Since launching on Jan. 20th, the Chinese app Deepseek has become the top-rated artificial intelligence model in the US. Even President Donald Trump praised the app for its incredible performance and affordable cost. He and tech investors noticed the AI app was stripping millions of dollars from its competitors including NVIDIA. So how was Deepseek made and why is it affecting US company sales?
Fueled by US-China Tech Exports Ban
In 2021, the Biden-Harris Administration issued an export ban against China’s chipmaking. According to then-president Joe Biden, the country’s technology threatens US security due to artificial intelligence development. Then in mid-2023, China introduced a counter-measure on US-partnered manufacturers based on the minerals used in their semiconductors. Any exports found containing gallium, germanium, antimony, and “superhard” metals will be disbanded.
The Chinese Ministry of Commerce saw these materials mainly produced in their own country. The organization believed that preventing them from shipping to the US decreases manufacturers’ sales tremendously, The ban was published on Dec. 2nd, 2024 during Joe Biden’s third semiconductor crackdown. Within those three years, Hangzhou engineers began developing Deepseek.
Challenging AI-Powered Limitations
During 2021, High-Flyer AI co-founder Liang Wenfeng invested in 10,000 NVIDIA A100 chips. From there, he and his team tested LLMs for Deepseek’s prototype modules. They wanted to have the AI trained using math functions to produce quantitive investments. Then in May 2023, Wenfeng released the app to gain feedback, earning $8 billion to invest in further research.
The app’s recent versions, Deepseek V3, and Deepseek-R1, are modded to use Mixed of Experts (MoE) and Reinforcement Learning. These two functions influence the AI to execute auxiliary-loss-free strategies to maintain load balancing. This also trains the model’s interference acceleration to detect and improve any failing function. Wenfeng has revealed to the press that Deepseek R1 has been tested on 2,000 NVIDIA H800 chips.
According to Scientific American, they have a low-powered render output unit (ROP). The A100 runs on 160 ROPs while the H800 uses 24 ROPs. R1 also stirred trouble in the AI market when Wenfeng revealed the project cost less than $6 million to create.
Why NVIDIA Stocks Sank From Deepseek
In recent news, NVIDIA reported an all-time low of $600 billion, which doubled the loss from September 2024 ($279 billion). The company has been an entrusted partner for producing powerful GPUs to optimize games and video through AI-assisted technology. At the start of 2025, CEO Jensen Huang’s presentation on NVIDIA’s RTX 50 raised questions.
NVIDIA manufactured their new AI chip, Blackwell that offers generative AI capabilities and claimed to have 1,801 TOPS operating the RTX 5090. The problem that turned most of their consumers away was the chip’s capabilities only showed video optimization. Then disappointment pushed further when Huang shared the 4-digit prices. Now with Deepseek entering the US AI market, NVIDIA will need to rethink Blackwell’s launch strategy.
Deepseek V3 and R1 are open-sourced and being offered for free through website registration. When Wenfeng was asked about AI’s intention in a 2024 interview, he shared what is needed for the AI industry. “Our principle is neither to sell at a loss nor seek excessive profits. The current pricing allows for a modest profit margin above our cost. We will not go closed-source. We believe that establishing a robust technology ecosystem matters more.”
Deepseek’s Recent Cyber Attack
On Jan 27th, Deepseek announced they are dealing with malicious cyber attacks. While Trump is an avid AI supporter, some people are trying to prevent it from taking off in the US. Wenfeng and the Deepseek team are taking precautious measures to limit registrations until further noticed. As for their competitors, including Sam Altman’s OpenAI and Chat-GPT, they will need to adapt to these circumstances that AI does not need a powerful CPU chip.
