DeepSeek claims it trained R1 AI model for just $294,000
If the Chinese AI company numbers are accurate, it indicates that high-level AI can be developed at a fraction of the previously assumed cost.
Chinese artificial intelligence company DeepSeek has claimed in a paper published in the academic journal Nature that it trained its latest reasoning model, R1, for only about $294,000.
Experts have long believed that creating models of this size and complexity requires hundreds of millions of dollars in computing power. If DeepSeek’s numbers are accurate, it suggests that high-level AI can be developed at a fraction of the previously assumed cost.
The Hangzhou-based firm explained that the final training run for R1 used 512 NVIDIA H800 chips. It also acknowledged that earlier development involved A100 chips, which are subject to US export restrictions. This admission has drawn attention, as chip access has become a sensitive issue in the ongoing technology rivalry between the United States and China.
NVIDIA developed the H800 as a customised chip for China following US export controls introduced in October 2022, which blocked sales of the more advanced H100 and A100 GPUs. The H800 is essentially a pared-down version of the H100, with some performance features limited so that it fits within the restrictions.
The details in the paper are generating debate not just because of the cost but also because of what it might mean for the future of AI. Cheaper training of advanced models could lower barriers for startups and reshape competition in the sector. It could also widen access to powerful AI tools that were once thought to be the preserve of only the largest technology firms.
However, many in the industry remain sceptical. Some industry experts have noted that the $294,000 figure only reflects the cost of the graphics processors used for the final training run. It does not take into account the price of data, energy, infrastructure, or the large teams of researchers and engineers needed to carry out such work. Independent experts have not yet verified DeepSeek’s claim, and some argue the true cost is likely to be far higher.
This is not the first time DeepSeek has sent shock waves through the market. In January, news of its progress caused a dramatic sell-off in technology stocks, wiping hundreds of billions of dollars from the value of major chipmakers such as NVIDIA. The new paper is likely to reignite debate about how quickly AI is advancing and whether cost estimates can be taken at face value.
Edited by Swetha Kannan