DeepSeek
For many, DeepSeek is a newcomer to the AI market. Originating from China, it is shaking up an industry previously thought to be dominated by established players.
This week, DeepSeek—a Chinese spinoff launching in 2023—and its LLM DeepSeek-R1 entered the big leagues. The global release of its model prompted an avalanche of reviews, commented tests, and, most importantly, benchmarks. While I’m always wary of benchmarks that don’t always represent real-world use cases, they do help establish a trend. And these show that DeepSeek-R1 is an equal, if not better, competitor to current market leaders such as OpenAI, Google, Meta, Anthropic, and others.
In less than a week, this app has become the most downloaded app on Apple’s App Store. It feels like we’re reliving what happened when ChatGPT was first released to the public in 2023.
But, you might say, it’s just another LLM, and a good one at that. But why make a big deal out of it?
Well, there are several reasons to be interested in this product.
First , it uses an open-source model. Yes, the LLM is available to any developer who would like to use it in their applications. Anyone can download a copy and run it on the infrastructure of their choice. DeepSeek also built its own model using open-source solutions (like Meta’s Llama and PyTorch). “They had new ideas and built them on other people’s work. Since their work is published and open source, everyone can benefit from it.” (Yann Le Cun, Chief AI Scientist at Meta)
Second , the development cost. DeepSeek claims it only cost them $5.6 million. This is a shock when big AI players spend tens, even hundreds of millions of dollars to train a single LLM.
Third , US export restrictions on high-tech products did not affect the efficiency of LLM development. In fact, DeepSeek only had time to acquire between 10,000 and 50,000 nVidia GPUs before the US government banned their sale to China. This is a negligible amount compared to labs like OpenAI, Google, or Anthropic, which operate with 500,000 GPUs or more. In other words, geopolitical constraints did not hinder Chinese innovation. On the contrary, it seems to have been able to bounce back from this lack of cutting-edge hardware and manage to deliver a solution that even performs better than its competitors. This reduction in hardware requirements also translates into lower power consumption for the DeepSeek-R1 LLM, making it one of the most energy-efficient GenAI solutions on the market today.
The market and the political world took note. NVidia’s stock, along with those of AI leaders, plummeted on the NASDAQ (-17% for NVidia). Trump himself stepped up to urge his compatriots to wake up and adapt. American tech giants got the message. Even Sam Altman, who had made a positive comment about DeepSeek-R1, got it.
And for Europe, it’s also a positive signal. With limited resources (frankly, will Europe ever be able to launch a €500 billion program like the American Stargate?) and access to technology controlled by our American “allies,” European tech companies can afford to dream of (finally) eclipsing their American competitors. All they have to do is follow the same paths as those paved by DeepSeek.
This kind of solution with a lower energy footprint is also a good sign that in the future our consumption could be reduced and, why not, even AI could become sustainable!
Now, let’s not get carried away either. This is still a 100% Chinese product with its own political and cultural biases, imposed by the government. Don’t go asking DeepSeek-R1 about Tibet, the Uyghurs, or the Tiananmen Square uprising.
Before it suffers the same fate as TikTok or Huawei, take advantage of it now before it’s gone!
