According to Reuters, Meta is testing an in-house chip specifically designed for AI training. This is a crucial part of the company's strategy to reduce its reliance on hardware manufacturers like Nvidia.
These chips are produced in collaboration with TSMC, the Taiwanese semiconductor giant, and are specifically designed to handle AI-specific workloads. Currently, Meta is conducting "small-scale deployment" tests, and plans to scale up production if the results meet expectations.
It's worth noting that Meta has previously deployed custom AI chips, but those were only used for running models, not training them. The report indicates that some past chip design projects at Meta have been canceled or scaled back due to failure to meet internal expectations.
This move has significant financial implications for Meta. The company projects capital expenditures of $65 billion in 2025, a large portion of which is allocated to purchasing Nvidia GPUs. Reducing even a small fraction of this expenditure through self-developed chips would be a significant achievement for the social media giant.
