A week ago, at midnight, I received a call from a close friend who was panicking and urging me to sell all of my Nvidia (NASDAQ: NVDA) stock by 10 a.m. the next day. Naturally, I too began to panic. Did they know something I didn’t? Was Nvidia doomed from the start? After reasoning with them and calming them enough to get a coherent explanation, I hung up the phone and sold none of my stock. I don’t trade on a whim at 1 a.m.
This frenzy and the corresponding 17% Nvidia stock price drop are a reaction to moves made by DeepSeek, a Chinese artificial intelligence start up that rivals OpenAI. Specifically, it is a reaction to the release of DeepSeek’s open-source AI large language model and their subsequent research claim that the model training only cost six million dollars using Nvidia’s chips.
While at first glance this seems like a non-issue, the problem for Nvidia investors lies in the type of chips that DeepSeek used. DeepSeek’s new model was created using Nvidia’s reduced capability H800 chip, proving to developers that ChatGPT-like models can be produced without the expensive and higher-powered chips previously thought necessary. This could mean a significant reduction in profit for Nvidia due to a drop in demand for its more expensive chips. If AI developers don’t need to pay extra for a more powerful chip, why would they?
Despite the impending sense of doom for Nvidia fans, some analysts stay optimistic about America’s chip darling. I’m with them. I don’t think that this is the end of the fairytale for Nvidia, other chip makers or the American AI market. There is definitely more to the story.
The first red flag in the DeepSeek case comes from the 6-million-dollar claim. While 6 million dollars sounds cheap compared to the $78 million spent on training ChatGPT-4, it also sounds too good to be true. That is because this number primarily considers pre-training costs, neglecting the research and development, infrastructure and hardware costs that went into DeepSeek’s model. An independent research study by SemiAnalysis claims that the Capital Expenditure for DeepSeek could be up to $1.6 billion, a stark difference from the six-million-dollar cost that has been dominating the headlines. For this reason, I don’t think it is right to jump to the conclusion that DeepSeek has completed some feat of engineering that will upend the chip market.
Aside from the price, it is important to recognize the technological impact of using the H800 chip versus Nvidia’s higher powered offerings. While similar to the powerful H100 chip, the H800 chip was designed specifically to skirt export regulations in the Chinese market, and it performs less efficiently than the H100. While DeepSeek might have been able to work around the limitations for its first model, the reduced capability chip might present problems when it comes time for scaling and advancements.
On the business side, there is some doubt around the likelihood of AI developers and businesses hopping on the DeepSeek-cheap-chip bandwagon. Even if DeepSeek can continue to scale and advance using the H800 chips, the H100 chips will always be more powerful and efficient. If Nvidia’s biggest customers like Meta Platforms and Tesla want to keep improving to edge out the competition, I don’t expect that they will be willing to level down chip-wise. Realistically, these big revenue drivers for Nvidia probably won’t be going anywhere or changing their methods any time soon. They have the money to spend on better chips.
On top of this, DeepSeek’s accomplishment with the cheaper and simpler chip might actually drive revenues up, rather than down. Considering that the H800 chip is less expensive despite similar capabilities, this new knowledge could actually drive-up demand for Nvidia’s cheaper chips, encouraging other players to try out the new method and increase sales overall.
In a particularly volatile industry with constantly changing technology, it is easy to get caught up in the headlines and make impulsive decisions. In this case, though, there is so much more to consider than just what you see at first glance. DeepSeek’s announcement is far from a eulogy for Nvidia. Rather, this event is another twist in the developing story of artificial intelligence. Who knows where it’ll go? I don’t. But I do believe that Nvidia is safe for now, and despite what the midnight phone call might suggest, it certainly isn’t necessary to abandon all hope for what once was a promising company.