DeepSeek: The People’s ChatGPT
As of this Wednesday afternoon (written on the week of Jan 29th), the situation in China and the U.S. is significant: The U.S. stock market began the week with substantial losses — on Monday, Nvidia, the largest chip designer, experienced a 17% drop in its shares. Meanwhile, China was during its Lunar New Year holiday. This year marks the Year of the Wooden Snake, a symbol in Chinese tradition that embodies extension, flexibility, resilience, and diplomatic “maneuvering.” Fittingly, it appears that DeepSeek, a new large language AI model emerging from China at the end of 2024, will shape this new year. The recent decline in U.S. stock markets can be associated with DeepSeek’s viral breakout on social media this week.
Since 2022, the U.S. has been imposing a high-tech chip embargo on China, which was expanded last week. The latest restrictions limit access to the most advanced semiconductor chips to just 18 countries, including members of the Five Eyes intelligence alliance — Canada, New Zealand, Australia, and the UK — along with key Asian chip-producing nations and ten NATO allies considered close to the U.S. However, Turkey and 17 other NATO members are subject to export restrictions. The U.S. believes China could bypass these restrictions by sourcing AI-training chips through third countries. Despite the embargo, there are claims that DeepSeek trained its new model using chips acquired through illicit means.
DeepSeek delivers results comparable to GPT-4o, the leading large language model powering ChatGPT. The difference? DeepSeek is twice as fast and operates with 90% less data and computing power. In other words, DeepSeek has achieved what U.S.-based OpenAI did but with significantly weaker processors and at a fraction of the cost. You could call it the “People’s GPT” from the People’s Republic of China. Just as China previously scaled Western-developed solar energy and battery technology to drive down prices, it has now done the same with large language models, making them more efficient, scalable, and cost-effective.
Unlike closed-source models like GPT, whose code remains private, DeepSeek is open-source. More precisely, it is built upon LLaMA, a model initially developed by Facebook. Some reports even claim that DeepSeek’s training data included text generated by GPT-4, as the model allegedly refers to itself as “GPT-4” when answering specific questions.
When GPT launched two years ago, many anticipated that China would quickly produce a rival — but they were disappointed. After all, AI models are, at their core, statistical systems. Without enough variation in training data, the results suffer. Despite its technical prowess, China’s limitations on free speech meant that its large language models struggled with diversity in their datasets, resulting in subpar performance. Then came the U.S. chips embargo.
DeepSeek is owned by High-Flyer, a Chinese hedge fund. The world’s top AI talents often work in finance, where the money is, and High-Flyer initially focused on AI models for financial applications. Later, they decided to venture into large language models — not for consumer use but for enterprise applications. This allowed them to avoid China’s strict content regulations.
One significant challenge for Chinese firms competing with U.S. giants like OpenAI is gaining access to advanced chips due to the American chips embargo. However, such restrictions often drive innovation. The blockade has pushed China to develop a more efficient method for constructing an advanced AI model that requires significantly less processing power. Similarly, when the U.S. imposed an arms embargo on Turkey following the 1974 Cyprus Peace Operation, it encouraged the growth of Turkey’s domestic defense industry.
The chips embargo on China is now accelerating its domestic AI development. Of course, AI depends on three key inputs: human expertise, data, and chips. While China faces chip shortages, it possesses an abundance of skilled talent and limitless data.
Thanks to open-source models, China has an easier path to refining and scaling AI systems initially developed in the U.S. While the chip embargo may slow progress, it won’t halt it — and may even encourage further breakthroughs.
Technology is becoming more politicized than ever before.
Who stands to gain from these developments? China is benefiting, but I don’t think American firms will necessarily lose. Companies that provide computing power — like Microsoft and Amazon — or chip manufacturers like Nvidia could be the ultimate winners. According to Jevons’ Paradox, increasing efficiency in resource use leads to higher demand for that resource. When the steam engine was invented, energy consumption became more efficient, but total energy demand still soared. Similarly, AI models will drive demand for computing power, benefiting chipmakers and cloud providers. The real question is whether companies like OpenAI, which focuses solely on model development, will survive this shift. Integrating with established tech giants like Microsoft might accelerate as AI models become commoditized.
Reading Recommendation: Azeem Azhar’s weekly newsletter
This article is a translated version of “Halkın ChatGPT’si DeepSeek” which was initially published in Economic Daily (Nasıl Bir Ekonomi Gazetesi) in Economic Daily on January 31, 2025.
