DeepSeek: what’s the Chinese ChatGPT all about?

DeepSeek is a Chinese AI start-up that was launched in 2023. Its model is considered to be as powerful as ChatGPT, yet requires significantly less computing power and development costs.

The model is partly based on an open source approach: a document describes its architectural details and its parameters are accessible to the public. The application has thus sparked a little revolution in the world of AI. Just a few days after its release, DeepSeek dethroned ChatGPT and became the most downloaded application in Apple’s App Store.

Why is DeepSeek important?

DeepSeek challenges the prevailing narrative around artificial intelligence, which has largely been shaped by the big US tech companies. So far, these companies have justified massive investments by using several strategic arguments:

1. AI as a geopolitical challenge: the idea that AI will eventually exceed human intelligence and emerge as a tool of technological dominance prompts immediate and massive investment. In the United States, this manifests in the desire to stay one step ahead of China, perceived as a rival in a race against time for AI development.
2. Computing power as a prerequisite for progress: according to this narrative, a more advanced AI would require increasing amounts of computing power, whereas the ecological impact is viewed as an unavoidable collateral damage.
3. American Big Tech dominance: developing advanced AI models would be impossible without the financial resources, the expertise and the exclusive access to advanced technologies that the American giants have at their disposal.

DeepSeek has challenged these assumptions by developing a model that is as powerful as ChatGPT’s models in a relatively short time and partially releasing its architecture as open source.

Its success is not just a technological breakthrough: it challenges the idea of American AI supremacy and calls into question the rationale of competition driven by proprietary AI models such as ChatGPT.

What is revolutionary about DeepSeek?

DeepSeek shows that by optimising algorithms rather than increasing computing power, it is possible to design equally efficient models at a lower cost, and this from both economic and environmental perspectives.

Furthermore, DeepSeek’s market launch coincided with the US government’s announcement of its StarGate Project, a USD 500 billion investment plan in AI infrastructure. However, according to the available information, DeepSeek’s training incurred a cost of less than USD 6 million (compared to USD 100 million for ChatGPT) and required ten times less computing power.

This does not mean the end of the race for the highest computing power, but rather it will push US companies to invest more in the development of more efficient systems.

So this breakthrough has a direct impact on big American companies.

On the one hand, it calls into question their technology monopoly and the justification for their colossal investments, while also highlighting that these companies have not prioritised the development of more economical algorithms, despite the environmental costs of AI.

On the other hand, it creates a competitive pricing pressure. While OpenAI was considering price increases to cover its costs, DeepSeek offers a much cheaper alternative (20 to 50 times cheaper). This will force OpenAI to rethink its business model or offer models with significantly higher power in order to remain competitive.

How does DeepSeek impact the share prices of big tech companies, such as NVIDIA?

NVIDIA’s stock price dropped 17%, leading to a USD 593 billion market value loss for the world’s AI chip leader. This decline can be attributed to a loss of investor confidence.

Until now, NVIDIA has been regarded as a key player in the development of AI, particularly due to its high-end chips. However, due to export restrictions, these chips are not available in China.

Despite using less advanced chips, DeepSeek has succeeded in training a high-performance model. This suggests that software innovations and the optimisation of algorithms can compensate for limited access to advanced hardware.

This is good news for Switzerland, as it has also been excluded from Washington’s list of allied countries who have unlimited access to chips required for artificial intelligence.

It is therefore conceivable that, if other AI players succeed in developing powerful models with less computing power, this could reduce the demand for NVIDIA’s ultra-powerful chips in the long term.

However, one could also argue that more companies may enter the market if AI development becomes more affordable and chip demand grows accordingly.

All in all, NVIDIA’s plunge reflects the market’s reaction to growing uncertainty, rather than a direct challenge to its dominant position. However, this development clearly shows that the future of AI is not only based on the raw computing power of chips, but also on algorithmic innovation.

What is DeepSeek’s impact on AI?

If our information is accurate, DeepSeek proves that it is possible to achieve performance comparable to the most advanced models while consuming much less energy.

This should encourage US companies to optimise their models instead of constantly increasing computing power, which would have a positive impact on the ecological footprint of AI.

Moreover, by making its model partially accessible, DeepSeek promotes a more open and collaborative approach to AI development. This approach stands in contrast to the proprietary model adopted by companies such as OpenAI, which concentrates power in the hands of a small group of dominant players.

Ultimately, DeepSeek’s success shows that a young start-up with a relatively small team can compete with the US AI giants, challenging their monopoly status and potentially leading to a more diverse and cost-effective offering for users.

And finally, DeepSeek, like ChatGPT’s o1, is a “chain of thoughts” model, i.e. a model that mirrors human reasoning by breaking down problems into smaller steps. This means that although it requires less energy for development (training), it consumes more energy during use (inference energy) than a conventional system.

DeepSeek is a Chinese startup. What are the implications of this?

There are two main aspects to consider: the ideology that the model propagates and the use of user data.

DeepSeek is developed in China, which has implications for what information is classified as “acceptable”. Each model is adapted to ensure as much as possible that no content classified as offensive is generated by the system.

As expected, DeepSeek avoids certain sensitive topics, such as Tiananmen or the Uyghurs. However, rather than providing false information, the system replies that it cannot comment on the subject, an approach that is more transparent.

For less overt historical or geopolitical issues, it is likely that the template reflects a Chinese rather than an American or Western perspective, potentially conveying inaccurate information, whether intentionally or unintentionally.

Having said that, all AI models have biases, including those developed in the United States, where historical and political narratives are also influenced by national perspectives. Since DeepSeek is (partially) open source, users can retrain the model with more diverse data to mitigate certain biases and adjust the system’s prevailing perspective.

Why is it important that DeepSeek is open source?

The fact that DeepSeek is partly open source also helps mitigate certain concerns relating to data security. In fact, it is possible to install and run the model on your own server and thus avoid sharing data with China. However, it is probable that most users will rely on the Chinese servers provided.

It is possible that DeepSeek has opted for open source precisely in order to enable such customisations and thus appease the international community about the use of the model. On the other hand, concerns may arise as to whether the tool, once widely adopted, could be used for propaganda purposes.

Although the same concerns may arise with US models, the risk could appear greater with a Chinese system. Meanwhile, the concerns over data use are similar, whether the systems are Chinese or American.

What do sceptics say about DeepSeek?

It is important to note that, according to some observers, the information shared by DeepSeek (training costs, number of chips, etc.) does not correspond to reality, but rather reflects a marketing campaign to enter the global market.

Furthermore, it appears that DeepSeek may have used ChatGPT to train its own AI, i.e. it has used data produced by OpenAI’s technology.

Ironically, OpenAI expressed concerns that its content had been stolen to train DeepSeek – while OpenAi also admitted that it had used proprietary material without permission to train its own model (see articles from The Guardian or Futurism).