Why DeepSeek AI has brought down the giants of Silicon Valley

The revolution in artificial intelligence (AI) models by China’s DeepSeek is based on one key factor: they do the same thing, but for less. They generally perform just as well as Silicon Valley’s market leaders, but they are much less expensive to create and use. OpenAI, Anthropic, and Google, the creators of the most famous models, and Nvidia, the company behind the sophisticated chips used by these companies, have seen their apparent advantage collapse in just a few days. Fewer experts today believe that training and using an AI model requires the power of a tech giant. Nvidia’s monumental stock market crash is an illustration of that belief.

DeepSeek’s chatbots work the same as any other. Users register with an email and password on the company’s website and can start asking it whatever they want. It sometimes takes a little longer to respond, it doesn’t always answer sensitive political questions about Tiananmen Square or Taiwan, and it works pretty well in Spanish. The best thing, however, is that it’s free. DeepSeek’s business model is based on charging users who require professional applications.

The Chinese company’s latest model is R1, which caused the collapse of the Western AI edifice. It reasons in the same way as OpenAI’s most advanced creation, o1. But in the middle of the Christmas holidays, the Chinese company launched their V3 model, equivalent to the more common ChatGPT-4, which already included the technical advances that have caused the stir.

DeepSeek has taken training a new model from over $100 million for ChatGPT-4 to less than $6 million. Andrej Karpathy, co-founder of OpenAI, former head of AI at Tesla, and one of the most respected experts in the industry, described that budget as “a joke” and added: “You have to ensure that you’re not wasteful with what you have, and this looks like a nice demonstration that there’s still a lot to get through with both data and algorithms.” DeepSeek’s latest model is so efficient that it required a tenth of the computing power of Meta’s comparable model.

These AI models are extremely expensive to train and use. DeepSeek has managed to reduce costs in both cases. To train them, it has made its model learn from previous ones. This is something that the terms of service of these companies do not allow, but it is very difficult to detect and avoid. Just as the original OpenAI models took all the data from the internet without paying for it, now a Chinese model uses closed models at a minimum cost. This is achieved with a technique called distillation, which allows it to ask questions of the previous model and save the answers, thus learning and training itself. This technique implies that the models can learn from each other, which throws us into another scenario of even faster and unknown evolution.

The other technique DeepSeek uses to cut costs is a system called “experts.” Every time a user asks a question, the model decides whether to activate its medical expert, translator, lawyer, or physicist. Classic models activate all of them at once, which wastes energy and computing power. It’s like needing a huge power plant of your own to run an elaborate assembly line. DeepSeek has done the same thing by connecting its small factory to a power outlet and a handful of generators.

Suddenly, users who have been paying around $20 to use OpenAI models can now get almost the same results for free. The expansion of AI is reaching its commoditization stage faster than expected, when new technology becomes a common product accessible to everyone.

All this has also taken place in a company in Hangzhou, a city near Shanghai. The Biden administration banned the export of the newest chips to limit Chinese competition. The Asian giant’s response has been simple: we don’t need them. And they have achieved this with open source, showing the guts of their achievements with documentation published on the internet. Making a virtue out of necessity.

The end of this early stage of AI appears to be bad news for the pioneering giants, who are witnessing a good part of their advantage disappear. But the other big companies, such as Amazon or Apple, which were not at the forefront of the AI race, can clearly benefit and offer the same benefits as the rest at a fraction of the cost.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition