Data on AI Chip Sales (epoch.ai)

A new open database has been launched that compiles estimates of AI chip sales, tracking the number of dedicated AI accelerators shipped by major manufacturers over time. This comprehensive dataset leverages financial reports, company disclosures, and analyst estimates to provide insights into the compute capacity, power usage, and spending associated with various AI chips, including those from industry leaders like Nvidia, Google, and Amazon. The database is significant as it offers a detailed look into the growing AI hardware market, which is critical for understanding trends in AI capabilities and the infrastructure supporting machine learning advancements. The dataset categorizes sales by chip manufacturer and model, using Nvidia's reported revenues as a benchmark for accuracy, particularly for its Hopper and Blackwell GPUs. It employs an innovative metric, H100e (H100-equivalent), to convey compute capacity more intuitively by equating each chip's operations to Nvidia's widely utilized H100 GPU. While the dataset faces challenges regarding sales transparency and the complexities of chip performance metrics, it serves as a vital tool for researchers, developers, and stakeholders aiming to navigate the rapidly evolving AI landscape.