Sorry, but DeepSeek didn’t really train its flagship model for $294,000

Recent reports suggesting DeepSeek's R1 AI model was trained for a mere $294,000 are inaccurate, as this figure only accounts for the reinforcement learning phase. The foundational DeepSeek V3 model required an additional 2.79 million GPU hours, bringing the estimated total training cost closer to $5.87 million. This correction indicates that the capital expenditure for developing advanced AI models by Chinese firms is comparable to Western counterparts like Meta's Llama 4, challenging previous perceptions of significantly lower development costs and providing a more realistic view of investment requirements in the AI sector.

Analysis

The widely circulated claim that Chinese AI firm DeepSeek trained its R1 model for only $294,000 is a significant misinterpretation of the facts. This figure exclusively covers the post-training reinforcement learning (RL) phase, which constitutes a small fraction of the total effort. The foundational pre-training of the prerequisite DeepSeek V3 model actually required 2.79 million GPU hours on 2,048 Nvidia H800 GPUs, at an estimated cost of $5.58 million. Consequently, the combined cost to develop the R1 model is closer to $5.87 million, approximately 20 times the misleading headline figure. This level of expenditure places DeepSeek's resource commitment in a range comparable to Western counterparts, such as Meta's Llama 4, which required between 2.38 million and 5 million GPU hours. While DeepSeek V3 is a larger model than Llama 4 Maverick, it was trained on significantly fewer tokens (14.8 trillion vs. 22-40 trillion), indicating different strategic trade-offs between model size, compute hours, and data volume. Furthermore, the cost estimates are based on conservative GPU rental rates of $2/hour and do not account for the substantially higher capital cost of purchasing the hardware (estimated north of $51 million), nor do they include R&D, data acquisition, and other operational expenses. The analysis dismantles the narrative of hyper-efficient, low-cost Chinese AI development, revealing that building state-of-the-art foundation models remains a highly capital-intensive endeavor globally, subject to hardware supply chains dominated by firms like Nvidia and the associated geopolitical pressures.

Key Decisions for Investors

• Investors should disregard claims of dramatically lower AI training costs from Chinese firms and maintain the thesis that building competitive foundation models requires massive, ongoing capital expenditure, reinforcing the market position of well-capitalized tech leaders and key hardware suppliers.
• The analysis confirms both the critical demand for Nvidia's (NVDA) advanced GPUs in China and the significant geopolitical headwinds, including local competition and regulatory risks; therefore, investors should closely monitor NVDA's exposure and revenue stability in the Chinese market.
• The direct comparison with Meta Platforms (META) validates its AI development spending as being in line with global competitive standards, providing justification for the firm's substantial R&D budget allocated to its Llama model series.
• This incident highlights the need for rigorous due diligence on AI-related corporate disclosures, particularly by distinguishing between partial (e.g., reinforcement learning) and complete (e.g., pre-training) cost reporting to accurately assess a company's true capital efficiency and investment scale.