Google & Meta Team Up on AI Chips
In a direct challenge to Nvidia's market dominance, Google has signed a multibillion-dollar deal to supply Meta with its next-generation AI accelerator chips. The pact will see Meta deploy Google's custom silicon in its data centers, signaling a major industry push to diversify the AI hardware supply chain.
This multi-year, multi-billion dollar agreement will see Meta leasing Google's Tensor Processing Units (TPUs) to train and run its AI models. The deal is a significant move for Google as it aims to capture a larger share of the AI chip market and position its TPUs as a viable alternative to Nvidia's dominant GPUs. There are also discussions for Meta to purchase TPUs outright for its own data centers as early as 2027. Nvidia currently holds a commanding lead in the AI accelerator market, with an estimated 80-85% market share. This dominance is largely attributed to its powerful GPUs and its CUDA software, which has become an industry standard for AI development. The AI chip market itself is projected to grow from $20 billion in 2020 to over $300 billion by 2030. This partnership is part of a broader strategy by Meta to diversify its AI hardware suppliers. The company has also recently signed massive deals with Nvidia for its next-generation GPUs and with AMD for up to $60 billion in AI chips. Meta has committed to capital expenditures ranging from $115 billion to $135 billion for 2026, a significant increase from the $72 billion spent in 2025, primarily for AI infrastructure. Google's latest AI accelerator is the seventh-generation TPU, codenamed "Ironwood," which is designed specifically for inference—the process of generating answers from a trained AI model. Google claims Ironwood offers more than four times better performance per chip compared to the previous generation. These TPUs power Google's own major services like Search and Gemini. Meta has also been developing its own custom silicon, the Meta Training and Inference Accelerator (MTIA). However, the company has faced challenges, scrapping a more advanced version of its training chip due to design roadblocks. While its first-generation inference chip is in use, the difficulty in developing a competitive training chip underscores the complexity of the field and the rationale behind partnering with established players like Google.
