There is no doubt, the event of these days is the shocking announcement of the launch of DeepSeek, the low-cost Chinese competitor to OpenAI, if only for the turmoil created in the stock markets, dragged down by NVIDIA’s sudden drop. Much has been written about the event, including very detailed technical specifications: what interests us is to focus on the key points of this affair to support that, as often happens, there may have been some exaggerations.
To explain the meaning of this story, let’s use an example. We are all used to ordering a list of information according to a criterion, for example, arranging a list of names in alphabetical order on an Excel sheet. There are algorithms that perform this simple operation. In its earliest version, the sorting algorithm required approximately 
elementary operations (or flops), where 
is the number of objects to be sorted. Therefore, to sort a thousand names, the algorithm required about 1 million operations. This algorithm was soon refined, reducing the number of operations to about 
, requiring only 10,000 operations to sort 1,000 names alphabetically. This is a typical innovation in computational science, and artificial intelligence is no exception.
So, what happened with DeepSeek? Fundamentally, three things:
1) A replication of OpenAI’s capabilities at much lower costs;
2) It was released under an MIT license and is therefore open source;
3) It provides APIs that make it very easy to use, as well as free.
The first point is the most interesting for us because it allows us to clarify some aspects. On social media, the information circulated that DeepSeek essentially emulated OpenAI at a cost of only 5 million dollars. In reality, this estimate is based on the training of version V3, which uses the Mixture-of-Experts model, inherently requiring less computational power. But beyond being an estimate, nothing is said about the learning costs of R1, which is the release we are interested in. Furthermore, it seems – but we will have more precise information in the coming days or months – that the small Chinese startup extensively used (and it is still being evaluated whether correctly) OpenAI’s APIs to “distill” the knowledge of the American platform. In practice, it leveraged OpenAI’s knowledge and computational methods as a “teacher” to help the “student” DeepSeek learn. This process, which is still under examination, would explain very well why the Chinese platform was able to replicate OpenAI’s performance at significantly lower costs.
This does not detract from the fact that, in many aspects – from the intelligent use of long-known Machine Learning models to the optimization of memory usage – DeepSeek represents a technological innovation comparable in nature to the sorting algorithm described earlier. Therefore, it is fair and correct to speak of actual innovation, which has been made available to everyone for free. These two aspects are truly disruptive.
That said, we believe that the market reaction was marked by hysteria, which partially subsided once new information arrived and reflection took over. Certainly, the use of distillation models (still not fully understood in all their technical aspects) presents us with the classic case where America once again innovated, and China replicated (albeit in an innovative way). And Europe? Well, as always, it stays on the sidelines, thinking about how to regulate cases like this.
Disclaimer: This article expresses the personal opinion of the contributors of Custodia Wealth Management who drafted it. It is not intended as investment advice, personalized consultancy, or an invitation to engage in financial instrument transactions.
