DeepSeek, the AI-focused Chinese company, has disrupted the industry with an unprecedented price-to-performance ratio. This has caused Wall Street to crash for companies like NVIDIA, which lost $600 billion in market capitalization in a single day. However, experts have cast doubt on the company’s training investment figures. Now, an analyst firm says that DeepSeek AI has actually spent $1.6 billion on its NVIDIA GPU fleet.
As a reminder, DeepSeek claimed that training its most advanced R1 model cost only $6 million. This is just a small fraction compared to the hundreds—or billions—that other big companies have spent so far. However, a report by SemiAnalysis claims that the figure only corresponds to the GPU time needed for pre-training. The Chinese company has not accounted for the research stages, model refinement, data processing expenses, and general infrastructure costs, the report says.
DeepSeek AI spent $1.6 billion on its NVIDIA GPU fleet, SemiAnalysis claims
DeepSeek claimed it achieved the cost reduction in part by using older NVIDIA H800 chips for AI training instead of newer equipment. But SemiAnalysis says the Chinese firm has a fleet of 50,000 Nvidia Hopper GPUs that could include 10,000 H800s, 10,000 H100 units, and additional purchases of H20 hardware. This is in line with experts who called the alleged $6 million spending “likely a fictional story.”
In fact, this isn’t the first time we’ve heard about DeepSeek’s 50,000 NVIDIA Hopper chips. Scale AI CEO Alexandr Wang spoke about it last month, receiving backing from Elon Musk.
Obviously
— Elon Musk (@elonmusk) January 27, 2025
In total, SemiAnalysis estimates actual spending on the company’s AI models at $1.6 billion. Although DeepSeek may have come out of nowhere, it has some serious financial muscle behind it. It is backed by the Chinese High-Flyer fund, which has a valuation of $8 billion. High-Flyer launched DeepSeek in 2023 as a separate company focused primarily on AI development.
The keys to the company’s rapid progress
Beyond the potentially bogus AI training cost numbers, the report highlights strategic investments and a strong workforce as the keys to DeepSeek’s success. The company exclusively hires Chinese engineers with strong problem-solving skills. They come primarily from institutions like Peking University and Zhejiang University. Plus, the firm offers fairly high salaries, with some engineers earning up to $1.3 million annually.
DeepSeek also has full control over its data centers. This gives it a direct advantage over other startups that often rely on cloud AI hardware providers. This increases the efficiency of its developments by offering immediate interaction without bottlenecks in between.
