ChatGPT Introducing a New Opponent: Meet DeepSeek

The Chinese small company DeepSeek surprised the biggest players of Silicon Valley with its new AI virtual robot that equals the performance level of OpenAI's chat assistant, ChatGPT, while costing much less to develop.

To date, DeepSeek has surpassed ChatGPT on Apple's App Store as the best free-downloaded application in the United States, the UK, China, and many other countries. It works like a typical chatbot: you enter a request and its open-source model generates a response. The underlying DeepSeek-V3 model contains 671 billion parameters, allowing the application to "think" before solving the problem.

Unlike OpenAI's model that can only run on its own servers, DeepSeek can operate locally on high-end computers and many servers equipped with GPU. The company claims that its model corresponds to OpenAI's o1 on certain benchmarks. Unlike many other AI chatbots, DeepSeek also transparently shows its reasoning and how it derived a response. However, as a Chinese application, DeepSeek censors certain topics such as the Square Tiananmen.

The app, launched on January 20th, is also available in the Google Play store. DeepSeek gained sudden success thanks to its ability to match or surpass established AI models. Its research article published on Monday reveals how the training of DeepSeek-V3 was also costly.

While his statements have not yet been verified, DeepSeek seems to have used only 2,048 specialized Nvidia H800 chips for training R1, compared to more than 16,000 Nvidia chips for training the most advanced models of OpenAI. DeepSeek claims that drastically reducing the number of top-tier GPU chips needed for AI learning allowed it to spend only $5.6 million to train R1. Compare this to OpenAI, which spent over $100 million to train its GPT-4 model comparatively in size. However, the company has not quantified DeepSeek's energy consumption compared to competitors.

R1 is based on the large language model DeepSeek V3 (LLM), which the company claims corresponds to OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet. DeepSeek's achievement is particularly impressive in light of the US commercial sanctions imposed on powerful sophisticated Nvidia chips used for AI processing. With the biggest players in AI like Nvidia, OpenAI, Meta, and Microsoft investing billions in their AI data centers, we will see many headlines this week questioning the lead of the United States in AI.

It will be fascinating to see how DeepSeek's progress affects the Trump administration's billion-dollar Stargate project - supported by OpenAI, Softbank, and Oracle - with the goal of investing $50 billion in new AI infrastructure and data centers for OpenAI over the next four years. This will be particularly interesting when OpenAI transforms from a non-profit organization into a profit-making company, as DeepSeek has open-sourced its AI models.

Meta has also open-sourced certain aspects of its AI technology, such as the large language model Lama. However, DeepSeek is a new entrant attracting much attention, and the fact that developers can freely build on DeepSeek's technology could give price competition to OpenAI.