Social media giant Meta is poised to begin production of its latest generation of AI-specific chips in September. The move, first reported by Reuters citing an internal memo, signals Meta's deepening commitment to developing its own hardware infrastructure to power its vast artificial intelligence operations. By creating its own chips, the company aims to curb the significant costs associated with purchasing Graphics Processing Units (GPUs) from external suppliers such as Nvidia and AMD, a trend increasingly adopted by major tech firms.
These new chips are part of Meta's Meta Training and Inference Accelerator (MTIA) programme, which has been in development since 2023. The company detailed four new chips in March, with some already in deployment or scheduled for rollout this year and next. Meta is taking a modular design approach, anticipating that the rapidly evolving nature of AI will necessitate adaptable hardware. While Meta will continue to procure GPUs from traditional chipmakers, the MTIA chips are intended to handle specific tasks, including training models for ranking and recommendation algorithms, broader AI workloads, and inference for its various applications.
The manufacturing process will see Taiwan Semiconductor Manufacturing Company (TSMC) producing the chips, with design collaboration from Broadcom. Additionally, Meta is sourcing essential components from various suppliers, including RAM from Samsung, storage from SanDisk, and fibre-optic equipment from Sumitomo Electric. This complex supply chain underscores the intricate nature of modern chip production and the global dependencies within the technology sector.
Meta's investment in proprietary AI hardware is part of a much larger financial outlay into artificial intelligence. The company announced in April that it anticipates capital expenditures between $125 billion and $145 billion this year, with a substantial portion dedicated to its AI initiatives. This includes striking data centre and power deals globally, alongside securing computing capacity to train and deploy its new Muse Spark series of AI models. The company reportedly plans to deploy 7 gigawatts of compute capacity this year, with intentions to double that figure next year.
Meta is not alone in this strategic shift. The escalating demand for AI compute power and the associated costs have prompted several tech giants to develop their own silicon. OpenAI recently unveiled an inference processor developed with Broadcom, while Anthropic is reportedly exploring chip development with Samsung. Amazon and Google have long been producing their own chips for AI training and inference, highlighting a broader industry trend towards vertical integration in the face of unprecedented AI growth and competition.