Meta Scraps In-House AI Chip, Turns to Partners
Meta has reportedly scrapped its advanced in-house AI chip project, opting instead to deepen its reliance on external partners like Nvidia, AMD, and now Google. The move signals a strategic pivot toward speed and scale through partnerships rather than pursuing vertical integration for its AI hardware.
This move follows the termination of two of Meta's most advanced internal training chip projects, internally codenamed "Iris" and "Olympus". The "Olympus" chip was ultimately deemed too risky due to insufficient software stability and a design that was too complex for reliable mass production, especially amid intense competition from OpenAI and Google. The pivot to external partners includes a multi-year, multi-billion dollar deal to lease Google's Tensor Processing Units (TPUs), marking the first time Google's TPUs will be used in a hyperscale data center outside of its own cloud services. Meta is also reportedly negotiating to purchase TPUs directly for its own data centers starting as early as 2027. This new Google partnership complements massive, long-term deals with Meta's existing suppliers. The company has a multi-year agreement with Nvidia for millions of its Blackwell and Rubin GPUs and a separate five-year partnership with AMD, valued at up to $100 billion, to deploy 6 gigawatts of its Instinct GPUs. The shift away from in-house development comes amidst a colossal increase in infrastructure spending. Meta has forecasted capital expenditures between $115 billion and $135 billion for 2026, a significant jump from the $72 billion spent in 2025. CEO Mark Zuckerberg has indicated plans to invest at least $600 billion in U.S. data centers and related infrastructure by 2028. Meta's custom silicon effort, the Meta Training and Inference Accelerator (MTIA) program, has faced challenges before. Its first-generation inference chip, MTIA v1, was designed in 2020 and successfully deployed for recommendation workloads but showed efficiency gains only in low-to-medium complexity tasks, lagging behind GPUs in more complex applications. The company had also scrapped an even earlier inference chip after it underperformed in testing.
