Meta Scraps In-House AI Chip After Setbacks
Meta has scrapped its advanced in-house AI training chip after facing significant performance and reliability roadblocks. The failure is a major setback for its 'Meta Compute' ambition, a gigawatt-scale blueprint for vertically integrated infrastructure, highlighting the immense difficulty of custom silicon development for hyperscale AI.
The now-canceled chip, codenamed "Olympus," was part of Meta's broader "Meta Training and Inference Accelerator" (MTIA) program. This initiative previously saw another internally developed training chip, "Iris," also get discarded. The primary goal of the MTIA program is to create custom silicon specifically for Meta's AI workloads, aiming to improve efficiency for tasks like content recommendation on its platforms. The decision to scrap Olympus stemmed from its complex design, which raised concerns about the ability to manufacture it in large quantities. Executives worried that the software for the chip would not be as stable as Nvidia's offerings, posing a significant risk to the company's ability to train new AI models and stay competitive with rivals like Google and OpenAI. This setback occurs within the context of Meta's massive AI infrastructure expansion, a project dubbed "Meta Compute." The company has committed to spending between $115 billion and $135 billion in 2026 alone on this initiative, a significant increase from the $72 billion spent in 2025. The goal is to build data center capacity measured in tens of gigawatts within the decade. In response to the internal chip development challenges, Meta has aggressively turned to external vendors to secure its AI future. The company has inked a multi-billion dollar deal with Google to lease its specialized AI chips. This move helps Meta diversify its suppliers and reduce its heavy reliance on a single provider. Furthermore, Meta has deepened its long-standing relationship with Nvidia, signing a multi-year agreement to purchase millions of its next-generation "Vera Rubin" GPUs and "Grace" CPUs. This partnership is a cornerstone of Meta's strategy to deploy the massive computing power required for its AI ambitions. To further bolster its hardware supply chain, Meta has also entered into a significant partnership with AMD. This deal, reportedly worth up to $100 billion, will see Meta deploy up to 6 gigawatts of AMD's Instinct GPUs to power its next-generation AI infrastructure.
