Meta is preparing to launch the production of its next-generation AI chips in September, as revealed in an internal memo cited by Reuters. The company is collaborating with Broadcom for design and TSMC for manufacturing. To complete the hardware stack, Meta is sourcing RAM from Samsung, storage from Sandisk, and fiber-optic components from Sumitomo Electric.

These chips are part of the Meta Training and Inference Accelerator (MTIA) program. To keep pace with the rapid evolution of artificial intelligence, Meta is utilizing a modular chiplet architecture, allowing each generation to incorporate new hardware technologies and workload insights on a shorter release cycle.

The MTIA chips will handle inference for Meta's applications, as well as training for recommendation and ranking algorithms. While the company will continue to purchase GPUs from Nvidia and AMD, these custom chips are intended to reduce reliance on external GPU suppliers and lower overall costs.

This hardware push coincides with massive infrastructure investment. Meta expects capital expenditures to reach between $125 billion and $145 billion this year. This funding supports the deployment of the Muse Spark model series, with plans to implement 7 gigawatts of compute capacity this year and double that amount in 2027.