Meta plans to deploy two new generations of custom AI chips in 2027 as it tries to reduce the cost of running models across its services. Its roadmap, published March 11, schedules MTIA 450 for mass deployment in early 2027 and MTIA 500 during the same year.
Both prioritize inference, the work a trained model does when it responds to a request. Meta says designing around that workload can be more economical than using chips built primarily for large-scale model training. The goal is to make repeated AI use cheaper at the company's scale; the roadmap does not establish a guaranteed saving for each request.
The custom hardware also depends on outside partners. Meta's expanded agreement with Broadcom covers chip design, advanced packaging and networking. In its April 14 announcement, Broadcom said the initial commitment exceeds 1 gigawatt, with plans to support MTIA generations through 2029. That figure describes a planned rollout, not capacity already operating.
Meta also uses Taiwan Semiconductor Manufacturing Co. to fabricate its processors, Reuters reported in March. The same report noted that Meta had signed agreements in February to buy tens of billions of dollars worth of chips from Nvidia and AMD. Meta describes its approach as matching different accelerators to different workloads, leaving a role for commercial chips alongside its own designs.
Making specialized hardware introduces a timing problem. Meta says chip development can take about two years, during which the AI workloads designers anticipated may change substantially. Its response is to use modular components and shorter development cycles. The MTIA 400, 450 and 500 share chassis, racks and networking infrastructure, according to the technical roadmap, so a new chip generation can fit into the same physical systems.
