DeepSeek - Shaking Up the AI Marketplace Without Redefining AI

All eyes are on DeepSeek, the emerging AI star from China. But how does DeepSeek revolutionize the world of artificial intelligence? The short answer is that it doesn’t. It does not move the frontier of what AI can do for humanity or companies. It is unlike the moment when ChatGPT emerged and made large language models (LLMs) usable for everyone inside and outside of IT. From an AI technology perspective, DeepSeek’s LLM resembles many existing LLMs on the market. There are other reasons for all the buzzes.

First, DeepSeek has trained an LLM with significantly less effort in terms of energy, hardware, and costs. They demonstrated to the world that you can enter the LLM market and attack the big players. And, if you can provide the same service but have only a fraction of the costs, you either enjoy much higher profits than your competitors or drive them out of business (or both). Secondly, DeepSeek is a Chinese startup. Thus, China challenges the long-standing US dominance in AI, making it a new "Sputnik moment".

In contrast to the big change in the global scene, the impact on individuals and companies – in contrast to the emergence of ChatGPT – is limited. DeepSeek may lower the “costs per token” for AI usage—essentially, what customers pay for licensing and using LLMs. It does not allow new ways of using AI and does not impact the effort to make LLMs functional in companies. For example, if developing an FAQ agent and training it with specific company data took up to now 50 person days, DeepSeek’s advancements do not change this number (though running the agent might become cheaper for high-volume use cases somewhere in the future). Plus, most companies do not pay so much for LLMs per se but for services that incorporate LLMs. Think of all the Copilots: packaged, ready-to-use productivity tools that combine LLMs with vendor-specific data. This service pricing might not be under pressure if cheaper LLMs emerge.

To conclude: DeepSeek’s economic disruption mainly impacts the "AI supply chain," i.e., companies like Nvidia. Let us assume DeepSeek’s approach reduces training effort by 90% or even only by 50% (there is an ongoing dispute about the actual numbers). Then, the world might need significantly fewer chips, energy, and data centers than today (though IT has always found a way to use free computing capacities). So, while DeepSeek might not be a transformative disruption like ChatGPT, it might present an economic challenge for AI suppliers and, indeed, challenge the US AI dominance.

?

?