Meta & Google Team Up on AI Chips

This multi-year, multi-billion-dollar agreement will see Meta lease Google's custom-designed Tensor Processing Units (TPUs) to train and run its next-generation AI models. The deal provides Meta with long-term, stable access to a significant source of computing power, diversifying its hardware supply chain. There are also discussions for Meta to potentially purchase TPUs directly for its own data centers as early as 2027. Google's TPUs, first developed in the early 2010s and used internally since 2015, are specialized chips (ASICs) optimized for the mathematical operations common in machine learning. Initially designed for Google's TensorFlow framework, later generations of TPUs have become powerful enough for both training and inference in large-scale AI models. Google only began making its TPUs available to third-party companies in 2018. The move is part of Meta's broader strategy to diversify its hardware suppliers and mitigate reliance on Nvidia, which currently holds a dominant market share of over 90% in AI data center chips. This partnership follows Meta's recent multi-billion dollar deals with AMD for its Instinct GPUs and a multi-generational partnership with Nvidia for its chips. Meta's decision to partner with Google also comes after the company faced challenges and ultimately scrapped its most advanced in-house AI training chip project, codenamed "Olympus." While Meta continues to develop its own "Meta Training and Inference Accelerator" (MTIA) program for less complex tasks, the deal with Google provides access to high-performance chips for its most demanding AI workloads. For Google, this partnership represents a significant step in commercializing its custom hardware and challenging Nvidia's market dominance. The deal is expected to bolster Google Cloud's revenue and validates its strategy of offering its specialized AI infrastructure to external companies, even its direct competitors. This collaboration is indicative of the immense and growing demand for AI computing power, which has become a critical bottleneck in the development of advanced AI models. The high cost and tight supply of Nvidia's GPUs have prompted major tech companies to seek alternative solutions and build more resilient supply chains. The total investment in AI infrastructure is skyrocketing, with some projections suggesting the AI accelerator market could become a trillion-dollar industry within the next five years. Meta itself has committed to spending hundreds of billions on AI infrastructure by 2028, including the construction of massive new data centers.