March 13 - According to a Reuters report, Meta (META, Financial) is quietly testing its first in‑house AI training chip in a bid to slash infrastructure costs and challenge Nvidia's (NVDA, Financial) dominance in the AI hardware market. The chip, part of Meta's Training and Inference Accelerator (MTIA) series and manufactured by Taiwan's TSMC, recently completed a crucial “tape‑out” phase, a key early milestone in chip development.
Meta is looking to spend up to $119 billion on AI in 2025, with about $65 billion allocated to AI infrastructure. As it has spent billions on Nvidia GPUs, Meta is now reducing its reliance on external suppliers and trying to optimize its own systems for jobs such as recommendation algorithms and generative AI.
Chief Product Officer Chris Cox described the project as a “walk, crawl, run situation,” acknowledging past setbacks while highlighting the chip's potential. If successful, Meta plans to fully integrate the chip into its AI training platforms by 2026, potentially reshaping cost structures and altering the competitive landscape in favor of greater AI independence. However, any misstep could result in significant delays and cost overruns.
