The new Chinese AI has disrupted the US stock market – here’s what you need to know about DeepSeek.
DeepSeek's AI Model Shakes Up the Market
The launch of a new AI chatbot from a previously little known Chinese start-up has sent shockwaves through the US stock market. The new open-source AI model from DeepSeek – a tiny company which is said to employ no more than 200 people – initially wiped almost $600 billion off American chipmaker Nvidia’s stock value. Claiming to offer an app which functions in a similar manner to OpenAI’s ChatGPT and Google’s Gemini, developers behind DeepSeek say it offers users a similar experience – but for a fraction of the cost. Here’s what you need to know about DeepSeek’s new AI model and why it has caused such a stir.
DeepSeek: The Startup Behind the Shockwave
Founded in Hangzhou, China in 2023, DeepSeek is a startup company run by Liang Wenfeng who previously co-founded High-Flyer, one of the country’s top hedge funds which uses AI to analyse the data behind financial investment decisions. As well as his background in finance, Liang is a graduate from Zhejiang University with degrees in electronic information engineering and computer science. In an interview with a Chinese news outlet last year, Liang said: “Our principle is not to lose money, nor to make huge profits … our starting point is not to take advantage of the opportunity to make a fortune, but to be at the forefront of technology and promote the development of the entire ecosystem.”
In 2021, Liang reportedly purchased around 10,000 Nvidia graphics processing units (GPUs), a year before Joe Biden announced export restrictions on high powered chips heading to countries such as China. Introduced with the aim of stopping the country from being able to access the equipment needed for rapid AI development, the export of the Nvidia A100 chip was banned. The US company then developed the H800 chips for the Chinese market, which DeepSeek trained its AI model on, before they were also banned in 2023. Chips such as these are key in building and running AI systems – which is why Nvidia’s stock value was so high and why it took such a plunge when a paper about DeepSeek’s newly launched R1 model hit the news last week.
A "Sputnik Moment" for AI?
The reason why people are calling DeepSeek’s AI a “Sputnik moment” comes down to the company’s claim that they were able to create their R1 model at a fraction of the cost of industry leaders such as OpenAI. But it’s not only that. According to key benchmark figures released by DeepSeek, their open source R1 model is not only comparable to OpenAI’s o1 model – which was released at the end of last year – it even outperforms it in a few key areas. On top of that, it’s free to use whereas o1 can cost anywhere from $20 per month to $200 per month for the Pro version. While benchmark testing isn’t everything, DeepSeek has shaken the foundations of Silicon Valley and signalled to AI developers – and more importantly their investors – that more money and greater resources don’t necessarily result in better AI.
Impact on the Industry
As well as hitting companies such as Nvidia, which provide both hardware and software for AI development, it will also impact companies such as OpenAI who are trying to sell services. Concerns have also been raised about DeepSeek’s presence in the space potentially undermining a $500 billion investment into the Stargate Project, which aims to develop AI infrastructure in the United States, by OpenAI, Oracle and SoftBank.
Nvidia's Plunge and DeepSeek's Advantage
The release of DeepSeek’s AI saw the share prices of US-based tech giants plummet. In particular, chip-maker Nvidia’s share price suffered the biggest ever one-day loss in market value in Wall Street history – around $589 billion. The firm has enjoyed sky-high stock value in recent years, becoming one of the most valuable companies in the world, even ahead of Apple, because of AI chips. Essentially, Nvidia has had a monopoly on training and developing complex AI technology systems for around a decade. Why? Well, a few reasons. The first is Nvidia’s compatibility with Linux operating systems compared to chip manufacturers such as AMD. The second comes down to machine learning algorithms being optimised for software proprietary to Nvidia – which is now industry standard software – that can only be run on Nvidia GPUs. The third, and perhaps most important reason, is down to Nvidia’s existing infrastructure which allows developers to connect GPUs together at scale – which was thought to be increasingly important as AI models get larger and more popular. But DeepSeek has proven that market-leading performance can be achieved with limited resources.
The Role of Export Restrictions
One of the reasons why DeepSeek’s AI was such a shock is, in part, due to restrictions on GPU exports from the US. The restrictions saw countries including China banned from acquiring more powerful AI chips, such as those used by companies such as OpenAI, Meta and more. The ban forced DeepSeek to seek creative solutions to a lack of processing power, which in turn was critical in them innovating a more efficient AI model. The release of R1 then created the right conditions for a dip in confidence when it comes to Nvidia’s hold over the AI market. That being said, DeepSeek still used Nvidia chips when creating R1 and the firm has already claimed back some losses. It is also worth highlighting that, overall, Nvidia shares are up 14% across six months and more than 100% across the past year.
Benchmark Testing and the Future
As mentioned, benchmark testing by DeepSeek found that the R1 model was capable of matching the performance of OpenAI’s o1, at a fraction of the cost and with fewer resources to hand. Benchmarking can only demonstrate so much, but it does give a good indication of the potential of DeepSeek’s R1 model.
