In most (not all) technologies, the U.S. is perceived as the global leader in tech innovation. Places like Silicon Valley in California and Route 128 in Massachusetts host some of the world’s top companies like Apple, Nvidia and universities like Stanford and MIT.
These innovative tech companies, whether tech giants or newly formed startups, fuel the stock markets like the Nasdaq and the New York Stock Exchange. The historical profit and growth for American tech companies are proven. They comprise the top of the S&P 500, and the top ones are collectively known as the "Magnificent Seven."
Their valuations measured through their price to earnings ratios are quite high, but investors have learned that U.S. tech startups can have parabolic growth rates.
In the field of artificial intelligence, U.S. companies like Open AI, Google, Meta and others have historically led in terms of mindshare. A $500 billion "Stargate" AI infrastructure investment was recently announced by Softbank CEO Masayoshi San, Open AI CEO Sam Altman and Oracle Chairman Larry Ellison and President Donald Trump.
But few days later, a brick was thrown onto the U.S. AI glass castle. DeepSeek, a Chinese AI startup, announced they had met or exceeded several of Open AI’s performance benchmarks, using fewer older processors at a cost of less than $6 million.
Necessity is the mother of invention. Since China is subject to U.S. export controls on advanced semiconductors, which includes Nvidia’s latest AI processors like the Blackwell Graphic Processing Unit, DeepSeek had to find a way to train at the same level as the U.S. but with constrained computing resources. How DeepSeek was able to acquire older Nvidia H100 processors for training despite the U.S. ban on them is still not fully understood, as well as their actual training data. But the important point is that they were able to meet or exceed some American AI performance levels using older and fewer processors.
The Chinese were forced to innovate because of the lack of needed state-of-the-art Nvidia GPUs, which from conventional thinking was needed to match the performance of U.S. AI. According to their conclusion in their DeepSeek published journal paper, “DeepSeek-R1 is more powerful, leveraging cold-start data alongside iterative Reinforcement Learning (RL) fine-tuning. Ultimately, DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on a range of tasks.”
DeepSeek also uses a smaller model as opposed to that used by U.S. AI tech giants, which require more processors. Plus, they have made DeepSeek open source, which avoids the recurring fees that their U.S. competitors charge.
This caused a sudden loss of confidence in the stock market for U.S. AI companies, especially Nvidia, whose parabolic growth depends on the assumption that tech giants, other corporates and joint projects like Stargate will buy their cutting edge GPUs for training several AI models to the point where they can reach higher levels.
After all, if you can reach the same (or better) AI performance metrics using older GPU processors with a more efficient Chinese algorithm, why bother with higher end expensive Nvidia GPU processors? Why spend $500 million when you can spend less than $6 million on an open source AI model?
The immediate impact is that price to earning valuations for U.S. AI appear to have tempered somewhat. It also puts into question how far the U.S. lead is in AI and possibly other technologies. In a way, though, more reasonable price to earning valuations strip away the hype and may ironically attract more investors in the future.
Investors are now asking whether they should allocate more of their capital as they have done in the past to U.S. AI tech high flyers, or look elsewhere and do more diligence on innovations abroad in order not to be blindsided like they did in this case.
