It took about a month for the finance world to understand the significance of DeepSeek, but when it did, it did so by knocking more than half a trillion dollars, around the value of an entire Stargate, off Nvidia’s market cap. Not only was it just Nvidia in the crosshairs: but Tesla, Google, Amazon, and Microsoft also took a hit.

Why DeepSeek’s AI Breakthroughs Are Shaking Up the Industry and Challenging Big Tech’s Dominance

DeepSeek’s pair of AI models released in rapid succession has kept it on a level with the best of American labs, according to Alexandr Wang, the chief executive of Scale AI. Amazingly, DeepSeek managed to train these models at a much lower rate than its U.S. counterparts with one recent model costing only $5.6 million, the cost of an American top AI expert salary. By contrast, other AI models, including OpenAI’s GPT-4, cost companies more than $100 million to train. This shift in cost-efficiency could have massive industry-wide implications.

Within weeks, DeepSeek released two breakthrough models. Never mind that the reasoning model (v3) set the world on fire on Christmas Day; just last week the second of all the genAI models, R1 (released the week prior) was described by the likes of Marc Andreessen as an astounding breakthrough. These developments beg to be questioned, whether tens of billions in computing power are really needed to win the AI race.

Founded in 2023 by Liang Wenfeng, DeepSeek is the top AI startup in China. Liang, who began purchasing Nvidia GPUs in 2021, before U.S. sanctions went into effect, is focused on probing the limits of Artificial General Intelligence, or AGI, in a way that OpenAI cannot. DeepSeek’s approach — building on the work of existing open-source models — meant that it was able to train with a much lower cost than such rivals, with a saving in training efficiency. In fact, it estimates the cost of training R1 at $5.6B, which is an astonishing 95% less than what OpenAI charges per similar models.

The success of DeepSeek calls into question the conventional wisdom that huge investments in computing are necessary for advances in AI. The startup’s meteoric rise threatens to displace big tech firms and suggests that technical innovation, rather than aggressive fundraising, can power competitive AI models. Though skepticism has been voiced about DeepSeek’s promises, the swift reproduction of OpenAI’s advances indicates that even the most hermetically sealed A.I. techniques are not immune to upheaval. And because it’s using open-source models, DeepSeek’s success could be a watershed in how broadly available advanced AI becomes.

This success also underscores the rising competition between China and the U.S. in AI innovation. While the U.S. bets on raw power, DeepSeek’s ability to innovate with optimization and open-source technology shows that China is closing the gap. It also challenges the narrative that only tech giants can lead in AI, suggesting that smaller companies could disrupt the industry with more affordable and efficient approaches. The future of AI could be far more decentralized and accessible than previously thought.