Spreadsheets explains how FLOPs estimate DeepSeek training cost

The author explains the fundamental concepts of floating-point operations, or FLOPs, and how they drive artificial intelligence model training compared to traditional software. Using GPT-2 as an example, the article demonstrates how to count FLOPs in a browser and analyzes DeepSeek's reported $5.3 million training cost. The author argues that comparing this single final training run directly to total development costs is misleading because it excludes research, experimentation, and failed attempts. Open article →