DeepSeek spent US$6 million training its V3 model. OpenAI spent about US$100 million on GPT-4 in 2023. That is a 94 percent cost difference.
The numbers come straight from the company itself, and they sit at the center of why the AI industry is paying attention to a Chinese startup barely a year and a half old. Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. — DeepSeek for short — launched its chatbot and the DeepSeek-R1 model on January 15, 2025.
The company was founded in July 2023 by Liang Wenfeng, who also runs High-Flyer, a Chinese hedge fund. That timeline matters. Most AI labs take years and billions of dollars to reach this point.
DeepSeek did it in 18 months, on a budget that looks like pocket change next to its competitors. The V3 model’s training cost is not the only number that stands out.
DeepSeek says its model used roughly one-tenth the computing power that Meta needed for Llama 3.1, a comparable open-weight model. Computing power is the single biggest expense in large language model development — it is why companies like OpenAI and Google spend fortunes on data centers packed with Nvidia chips. DeepSeek appears to have found a way to do more with less.
The company describes its models as “open-weight.” That means the exact parameters — the mathematical settings that make the model work — are shared publicly. The training data is not. Developers can download the model, modify it, and build their own applications on top of it.
That is a different approach from OpenAI, which keeps GPT-4’s weights secret and charges for access through an API. It is closer to what Meta did with Llama, but at a fraction of the cost.
DeepSeek-R1’s responses have been compared to those from GPT-4 and OpenAI’s o1 model. That is a high bar. GPT-4 was released in March 2023 after years of development and billions in funding.
DeepSeek-R1 arrived in January 2025, trained on a hedge fund’s budget, and reportedly holds its own in performance. The phrase “upending AI” has been used to describe what the company has done.
The implications are straightforward. If a small team with limited resources can build a competitive model for US$6 million, the economics of the AI industry shift. The barriers to entry that kept most companies out — the need for massive capital, huge data centers, and thousands of GPUs — suddenly look lower.
More players can enter the field. The current leaders may find their advantages shrinking. DeepSeek’s training efficiency also raises questions about how much of the industry’s spending is actually necessary.
OpenAI spent US$100 million training GPT-4. DeepSeek spent US$6 million training V3.
The gap is enormous. If the results are genuinely comparable, then a lot of the money flowing into AI development may be waste. That is not a comfortable thought for the companies that have raised billions on the assumption that bigger budgets mean better models.
Liang Wenfeng started DeepSeek in July 2023. By January 2025, his company had released a model that competes with the best in the world.
The hedge fund background may explain part of it — financial firms are used to finding efficiencies, cutting costs, and getting maximum return on investment. DeepSeek applied that mindset to AI research and produced something that looks like a bargain. Whether DeepSeek can sustain this advantage is unclear.
One model does not make a company. But the numbers are public. The model is available.
And the industry is taking notice.
