DeepSeek Disruption

The stock market was rocked over the past weekend by the announcement of a new artificial intelligence system called DeepSeek, which threatens the business models of the entire U.S. AI industry. The chipmaker Nvidia saw its stock price fall 17% on Monday, and analysts immediately questioned the value of billion-dollar-plus investments in AI by OpenAI, Softbank and Oracle.

What is DeepSeek and why did it send shockwaves through the AI community?

The short answer is that the Chinese engineers who founded DeepSeek arrived at a more efficient way to manipulate the fantastic amount of information and computer processing power necessary for AI to function. Instead of using one massive system trying to know everything about everything, DeepSeek broke up its analytical engine into AI specialists—known to techies as ‘inference time computing.’ While traditional models like GPT-4 use 1.8 trillion calculations at every moment, answering every query and request all at once, DeepSeek only has to expend 37 billion at a time, each focused on a specific set of queries. The specialists can be trained more far more cheaply, they are twice as fast and use 75% less memory.

Put in dollar terms, DeepSeek has 200 employees, vs. 4,500 at OpenAI. DeepSeek was developed for roughly $5 million, compared with the $6.6 billion capital raise for OpenAI.

DeepSeek is open-source, which means anyone can build on it—and it will rapidly take market share away from the clumsier AI systems in today’s market. But it also uses a model that others can copy—breaking down different types of projects, queries and models into smaller, more nimble AI specialists.

The selloff of Nvidia came when tech experts (and analysts who talk with them) realized that AI services can be offered without the need for massive raw computing power—the kind of chips that Nvidia specializes in.

However, DeepSeek does use Nvidia chips as part of its server array. And there is no reason to imagine that the international AI cohort of developers can’t pivot to the new model and gain new efficiencies that nobody had thought of before. Already, Alibaba has announced a similar model; its 2.5-Max AI model, with an array of specialists, claims to be faster and more efficient than DeepSeek.

DeepSeek is not the end of the AI boom, but did manage to prick what some saw as an AI bubble.

Sources:

https://www.bbc.com/news/articles/c5yv5976z9po

https://www.bbc.com/news/articles/c5yv5976z9po

https://www.cbsnews.com/news/what-is-deepseek-ai-china-stock-nvidia-nvda-

https://www.computerworld.com/article/38130