According to a Reuters report citing an internal memo, Meta plans to start producing the latest generation of its AI-focused chips in September. The move aims to reduce GPU costs amid an unprecedented shortage of components.
What and when
The memo states that at least one of the chips cleared testing in about six weeks. Manufacturing is slated to begin in September; chip design involves Broadcom, while production will be handled by Taiwan Semiconductor Manufacturing Company (TSMC).
Meta detailed four new chips under its Meta Training and Inference Accelerator (MTIA) program in March. Some of those chips are already in deployment, and others are expected to be rolled out this year or next. The company is using a modular, chiplet-based design to allow flexibility as AI workloads and requirements evolve quickly.
Suppliers and components
Reuters notes that Meta is sourcing RAM from Samsung, storage from Sandisk, and fiber-optic equipment from Sumitomo Electric. Broadcom is collaborating on chip design and TSMC will manufacture the chips.
Intended use and significance
Meta intends to use the MTIA chips for training models used in ranking and recommendation systems, for broader AI workloads, and for inference in its applications. The internal development is expected to lower the company’s need to buy GPUs from external suppliers like Nvidia and AMD, although Reuters says Meta still expects to spend substantially with those vendors.
Large-scale investment and compute capacity
Meta has been investing heavily to secure compute capacity for its AI efforts. In April the company said it expects capital expenditures of $125 billion to $145 billion this year, much of which is directed toward AI. The memo Reuters cited indicates Meta plans to deploy 7 gigawatts of compute this year and to double that capacity next year.
Partnerships and market context
Meta has also signed deals to secure compute for recommendation systems with ARM, a multibillion-dollar agreement with AMD for Instinct GPUs, and a multibillion-dollar deal with Amazon to use the cloud provider’s homegrown CPUs for AI needs. Reuters points out that other industry players are also developing or procuring custom chips: OpenAI unveiled an inference processor built with Broadcom, Anthropic is reportedly considering chip development with Samsung, and both Amazon and Google develop their own AI training and inference chips. Numerous startups are also active in the space to meet rising demand.
Comment
Meta declined to comment to Reuters.
