Context: The $5.6M figure is DeepSeek’s own and, per CSIS, covers only the final successful pretraining run — excluding prior experiments, post-training, staff, and the underlying hardware. SemiAnalysis estimated DeepSeek/High-Flyer operates ~50,000 Hopper-class GPUs and spent ~$1.6 billion on GPU-server capex. So the marginal-run figure may be accurate while the ‘trained for $5.6M’ framing understates true cost by orders of magnitude.