DeepSeek development costs 'overstated'

In its report, DeepSeek said the total training cost was $5.576 million, largely due to the cost of renting Nvidia’s graphics processing units. The company also cautioned that this figure only reflects “formal training” and does not include costs related to previous research or experiments on architecture, algorithms, and data. This immediately attracted investors and caused global technology stocks to “evaporate” $1,000 billion in value in just one day on January 27.

Are DeepSeek's operating costs much higher than claimed?

Now, a report from semiconductor research and consulting firm SemiAnalysis offers a different take on DeepSeek’s development costs. The firm estimates DeepSeek’s hardware spending to be much higher than claimed, noting that the R&D and total cost of ownership were significant.

SemiAnalysis said that generating “synthetic data” for training models would require “significant amounts of computing power.” Additionally, companies would need to experiment, develop new architectures, collect and clean data, and pay employees.

DeepSeek Needs 50,000 GPUs Instead of Just Over 2,000 GPUs?

According to the calculations, experts said DeepSeek operates a large-scale computing system consisting of about 50,000 Hopper GPUs, with 10,000 H800 GPUs and 10,000 more powerful H100 GPUs, along with additional H20 GPUs. This number contradicts DeepSeek's previous claim of using only 2,048 Nvidia GPUs. The total capital expenditure for the server is estimated at about $1.6 billion, while the operating costs are up to $944 million.

SemiAnalysis believes 50,000 Nvidia GPUs were used by DeepSeek

DeepSeek was founded in 2023 by Liang Wenfeng, co-founder of High-Flyer, a Chinese AI-focused hedge fund. The startup grew out of the fund's AI research unit in April 2023, with the goal of developing large language models and achieving artificial general intelligence (AGI).

Interest in DeepSeek grew when the company released R1, a reasoning model that competes with OpenAI's o1, but is open source, allowing any AI developer to use it. However, like many other Chinese chatbots, DeepSeek also has limitations on certain topics.

OpenAI CEO Sam Altman praised DeepSeek's model, noting that DeepSeek is "clearly a great model." However, he also said there was evidence that DeepSeek had collected OpenAI's data to work in a "distilled" manner.

Source: https://thanhnien.vn/chi-phi-phat-trien-deepseek-bi-tang-boc-185250203151508438.htm