation training for Facebook and Instagram users. The subsequent models will primarily support generative AI inference production through 2027, with MTIA 450 arriving early next year and MTIA 500 following six months later.
Meta partnered with Broadcom to develop the semiconductors built on open-source RISC-V architecture, with Taiwan Semiconductor Manufacturing Corporation handling fabrication. The company takes an inference-first approach with MTIA 450 and 500 optimized for generative AI tasks before handling other workloads like ranking and training.
This aggressive roadmap follows reports earlier this year that Meta scaled back some in-house efforts to make high-end chips competing directly with Nvidia. The company now appears eager to dispel that narrative by announcing this accelerated development schedule.
Despite the custom silicon push, Meta continues spending billions on AI hardware from industry leaders. Recent agreements with Nvidia and AMD are worth tens of billions of dollars each, securing gigawatts of AI capacity over coming years.
Meta expects capital spending between $115 billion and $135 billion this year as it expands its global data center footprint. The company's portfolio approach combines internal chip development with purchases from established suppliers to scale infrastructure capacity.
