ChatGPT Creator Spent IDR 1.65 Trillion, China Imitates IDR 4.8 Billion in Capital
Jakarta, CNBC Indonesia- Chinese artificial intelligence (AI) developer DeepSeek revealed that it only spent around US$294,000, or the equivalent of Rp4.8 billion, to train its latest model, R1.
This figure is much lower than the training costs of AI models from US competitors, which are said to reach hundreds of millions of dollars.
This statement appeared in a peer-reviewed scientific article published in the academic journal Nature on Wednesday (September 17, 2025)
This marks the first time DeepSeek has detailed the estimated training costs of R1, a model focused on reasoning.
According to the report, R1 training used 512 Nvidia H800 chips and took a total of 80 hours.
Previously, OpenAI CEO Sam Altman said that training fundamental AI models cost more than US$100 million, although the company never specified the figure.
DeepSeek shocked global markets last January when it released its low-cost AI system, raising concerns among investors about the threat to the dominance of tech giants like Nvidia. This situation even triggered a sell-off in tech stocks.
However, after its release, the company and its founder, Liang Wenfeng, rarely appeared in public except to launch a few product updates.
The Nature article also revealed that DeepSeek used the A100 chip in its early stages of development before eventually switching to the H800.
This is noteworthy because the US has banned exports of advanced H100 and A100 AI chips to China since 2022.
However, Nvidia emphasized that DeepSeek only uses legally acquired H800 chips.
In a supplementary information document accompanying the Nature article, the company acknowledged for the first time that it owns A100 chips and used them in the preparatory phase of development.
“Regarding our research on DeepSeek-R1, we leveraged the A100 GPU to prepare for experiments with smaller models,” the researchers wrote, as quoted by Reuters on Monday (September 22, 2025).
After that initial phase, R1 was trained for 80 hours using a cluster of 512 H800 chips.
SOURCE : CNBC INDONESIA
