Meta rolls out in-house AI chips weeks after massive Nvidia, AMD deals

On Wednesday, Meta announced the launch of four custom-designed AI chips, enhancing its ambitious data center expansion strategy. This new lineup, part of the Meta Training and Inference Accelerator (MTIA) series, marks a significant step forward for the company as it aims to optimize performance and reduce reliance on external vendors. According to Yee Jiun Song, Meta's Vice President of Engineering, the development of these in-house chips, manufactured by Taiwan Semiconductor, allows the company to achieve better price-performance ratios across its data centers. "This also provides us with more diversity in terms of silicon supply and helps shield us from price fluctuations to some extent," Song explained. "It gives us a bit more leverage." The MTIA 300 chip, which has already been deployed, is designed for training smaller AI models that support Meta's content ranking and recommendation systems, including targeted ads on platforms like Facebook and Instagram. Future chips in the series—MTIA 400, MTIA 450, and MTIA 500—are intended for advanced generative AI tasks, such as creating multimedia content based on text prompts. Notably, these chips will not be used for training large language models, as clarified by Song. Meta is progressing swiftly, with the MTIA 400 chip already undergoing testing and set to be deployed soon, while the latter two chips are expected to become operational by 2027. "It’s uncommon for a silicon company to release new chips every six months, but we find ourselves rapidly expanding our capacity and investing heavily in capital expenditures, necessitating the deployment of state-of-the-art chips at all times," Song added. He noted that these chips are anticipated to have a useful lifespan of over five years. With significant investments in new data centers, including a massive facility in Louisiana and others in Ohio and Indiana, Meta is also exploring leasing opportunities at the Stargate site in Texas. This follows the withdrawal of OpenAI and Oracle from their plans to expand the AI data center there, as reported by Bloomberg. Tech companies, including Google, have increasingly turned to in-house silicon to mitigate the challenges posed by the expensive and often scarce GPUs from Nvidia and AMD. These large tech firms have developed application-specific integrated circuits (ASICs), which are typically smaller and more cost-effective but limited in their functionality compared to general-purpose GPUs. Google pioneered this trend with its Tensor Processing Unit in 2015, followed by Amazon in 2018. While Meta's MTIA chips are exclusively for internal use, the new models will feature enhanced high-bandwidth memory (HBM) to support generative AI inference tasks. However, the broader industry's push for AI has led to a memory chip shortage, raising concerns about the supply chain for Meta's ambitious silicon plans. Despite these challenges, Song expressed confidence in securing the necessary supplies, though he refrained from commenting on potential long-term contracts with memory vendors. In addition to the new chips, Meta has recently secured agreements to integrate millions of Nvidia GPUs and up to 6 gigawatts of AMD GPUs over the coming years. "The workloads are evolving rapidly, and we want to ensure we have options," Song concluded, highlighting the importance of flexibility in their chip strategy. Manufacturing for Meta's chips is handled by Taiwan Semiconductor, which operates primarily in Taiwan but has expanded operations to a new fabrication campus in Arizona. While Meta did not disclose specifics about production locations, a large portion of its dedicated engineering team is based in the United States, where most of its 30 operational and planned data centers are situated.