DeepSeek-R1 shocks AI world with $294K training
China’s technology company DeepSeek has sparked a global storm in the tech world after revealing the training cost of its new AI model, DeepSeek-R1. The company claims the model was trained at a cost of just $294,000 (around ₹2.45 crore), while American firms have long stated that such projects require hundreds of millions of dollars. This revelation has been published in a recent research report in the prestigious science journal Nature.
DeepSeek’s claim not only challenges the technological dominance of countries like the US but also sets off a fresh debate worldwide about the cost and accessibility of artificial intelligence models.
Training on Nvidia H800 Chips, Completed in Just 80 Hours
According to DeepSeek’s report, the R1 model was trained using 512 Nvidia H800 chips. Surprisingly, the entire training process was completed in only 80 hours. These H800 chips were specially designed by Nvidia for the Chinese market after the US banned the export of its advanced H100 and A100 chips in October 2022.
The US had taken this step as part of its strategic move to restrict China’s access to high-end chips. Despite this, DeepSeek has demonstrated that cutting-edge AI models can be developed at minimal cost, even with limited resources.
Early Experiments on A100 Chips
DeepSeek has for the first time acknowledged that initial experiments for the model were conducted on Nvidia A100 chips. However, the final training was entirely carried out on a cluster of 512 H800 chips. This approach not only cut costs significantly but also saved a great deal of time.
These figures are noteworthy, as OpenAI and other American firms often use hundreds of thousands of GPUs for training their large AI models, with reported investments running into billions of dollars.
Direct Challenge to US Tech Giants
In 2023, OpenAI CEO Sam Altman clearly stated that training a large AI model costs over $100 million (around ₹830 crore). In this context, DeepSeek’s claim of creating a model for just $294,000 has stunned the entire tech industry.
This is also why, when news of DeepSeek-R1 first surfaced in January 2025, global tech stocks saw a sharp decline. Investors grew concerned that if powerful models could be developed at such low cost, the long-standing dominance of US tech firms might come under threat.
Debate Over Distillation Technique
The biggest controversy surrounding DeepSeek’s model lies in its training methods. Several experts allege the company relied on the distillation technique, indirectly learning from American models.
Under distillation, an AI model is trained on the outputs of another model. Reports suggest DeepSeek admitted that some of its models were distilled from Meta’s open-source Llama model. In addition, data collected from the internet also “incidentally” contained responses generated by models such as OpenAI’s.
DeepSeek, however, argues that it used open-source and publicly available data, further emphasising its transparency by publishing the research paper in Nature.
Why Is This News Important?
- Advanced Model at Lower Cost: DeepSeek-R1 proves that building powerful AI models no longer requires billions of dollars. This is a significant step towards democratising AI.
- Challenge to US Dominance: US companies like OpenAI and Anthropic have so far positioned themselves as leaders based on resources and costs. DeepSeek’s model has shaken that foundation.
- Open-Source Advantage: One of DeepSeek-R1’s biggest strengths is its open-source availability. Any researcher or company can download it for free and build their AI projects at lower costs.
- Global Debate: This revelation has sparked fresh discussions worldwide about the cost, transparency, and accessibility of AI development.
Massive Popularity on Hugging Face
The DeepSeek-R1 model is receiving overwhelming support from the open-source community. Reports suggest it has already been downloaded more than 10.9 million times. On platforms like Hugging Face, researchers and tech companies are adopting it on a large scale.
This popularity highlights that technology is no longer confined to the US and Europe—China is now emerging as a new centre of innovation.
Doubts Among US Companies and Experts
American firms have questioned DeepSeek’s claims regarding cost and time. Representatives from Nvidia and US cyber experts argue that the stated resources are insufficient for training a large-scale model.
However, with Nature validating DeepSeek’s training process, the credibility of the company’s claims has gained substantial weight.
New Path for India and Global Startups
The impact of DeepSeek’s breakthrough goes beyond the US and China. AI startups in India and other parts of the world are now encouraged to believe that advanced models can be developed with limited budgets. For countries with restricted resources, this model opens new doors of opportunity.
A New Phase in the US-China AI Race
DeepSeek’s move extends beyond technology—it also carries geopolitical significance. The US has long been regarded as the leader in global AI power, but DeepSeek’s success places China on an equal footing.
If China can achieve such progress despite limited resources and US restrictions, the balance of power in AI technology could shift dramatically in the coming years.
Conclusion
DeepSeek-R1’s training cost of just $294,000 is reshaping conventional thinking. It proves that in the technology race, innovation and strategy matter more than money. The model is not just a warning for American firms but also an inspiration for countries like India, showing that world-class technology can be built even on a limited budget.
The world now waits to see whether models like DeepSeek-R1 will completely redefine the future direction of the global AI industry.
Also Read
