Google has entered into a multi-billion dollar agreement to lease its AI chips to Meta, directly challenging NVIDIA’s dominance in the AI training semiconductor market. While Meta has traditionally used NVIDIA’s GPUs for AI model development, it plans to utilize Google’s Tensor Processing Units (TPUs) for developing new AI models under this contract.
On February 26 (local time), the US IT specialist media outlet The Information reported that Meta has signed a multi-year TPU leasing deal worth billions of dollars with Google. Meta is also reportedly considering directly purchasing and implementing TPUs in its data centers from next year.
This agreement is significant for Google. While Google has provided TPUs to its cloud customers, this is the first instance of supplying them for large-scale technological firms’ AI training. Google’s potential to secure additional revenue via TPU sales could allow it to encroach on up to 10% of NVIDIA’s annual sales, which reached approximately $200 billion in the past 12 months.
Notably, Meta plans to use TPUs for AI model training rather than inference, which is remarkable given the prevailing belief that it is challenging to replace NVIDIA in this high-difficulty area. The Information also noted that Google is considering establishing a joint venture with private equity funds to promote TPU proliferation. This joint venture would purchase TPUs and lease them to other AI companies, using an SPV structure for funding, similar to NVIDIA’s strategy in expanding its ecosystem by investing in so-called “neo-cloud” businesses that lease GPUs.
Google, however, must carefully balance its strategy, as Google Cloud is a major customer of NVIDIA GPUs. Furthermore, as both TPUs and GPUs are produced by TSMC, the two sides will inevitably compete over production capacity.
Additionally, The Information reported that Meta has halted the development of its next-generation AI training semiconductor, “Olympus.” Previously, Meta had also paused some versions of its second-generation training chip “Iris,” indicating a series of revisions to its semiconductor roadmap. Meta had aimed to complete the Olympus design by the fourth quarter of this year, but the production process would have required at least nine additional months. Meta had planned to use technology from the startup Livos, acquired last year, for the core design of the training GPU, but complex architecture and software stability issues posed challenges. Consequently, Meta will continue to rely on external chips for AI model training for the time being.
[Reported by Ho-seop Won in Silicon Valley]