Meta Pivots From Custom AI Chips
Meta has reportedly scrapped its most advanced in-house AI chip project, opting to deepen its reliance on partners like Nvidia and AMD. The move signals the immense difficulty of competing at the highest end of chip design, even for a tech giant, while its engineers focus on scaling video processing with open-source tools like FFmpeg.
This isn't Meta's first attempt at custom silicon to hit a snag. The company's "Meta Training and Inference Accelerator" (MTIA) program previously scrapped an earlier inference chip that underperformed in testing, leading to a pivot back to Nvidia GPUs in 2022. This latest move cancels the more advanced "Olympus" training chip and an earlier second-generation chip codenamed "Iris" due to significant design challenges. The pivot to external partners involves massive capital outlay. Meta's planned capital expenditures for 2026 are set between $115 billion and $135 billion, a significant increase from $72.2 billion the previous year, with a large portion dedicated to AI infrastructure. The company has stated a goal of spending $600 billion on AI infrastructure by 2028, including the construction of 26 U.S. data centers. The scale of the new Nvidia deal is immense, described as a "multiyear, multigenerational strategic partnership." Meta will be purchasing millions of Nvidia's next-generation Blackwell and Rubin GPUs, as well as its Grace and Vera CPUs, making it the first company to deploy the Grace CPUs as standalone chips at scale. To avoid over-reliance on a single supplier, Meta is also making a huge investment in AMD. The company signed a multi-year agreement worth up to $60 billion for AMD's upcoming MI450-based Instinct GPUs, which are specifically tailored for AI inference workloads. This "Nvidia plus one" strategy is designed to hedge against supply chain risks. The agreement with AMD includes performance-based stock warrants that could allow Meta to acquire up to a 10% stake in the chipmaker. This signals a deeper strategic alignment, moving beyond a simple buyer-seller relationship to intertwine the long-term hardware and software roadmaps of both companies. Developing competitive, cost-effective custom chips in-house presents formidable challenges, even for tech giants with deep pockets. Beyond the sheer capital investment, it requires a long-term accumulation of specialized engineering talent, intellectual property, and the ability to navigate complex design and debugging processes to ensure power consumption and performance can justify the effort over buying from established players like Nvidia.
