Meta Scraps AI Chip, Bets Big on Nvidia

Meta's now-scrapped advanced AI training chip, codenamed "Olympus," was part of the company's long-term "Meta Training and Inference Accelerator" (MTIA) program. This initiative aimed to create custom silicon to reduce reliance on third-party vendors and optimize for Meta's specific workloads, particularly deep learning recommendation models. However, the project hit significant design and production roadblocks, leading to internal skepticism about its ability to compete with Nvidia's performance. This isn't the first setback for Meta's custom chip ambitions. An earlier inference chip was also scrapped after underperforming in tests, prompting a pivot to Nvidia GPUs in 2022. The Olympus chip, which was to be based on a 2nm chiplet architecture, was deemed too complex for mass production with concerns about software stability, ultimately leading to its cancellation. The decision to abandon Olympus coincides with a massive increase in Meta's planned capital expenditures, set to reach $115-$135 billion in 2026, a significant jump from $72.2 billion in 2025. This spending spree is largely dedicated to AI infrastructure to support large-scale models like their upcoming "Avocado" and "Mango." To fill the gap left by its in-house chip, Meta is pursuing a multi-supplier strategy. The company has inked multi-billion dollar deals to acquire millions of next-generation GPUs from Nvidia, including the upcoming Vera Rubin line. It has also committed to a deal with AMD that could be worth up to $100 billion for their Instinct GPUs and is even leasing Google's Tensor Processing Units (TPUs) for the first time. This hardware diversification is critical for Meta's focus on recommendation systems, which power content discovery and ads across Facebook and Instagram. These systems, along with generative AI, have different architectural needs than the large language models developed by competitors. By leveraging a mix of Nvidia, AMD, and Google hardware, Meta aims to avoid supply chain bottlenecks and gain negotiating power. The move highlights the immense difficulty and expense of developing custom AI accelerators that can rival the established ecosystems of players like Nvidia. While Google has seen success with its TPUs, which are optimized for its TensorFlow framework, creating a competitive software and hardware stack from the ground up presents significant challenges. Meta's focus now shifts to leveraging external hardware to power its next generation of AI-driven products.