China's Deepseek claims that the flagship AI system known as the R1 was trained for just $294,000.
Details are published this week in a peer-reviewed paper on Nature, which could encourage further discussion of Beijing's ambitions in the global artificial intelligence race. The Hangzhou-based company said the inference-focused model was trained using the 512 NVIDIA H800 chip. The hardware was specifically designed for China after the US banned the sale of more powerful H100 and A100 processors.
The paper, co-authored by founder Liang Wenfeng, marks the company when it first discloses such expenses.
Deepseek uses only a small portion of the cost of the US model
In January, the sale of high-tech stocks was made as Deepseek's cheap AI tools released a volatile global market could overturn established giants such as Nvidia and Openai.
However, Liang and his team have been making a point of notable appearance since then only sporadic product updates.
The reported price tag of $294,000 is in contrast to estimates from American companies.
In 2023, Sam Altman's Openai CEO said “it's well over $100 million for training basic models.” However, he did not give any particular breakdowns.
Training large language models involves long running banks of powerful chips and spending enormous amounts of power while processing text and code. Industry observers have long envisioned bills for such projects, ranging from tens of millions or millions of people.
That assumption is currently being challenged, and in the supplementary document, DeepSeek owns A100 chips and allows them to be used in early development before moving to the H800 cluster. According to the high-tech company, the model ran for 80 hours during the final training phase.
Nvidia claims that Chinese startups can only access H800 processors, but American officials remain skeptical. A few months ago, a US source told Reuters that Deepseek illegally owns a large number of H100 chips it exported to China.
Putting innovation under the microscope
The R1 is attracting attention as it may be the first major model to receive formal peer reviews as well as low training costs.
“This is a very welcome precedent. Without this sharing norm, risks can be very difficult to assess,” said Luis Stanstall, a machine learning engineer who hugged Face, who reviewed Nature Paper.
The review process encouraged DeepSeek to clarify technical details, including how its models trains and safeguards in place.
“We're excited to see the latest trends in our research and development,” said Huan Sun, an AI researcher at Ohio State University.
A key breakthrough for Deepseek was using a pure reinforcement learning approach. Instead of relying on human-curated examples of reasoning, according to the paper. This model was rewarded by solving problems correctly and gradually developing a unique problem-solving strategy.
The company says that the trial and error system allows the R1 to verify its behavior without copying human tactics.
“This model is extremely influential,” Sun added. “It's possible that almost all of the reinforcement learning tasks in 2025 were inspired in some way by R1.”
Deepseek refuses to make a claim for copy
Shortly after the release of the R1, speculation swirled that Deepseek relied on the achievements of rivals, particularly from Openai, to accelerate training. However, the company is currently denialing the claim completely.
In communication with the judge, Deepseek allegedly did not copy the example of inference generated by Openai. However, like most large language models, it was trained with Internet text. This means that some AI-generated content is inevitably included, and the explanation has convinced some reviewers.
“We cannot be 100% sure that R1 is not trained in the Openai example. However, attempts at replication by other labs suggest that reinforcement learning is sufficient in itself,” Tunstall said.
Deepseek says R1 is built to be excellent for inference-rich tasks such as coding and mathematics. Unlike most closure systems developed by US companies, it was released as an open weight model that researchers can freely download. It has already been downloaded over 10 million times as it hugs the face of the AI community site.
The company spent about $6 million in developing the base model on which the R1 is built, but even if it was added, its costs were not well below the amounts associated with its rivals. For many in this field, it makes the R1 attractive.
Sun and his colleagues recently tested a system on scientific data tasks, and found it to be the least accurate and best in terms of cost and performance.
The smartest Crypto Minds has already read our newsletter. want? Join them.
