Google and Meta Team Up on AI Chips

This partnership is part of a larger trend of tech giants, known as hyperscalers, developing their own custom silicon to reduce their dependence on third-party chipmakers. Companies like Amazon with its Trainium and Inferentia chips, and Microsoft with its Maia project, are all making significant investments in in-house chip design to optimize for their specific AI workloads and control costs. Google has been a pioneer in this space with its Tensor Processing Units (TPUs), first developed around 2013 to handle the growing computational demands of its own AI services like speech recognition. These custom-built chips, designed to accelerate machine learning tasks within Google's TensorFlow framework, have been used internally since 2015 and were made available to third-party cloud customers in 2018. The latest versions of Google's TPUs are liquid-cooled and can be connected into large clusters, or "pods," to train massive AI models. Meta has also been working on its own line of custom AI chips called Meta Training and Inference Accelerators (MTIA). The primary goal of the MTIA is to efficiently power the company's recommendation algorithms and other inference-based tasks across its suite of apps. While early versions of Meta's in-house chips struggled to compete with Nvidia's performance, the company continues to invest heavily in its own silicon as a long-term strategy. Under the new agreement, Meta will not only rent Google's TPUs but is also in talks to potentially purchase the chips outright for its own data centers in the future. This move to diversify its AI hardware supply chain comes as Meta also continues to purchase billions of dollars worth of GPUs from Nvidia and has a separate deal with AMD for their AI chips. The a multi-year, multi-billion dollar deal provides Meta with more options and leverage in a market where demand for high-performance AI chips far outstrips supply.