DeepSeek AI Causes A Huge Blow to America’s Monopoly on AI Technology

In 2022, THe U. S. Department of Commerce banned the export of artificial intelligence chips to China. The race for dominance in artificial intelligence technology was thought to be over by many experts. Without powerful chips like Nvidia’s A100, A800,  H100, and H800; China’s progress in AI technology suffered a blow. China took a different route to acquire these chips which some might call illegal. 

Among many contemporary AI software, DeepSeek AI has stirred the pot with their startlingly strong performance. DeepSeek has outperformed and competed with market-leading AI software like ChatGpt, Gemini, and Claude in different aspects. More surprisingly, where Google, OpenAI, and Meta spent hundreds of millions of dollars, it took only $5.57 million and 2 months to build this model according to CNBC. 

DeepSeek R1 is the latest version of Generative AI from DeepSeek. With 30% cheaper service, it offers almost identical quality. Third-party quality testers have proved that DeepSeek models have outperformed GPT-4.o and Claude 3.5-Sonnet in terms of math coding and scientific reasoning tasks. US companies like Meta, Google, and OpenAI used the most powerful chips available on the market, the Nvidia H100. DeepSeek used a lesser chip, the Nvidia H800 and got better results than expected. This outworldly success of DeepSeek R1 has changed the perception of how developers see AI engineering. 

What makes DeepSeek so much more than just another GenAI is an idea that has been here for a while: Open Source AI Software. Open source refers to the sharing of underlying code with other businesses and researchers. This lets people outside DeepSeek to build and distribute their own software or products using the DeepSeek technology. 

There have been many arguments about whether OpenAI or other US GenAI software should be made into open source or not. The US experts have always said that they do not need to make this software Open Source. However, DeepSeek has proved that there is a huge demand for open-sourced GenAI. Also, these software are cheaper than closed software like ChatGPT.

A team at UC Berkeley has developed a GenAI called Sky-T1 with less than $500 budget for computing power under the leadership of Dr Stoica according to the New York Times. They relied on open-source software as well as cloud hosting, cloud GPU, and efficient computational resources.

This poses a question to these big developers; “Are they wasting resources by using more when evidence says it is possible to train and build GenAI with much fewer resources?”

The UC Berkeley project proves that it is possible to build this software in a cost-efficient way. Albeit this project is nowhere near the scale of OpenAI, Gemini, or Claude. However, this project along with DeepSeek R1 proves that it is certainly possible to build Generative AI without spending billions of dollars. 

To train their AIs, the leading companies in the world like Meta, OpenAI, and Google use supercomputers that use around 16,000 chips. DeepSeek, on the other hand, said that they only used 2000 specialized chips to train their AI. With the US embargo on China regarding technology, the engineers had to work around the ban and make do with only a handful of resources compared to the leading companies. 

DeepSeek’s success comes from its astonishing reasoning capabilities. This reasoning was developed with the ‘Reinforced Learning’ (RL) method. Here, they rewarded the AI for making the correct decision and punished it for making a wrong decision. However, the AI did not know which was good or bad. After some decisions, DeepSeek learned to follow a path based on the results. 

Before Reinforced Learning, a few engineers fine-tuned the AI by pointing out the desired output they wanted. They also provided context about their prompts and the desired results. Then reinforced learning provides different outputs and the engineers rank the best ones. This process is repeated as long as the AI learns to provide satisfactory outputs consistently. Moreover, because of RL, DeepSeek-R1-Zero developed sophisticated capabilities. These include self-verification and reflection. The programmers say that they did not explicitly programmed it to do so but through RL, it learned how to self-verify and reflect. 

The result of this training is impressive to say the least. DeepSeek-R1-Zero has surpassed OpenAI’s o1 reasoning model. It outperformed o1 in the AIME 2024 mathematics benchmark with a 79.8 success rate. Similarly, it outperformed 96.3% of human coders demonstrating “expert level” performance on standardized coding tests. It achieved a 2029 Elo rating on codeforces while outperforming human coders. 

Despite this performance, DeepSeek is unbelievably cheap. It runs queries at just $0.14 per million tokens. In contrast, OpenAI spends $7.50 per million tokens making it 98% more expensive than DeepSeek. Its completely open-source code and training method are licensed under MIT, so it can be used and modified without restriction. 

DeepSeek’s better-funded competitors are now in hot water as experts are starting to ask questions.

Stability AI’s founder Emad Mostaque posted a provocative status on X saying “Can you imagine being a frontier lab that’s raised like a billion dollars and now you can’t release your latest model because it can’t beat DeepSeek?“.

There are even questions about OpenAI’s pricing model as DeepSeek is providing the same services with a fraction of their price.

Arnaud Bertrand compared their pricing in his X account. He wrote “Say you want to build a service that analyzes codebases for security vulnerabilities. A typical enterprise codebase might be 1 million lines of code or roughly 4 million tokens. That would cost $60 with OpenAI versus just $2.20 with DeepSeek. At OpenAI’s prices, doing daily security scans would cost $21,900 per year per codebase; with DeepSeek it’s $803.” which shows that DeepSeek is 97% cheaper than OpenAI while providing similar or better services.” 

DeepSeek has claimed it had been a victim of “Large-scale malicious attacks” hindering their services. Despite that, they have launched another model named “DeepSeek Janus-Pro-7b”. Janus-Pro is an image-generation AI model and it claims to outperform OpenAI’s Dall-E 3 and Stability AI’s Stable Diffusion in key aspects of image generation benchmarks. 

This remarkable rise of DeepSeek AI has affected the US economy deeply. On January 28th, Nvidia alone lost $600 Billion after the release of DeepSeek R1-Zero. The Shockwave was felt throughout Wall Street. Around $1 Trillion was erased from the Nasdaq’s market value and has severely rattled the economy of the United States of America. 

On the other hand, the consumers are loving DeepSeek AI very much. It has dethroned OpenAI as the most downloaded free app in the U.S. on the Apple store. Given that this happened within 2 months of its initial release makes this feat immensely applaudable. 

The war for technological domination has already started. The ban on chip export has backfired on the US. It forced China to build a much more efficient solution with just a fraction of OpenAI’s budget. They are now sharing this with the world at just 3% of OpenAI’s prices.

The new era of GenAI has begun and the ball is now in the US court. Tech enthusiasts are eager to see if they will fold under pressure or rise up with something even better than we can imagine.