DeepSeek Misled AI Markets: Actual Training Costs Are 400 Times Higher Than Reported

The unfolding debate regarding the training expenses associated with DeepSeek’s R1 model has sparked significant market volatility, revealing a layer of deception that was previously unrecognized. The actual financial implications are indeed more striking than initially claimed.

Revised Insights on DeepSeek’s Training Expenses Surpassing the Reported “$5 Million”

A comprehensive study by SemiAnalysis has challenged the prevailing narrative that DeepSeek’s R1 model has dramatically optimized efficiencies, thereby diminishing the need for robust computing resources, such as those offered by NVIDIA. Initially, industry analysts reported that DeepSeek’s expenses for their R1 model amounted to merely “$5 million”—a figure comparable to the costs associated with OpenAI’s GPT models. This revelation instigated a wave of panic among retail investors that adversely influenced the US stock market. However, as more information has surfaced, a deeper examination of the true costs involved is warranted.

To put things into perspective, DeepSeek began as a subsidiary of the Chinese hedge fund High-Flyer. According to SemiAnalysis, back in 2021, DeepSeek acquired 10, 000 units of NVIDIA’s A100 GPUs before stringent export controls were enforced. Following the parent company’s decision to spin off the project, DeepSeek rapidly expanded its operations, significantly increasing its computational capabilities.

DeepSeek hardware costs analysis — Image Credits: SemiAnalysis

The findings indicate that DeepSeek is equipped with around 10, 000 “China-specific” H800 AI GPUs, alongside 10, 000 cutting-edge H100 AI chips. Additionally, the organization has invested in NVIDIA’s H20 AI accelerators, creating a substantial “pool”of resources shared between DeepSeek and High-Flyer’s ventures for activities including trading, research, training, and inference. Cumulatively, the CapEx for DeepSeek is estimated at approximately $1.6 billion, with operational expenses rumored to reach around $944 million—figures that are up to 400 times greater than earlier market calculations.

It’s important to note that the initial cost estimate likely represented only a fraction of the total training costs associated with running the R1 model. Remarkably, DeepSeek has demonstrated an aptitude for attracting local talent by hosting recruitment events at prestigious universities, with selected employees receiving salaries exceeding $1.3 million. This strategy has empowered DeepSeek’s development team to engineer competitive solutions capable of standing toe-to-toe with industry giants like OpenAI. Unfortunately, the misreported financial figures provided the impetus for last week’s so-called black swan event in the market.

SemiAnalysis has undertaken extensive evaluations of DeepSeek’s AI model; readers are encouraged to explore these findings for further insights and intricate details.

Source & Images