Training Methods: R1 vs. o1
Comparing it again to OpenAI’s o1, both are chain-of-thought AI models which look to mimic human reasoning by breaking down problems step by step and then generating a response. But the differences between DeepSeek’s R1 and OpenAI’s o1 come down to how they were trained. R1 differs from o1 in its use of direct reinforcement learning, instead of supervised fine-tuning (SFT), which is more widely used. Developers use SFT to train AI models by showing them curated data, such as example problems, and then explaining how they should be solved, step-by-step to develop “reasoning” abilities. Meanwhile, direct reinforcement learning is when AI models are given problems with no solutions which they must then solve by testing various methods then learning or reinforcing behaviour to find the correct solution. This allowed developers to create an AI model – with limited SFT data – which could develop complex problem solving abilities. During training, researchers even observed the model during, what they call, an “aha moment” where it would learn to revise its thinking midway through solving a problem.
Accessibility and Efficiency
In addition, unlike other AI models, you can very easily access R1’s chain-of-thought processes – something which DeepSeek’s rivals such as Meta make difficult to access, or in the case of OpenAI, hide altogether. DeepSeek is also able to punch above its weight in terms of performance as it uses a “mixture-of-experts” system, breaking down one large model into a number of smaller submodels – or “experts” – which each then specialise in one specific type of data. This means that each expert needs less training and less power, which reduces demand overall and allows the model to run more efficiently.
Real-World Usage and Performance
When it comes to real world usage of DeepSeek versus OpenAI, so far there doesn’t appear to be one overall winner. In efficiency, DeepSeek beat OpenAI’s o1 model. In everything else? Time will tell, as user reports are already finding areas where the model performed worse in comparison to its rivals. DeepSeek’s model is innovative and competitive but it’s not quite an overall market leader, yet.
The Impact of Open Source
But DeepSeek’s recent innovations will give AI companies plenty to think about when it comes to their own AI models, and how money should be spent. Due to the open source nature of DeepSeek’s AI, developers around the world will be picking the model apart. This means that many of the Chinese company’s groundbreaking discoveries will soon be used to inform the approach of other firms which are developing their own AI.
Censorship Concerns
As it is a Chinese-developed service, there is a certain level of censorship which users have been encountering while using DeepSeek’s AI chatbot. Writing for The New York Times, China-based reporter Vivian Wang shared her experience using the chatbot. When engaging the chatbot in conversation about controversial or taboo subjects in China, she found that oftentimes the AI would “reason” itself through a thoughtful answer but when it finished, the reply would instead disappear. A message along the lines of: “Sorry, that’s beyond my current scope. Let’s talk about something else,” would appear or it would cut itself off before an answer was finished. Users outside of China have also experienced similar situations regarding the AI cutting itself off mid-answer or even disappearing replies.
Popularity and Caution
Since the launch of DeepSeek’s AI assistant, it has quickly risen to the top of the App Stores around the world ahead of more established rivals ChatGPT and Google Gemini. But experts do still warn users to approach the service with caution, particularly regarding misinformation, personal data and privacy information, as well as issues surrounding censorship.
Data Collection and Privacy Concerns
According to DeepSeek’s privacy policy, it will collect data from users including email addresses, phone numbers and the user’s date of birth, more standard details as well as user inputs – both text and audio – chat histories, and "technical information” which can be anything from the model of your phone and its operating system to your IP address and “keystroke patterns”. Data will then be stored in "in secure servers" in China, though it may be shared with others according to the policy, which also states that your data will be kept "for as long as necessary". Another area of concern within the policy also references personal identifiers being used to “help match you and your actions outside of the Service”. If you are using any AI chatbot – whether it’s ChatGPT or DeepSeek – users should avoid sharing any sensitive or private information because there is no real way of knowing where it may end up. Speaking on DeepSeek, Dr Richard Whittle from the University of Salford raised various concerns about data and privacy regarding the AI service, however also highlighted that “there are plenty of concerns around privacy with the existing US dominant models too”.
Investigation into Data Usage
There could be trouble on the horizon for DeepSeek, with OpenAI and Microsoft beginning an investigation into whether the Chinese firm has used data from the ChatGPT maker to build its own AI model. The American firms suspect that DeepSeek has integrated OpenAI’s own API into their new AI model, in a process called “distillation”. Distillation is a common technique among developers training AI as it allows them to extract data from larger more capable models to improve efficiency and reduce cost. Developers can integrate OpenAI’s work into their own applications. What they are not allowed to do, and what OpenAI say they have found evidence of, is DeepSeek distilling the outputs to build a rival AI model. It goes against OpenAI’s terms of service, with sources at the company telling outlets such as Bloomberg that Microsoft researchers detected large amounts of data being withdrawn through developer accounts in late 2024 which they believe is linked to DeepSeek. Speaking to the Financial Times, OpenAI have also claimed to have found evidence linking DeepSeek to the use of distillation, though they have yet to reveal any details.
Irony and Response
However, the situation does hold a certain amount of irony. OpenAI has previously come under scrutiny for using massive amounts of data and text which has been scraped from the internet - including copyrighted material - to train their AI models.
David Sacks, an advisor on AI and cryptocurrency, told Fox News: "There's substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI's models. I think one of the things you're going to see over the next few months is our leading AI companies taking steps to try and prevent distillation... That would definitely slow down some of these copycat models."
In a statement on the situation, a spokesperson for OpenAI has stated that they are taking “aggressive, proactive countermeasures” and that they will continue to work closely with the US government to protect models being built in the country.