Meta takes a crucial step in reducing dependence on external suppliers and gradually moving towards designing customized chips
According to a Reuters report on March 14, Meta, the global social media giant, is collaborating with TSMC to test its first self-developed AI training chip. If the test is successful, Meta will rapidly scale up production.
Previously, Meta had successfully completed its first "tape-out", which is an important step in delivering the initial chip design to the manufacturing partner for feasibility assessment and is a costly process that usually takes three to six months.
After Meta abandoned its first MTIA chip in 2022, it had to increase its procurement of Nvidia GPUs. Since then, it has ordered billions of USD worth of GPUs for training, inference, recommendation, and advertising services. If this test is successful, Meta hopes to use its self-developed chips to train the next-generation Llama large language model, thereby reducing GPU purchases.
This test is an important part of Meta's investment in AI. In 2025, Meta is expected to invest between USD 114 billion (CNY 825 billion) and USD 119 billion, of which up to USD 65 billion will be used for AI infrastructure construction.