DeepSeek claimed its R1 AI model was trained for $6M using 2,048 GPUs, but SemiAnalysis reports the company has invested $1.6B in hardware, including 50,000 Nvidia Hopper GPUs.
Originally launched by hedge fund High-Flyer, DeepSeek operates its own data centers and hires exclusively from mainland China, offering top salaries. It prioritizes algorithmic efficiency over brute-force scaling, developing innovations like Multi-Head Latent Attention (MLA).
The reported $6M training cost covers only GPU pre-training, excluding broader infrastructure and R&D expenses. Total AI investment exceeds $500M, reinforcing that major AI breakthroughs still require billions in funding.
