Google and Meta Form AI Chip Alliance

This partnership will see Meta renting Google's custom-designed Tensor Processing Units (TPUs) through its cloud division, rather than purchasing the hardware directly. The multi-year agreement is valued at several billion dollars and is intended to provide the massive computing power Meta needs to train and deploy its advanced AI models. The deal leverages over a decade of Google's work on custom chips. The TPU project began around 2013 when Google realized that the growing computational demands of its own AI services, like voice search, could require it to double its number of data centers. The first-generation TPU was deployed internally in 2015 to accelerate AI inference in products like Google Search and Maps. Meta has also been developing its own custom silicon, known as the Meta Training and Inference Accelerator (MTIA). The first version of this chip was announced in 2023, primarily designed to handle AI inference workloads for its recommendation models, which are crucial for services like Facebook feeds and ads. A next-generation MTIA is in development to further improve performance and efficiency for Meta's specific AI tasks. This alliance is a direct response to Nvidia's overwhelming control of the AI hardware market, where it holds an estimated 92-95% market share in AI data center chips. The high demand for Nvidia's GPUs has led to supply constraints and high prices, pushing major tech companies to diversify their hardware sources and seek more cost-effective, specialized solutions. The move is part of a larger industry trend where major technology firms, including Amazon and Microsoft, are investing heavily in creating their own custom chips. By designing silicon specifically for their software and workloads, companies aim to achieve greater performance, better energy efficiency, and more control over their technology stack, reducing their dependency on third-party suppliers.