Meta is preparing to begin production of its latest AI-focused chips in September, as it seeks to reduce reliance on costly GPUs amidst a global component shortage. According to an internal memo reported by Reuters, one of these chips has already cleared testing in roughly six weeks. Meta is collaborating with Broadcom on the chip designs and will manufacture them at Taiwan Semiconductor Manufacturing Company (TSMC). Other suppliers include Samsung for RAM, Sandisk for storage, and Sumitomo Electric for fiber-optic equipment.

The new chips are part of Meta’s Meta Training and Inference Accelerator (MTIA) program, unveiled earlier this year. This initiative takes a modular design approach, enabling Meta to adapt chip capabilities in response to the rapidly evolving AI landscape. Some versions of these MTIA chips are either already in use or scheduled for deployment by the end of this year or early next year. Meta aims to use these chips extensively for training AI models involved in ranking, recommendation systems, and broad AI tasks across its platforms.

Meta’s chip development reflects its broader strategy to mitigate costs and dependency on large chip providers like Nvidia and AMD, although it plans to continue substantial purchases from them as well. The company’s capital expenditure for 2026 is projected between $125 billion and $145 billion, much of which targets expanding AI infrastructure. Meta expects to deploy 7 gigawatts of computing power this year and aims to double that output in 2027, supporting AI models like Muse Spark.

This move aligns Meta with other major players in building custom AI chips, such as OpenAI and Amazon, both of whom are developing proprietary solutions to handle massive AI workloads. Meta has also secured significant deals with ARM, AMD, and Amazon to supplement its compute needs. The modular MTIA chips mark an important step in Meta’s ambitious efforts to scale AI capabilities while controlling costs amid growing computational demand.