The MTIA Revolution: From 300 to 500

Meta has officially unveiled a massive acceleration in its custom silicon strategy, announcing four new generations of its Meta Training and Inference Accelerator (MTIA) chips. This "sprint" aims to deploy a new chip roughly every six months through 2027, a pace unheard of in the traditional semiconductor industry.

While the MTIA 300 is already in production for ranking and recommendation tasks, the upcoming 400, 450, and 500 series are laser-focused on the high-demand world of generative AI inference.

Breaking Down the Roadmap

• MTIA 300: Already deployed, this generation handles the heavy lifting for Facebook and Instagram’s content ranking and ad systems.


• MTIA 400: Currently in lab testing, this chip marks Meta's first major move into generative AI, offering performance comparable to leading commercial GPUs.


• MTIA 450 & 500: Scheduled for mass deployment in 2027, these "behemoth" chips prioritize high-bandwidth memory (HBM). The MTIA 450 doubles the HBM bandwidth of its predecessor, while the 500 adds another 50% on top of that to eliminate inference bottlenecks.

Why Now? The Strategic Shift

By designing its own hardware in partnership with Broadcom and TSMC, Meta aims to optimize its data center efficiency and slash the astronomical costs associated with third-party GPUs like Nvidia’s.

These chips use a modular chiplet architecture, allowing Meta to iterate rapidly and drop new generations into existing server racks without costly infrastructure overhauls.

While Meta will continue to purchase billions in external chips, the MTIA series represents a long-term play for total control over its AI future.