The true cost of training DeepSeek’s R1 AI model has finally been revealed, and the numbers shocked the industry. In a report published in Nature, the Chinese company disclosed that the model cost just US$294,000 (≈ R$1.6 million) to train—powered by 512 NVIDIA H800 GPUs.
For comparison, OpenAI’s GPT-4 cost an estimated US$80–100 million, according to CEO Sam Altman. Taking the midpoint of US$90 million, DeepSeek’s R1 comes out to be over 300 times cheaper.
Why So Much Cheaper?
DeepSeek attributes its dramatic cost savings to a trial-and-error learning method, rather than relying heavily on human-labeled examples and demonstrations.
Most modern AI models learn through massive datasets curated by humans, making the process resource-intensive and hard to scale as tasks grow more complex.
Instead, DeepSeek R1 was trained by reward-based reinforcement learning, similar to how a child learns while playing a video game.
“As the child explores with their avatar, they learn by trial and error that some actions, such as collecting coins, yield points, while others, such as being hit by enemies, cause them to lose,” explained Yiming Zhang, a doctoral student at Carnegie Mellon University.
The model earned points for correct answers and lost them for wrong ones—allowing it to develop reasoning skills organically, without depending on endless pre-made tutorials. This approach enabled R1 to reach GPT-4-like reasoning abilities at a fraction of the cost.
The “DeepSeek Effect”
When DeepSeek R1 launched in January 2025, it made global headlines. The model is open-source and free, yet powerful enough to rival closed systems from OpenAI and Google.
The announcement shook markets. NVIDIA’s stock dropped nearly 18% in a single day, wiping out US$600 billion in market value—the largest single-company loss in U.S. stock market history.
Beyond finance, the launch had geopolitical consequences. Governments in the U.S., Canada, and Italy restricted the use of DeepSeek on official devices, citing privacy and security concerns.
