How DeepSeek caused panic in the US stock market and overshadowed chatGPT

This week, DeepSeek R1 artificial intelligence not only wreaked havoc on the US stock market, but it also hit OpenAI hard, outpacing it in the technology race. Its sudden exit caused market capitalization to collapse by $1 trillion, of which $600 billion Nvidia’s losses. Some companies have already started to recover, but it’s clear that DeepSeek R1 has had a huge impact on the AI industry and the major players in computing.

DeepSeek R1 has had a huge impact on the AI industry and the major players in computing.

The big question is, how did Chinese company DeepSeek achieve this result? We get into the details: what DeepSeek R1 is, why OpenAI has been accused of hypocrisy, and how the myth that the model is cheap to train caused a panic reaction from investors.

What is DeepSeek R1?

What is DeepSeek R1 and how is it different from DeepSeek V3?”

.

DeepSeek has unveiled two major AI models at once: V3 and R1. Despite their similarities, it was the DeepSeek R1 that became the main cause for discussion.

• DeepSeek V3 is a MoE-LLM (Mixture of Experts) with 671 billion parameters. It is a universal language model.
• DeepSeek R1 is an advanced AI with advanced reasoning capabilities, analogous to OpenAI o1, but with the additional feature of «talking to itself» before giving an answer.

.

DeepSeek R1 is based on DeepSeek-V3-Base and is available in a variety of options ranging from 1.5B to 70B parameters. A full version with 671B parameters is also available for download.

DeepSeek R1 can be used through the company’s web interface, like ChatGPT. However, DeepSeek’s servers are located in China, and some of the requests are subject to censorship. An interesting point: when running the model on local hardware, restrictions can be lifted by using special methods to «clean» the model, which is already being actively done by enthusiasts.

.

Did DeepSeek R1 steal OpenAI data?”

Despite the lack of formal charges, OpenAI and Microsoft have already launched an investigation to find out if DeepSeek used someone else’s designs. However, there’s an important nuance here: if even DeepSeek «stole» OpenAI’s data, it looks like utter hypocrisy.

And if DeepSeek has stolen OpenAI’s data, it looks like utter hypocrisy.

OpenAI has itself been repeatedly accused of illegally using content in training its models. Lawsuits filed include litigation with The New York Times, Intercept Media, Canadian News Outlets and other publications.

OpenAI Canadian News Outlets and other publications.

OpenAI has explicitly stated that «training AI on open data from the Internet is fair use» and that this principle is important for competition in the AI field. Now, if DeepSeek has used a similar strategy, OpenAI simply has no moral right to be outraged.

And if DeepSeek has used a similar strategy, OpenAI simply has no moral right to be outraged.

Nevertheless, some experts suspect DeepSeek of employing the method of «distillation» -training one model based on another through a constant exchange of questions and answers. US venture capitalist David Sachs claims there is «substantial evidence»that DeepSeek R1 learned by copying GPT-4o’s knowledge.

DeepSeek R1 is learning by copying GPT-4o’s knowledge.

There is, however, a serious objection to this theory: The OpenAI o1 model hides the logical chain of its decisions. Therefore, even if DeepSeek used it, there is simply nowhere to get the data to fully reproduce the reasoning mechanism.

Why are investors scared? Is it realistic to train an AI for $5.5 million?”

.

Financial markets were alarmed by DeepSeek’s claim that training the model only cost $5.576 million. If true, it undermines the entire business model of Nvidia, which makes money from selling powerful GPUs for AI training.

If true, this undermines the entire business model of Nvidia, which makes money from selling powerful GPUs for AI training.

The figure $5.5 million comes from whitepaper DeepSeek V3, however, and it reflects the cost of final training, not the entire development process. This amount does not include the costs of testing, previous versions of the model, and research.

Despite this, DeepSeek did use less powerful video cards Nvidia H800, which are 50% slower than the top-of-the-line H100, but still did well. This called into question the need for the most expensive GPUs for AI training, which caused Nvidia’s stock to plummet.

Nvidia’s stock plummeted.

In addition, a rumor broke out that DeepSeek had illegally purchased 50,000 Nvidia H100 chips, circumventing U.S. export restrictions. However, Nvidia and Singaporean authorities said that there was no evidence of a sanctions violation.

Singaporean authorities said they had not been able to prove that DeepSeek had violated the sanctions.

What does this mean for the future of AI?”

.

Despite the scandals, the DeepSeek R1 is a huge step forward for the industry. Cheaper models, powerful reasoning algorithms, and «open weights» (not to be confused with «open source») mean that AI technologies will become more accessible and efficient.

And it’s a huge step forward for the industry.

Already, enthusiasts are working to create Open R1, DeepSeek R1’s open data-based counterpart. Competitors, including Google and Meta*, are actively studying DeepSeek’s research to improve their own models.

What’s more, even OpenAI has recognized DeepSeek’s influence, pledging to make its AI models more transparent in terms of reasoning logic.

OpenAI has also recognized DeepSeek’s influence, pledging to make its AI models more transparent in terms of reasoning logic.

DeepSeek R1 can already be run locally, for example via LM Studio on a MacBook Pro with an M4 Pro chip using version 32B of Qwen.

Conclusion

DeepSeek R1 proved that you can train powerful AI without gigantic costs and without top Nvidia chips. This shook up the market, questioned OpenAI’s business model, and sparked data theft scandals.

But in the end, everyone wins: competition is accelerating AI development, and DeepSeek’s open research will help improve future technology for everyone.

So, in the end, everyone wins: competition is accelerating AI development, and DeepSeek’s open research will help improve future technology for everyone.

* Owned by Meta, it is recognized as an extremist organization in the Russian Federation and its activities are banned.