DeepSeek is not a game changer. Technology alone doesn’t make you a winner (picture: Shifter.no)
(This article was published first in Norwegian on Shfter.no)
On Monday morning 27.1.2025, the world woke up to what seemed like an AI-Ragnarok. Panic spread among investors, who dumped anything remotely related to artificial intelligence as if it were toxic. DeepSeek had taken the stage and promised performance on par with OpenAI, Meta’s Llama, Google’s Gemini, and all the other American generative AI models.
At the same time, it was rather striking how calmly the top executives of these companies reacted to the chaos. Even The Donald behaved surprisingly civilly, without theatrical threats or announcements of dramatic new measures.
And they were right.
At the time of writing, Microsoft and Google’s market values are roughly where they were before the crisis. Meta is near its all-time high. Nvidia (which was hit the hardest) is only six percent cheaper than it was a month ago.
All this for good reason.
Firstly, many of the techniques used by DeepSeek’s language and reasoning model are already well known. The big breakthrough, “Pure Reinforcement Learning,” which is supposed to allow the model to train itself and make it much cheaper to build, can also be copied relatively quickly. OpenAI has already begun experimenting with such models. In December, Meta launched Llama 3.3., a language model that also significantly reduces the need for computing power compared to previous models.
As expected, AI models are becoming increasingly simpler, cheaper, and more accessible.
Secondly, language is just one of many aspects of generative AI. The next milestone is realistic image generation, video, robotics, and emotion interpretation. The demand for computing power remains unstoppable. Nvidia’s AI chips, which are considered the most cost-effective in terms of computing power vs. price and the most energy-efficient AI chips, are still essential. New advanced applications based on artificial intelligence will emerge, and they too will require more and more efficient use of computing power. Nvidia’s CEO, Jensen Huang, can sleep well at night.
Thirdly, DeepSeek has its weaknesses. The potential surveillance risks from Chinese authorities are naturally problematic. From a hardware perspective, DeepSeek and other Chinese competitors face significant challenges once their stockpile of Nvidia chips is depleted and cannot be replaced due to the U.S. export ban on chips to China. Moreover, DeepSeek is still a research project. Whether its open-source model or organization will be able to scale is an open question.
Nevertheless, this development has consequences and will force the AI industry to split into three main segments.
On one side, we will find companies like Anthropic, OpenAI/Microsoft, Google, and Amazon, which will develop AI models with multiple applications for ever-larger markets. These companies will have to develop business models and user-friendly interfaces in fierce competition with each other, often delivered through their cloud services, which customers are already dependent on. Only a handful of these players will survive, but they will deliver solid returns to shareholders.
Another group of players will develop AI as a tool to support their core business. These companies will subsidize development with their cash flows and make their models open source and cheap to use to accelerate progress. This group includes companies like Meta and Tesla.
The downside of this strategy is that their models will be harder to adopt for the general public. Their applications will be more limited and tailored to their own corporate needs.
Nevertheless, better and cheaper AI means better products and services for their customers and, hopefully, higher profits and more satisfied shareholders.
Finally, we will see more open-source AI models like DeepSeek emerge, many of them originating from universities and research environments. These will either die quickly, become irrelevant over time, be acquired, or attract investors if they specialize in highly specific problems.
Based on these scenarios, some good opportunities may open up for Norwegian entrepreneurs.
Startups with expertise in generative AI can develop user-friendly interfaces and support services for open-source models. The company Red Hat did exactly that with the Linux operating system and sold to IBM for 284 billion NOK in 2018.
Those looking to incorporate artificial intelligence into their products can also rejoice. For most, the most common AI applications will become increasingly cheaper, better, and more accessible.
A third opportunity is to develop AI models with highly specific applications for highly profitable niches. In such cases, hands-on experience and excellent contacts in the selected niches will be critical. Investors will have laser focus on the team.
The biggest challenge will be finding the right talent and expertise in a country already suffering from an acute shortage of IT specialists. If the situation does not change soon and politicians fail to take these opportunities seriously, our society risks seeing them come to life in another country—or even on another continent.