“There are things known and there are things unknown and in between are the doors of perception.” — Aldous Huxley
I’m Huxley Westemeier (26’) and welcome to “The Sift,” a weekly opinions column focused on the impacts and implications of new technologies.
______________________________________________________
In September, I wrote an article about OpenAI’s ChatGPT o1 preview model. OpenAI had lofty claims stating that o1 could score 83% (the previous model’s accuracy was 13%) on the challenging qualifying exam for the International Math Olympiad. Four months later, I was expecting to write about o1’s public release or how powerful it has become. OpenAI even announced the successor to o1 dubbed ‘o3 preview’ during its December “12 Days of Shipmas” campaign, but it has yet to go public and there has been no additional information released. A new model that is more accurate, efficient, and significantly cheaper, was released on Jan 20 by Chinese startup DeepSeek, not OpenAI.
In the week since DeepSeek’s public launch, the stock price of hardware giant Nvidia (which produces chips that run AI models) fell by over 16% by Jan, 27.
Why?
Allow me to explain.
It might seem counterintuitive that a more efficient model would lead to a decrease in demand for processors such as Nvidia’s, but it’s important to consider the unique properties of DeepSeek’s R1 model. It’s entirely open source. Anyone can go to DeepSeek’s website, find the resources necessary to download the R1 and other models and run them locally on their own devices. The company even provides links to distilled models- models that have been simplified or edited in order to run on a lower-powered device like a laptop, which doesn’t have enough power or storage space to effectively run the same models as a massive server farm.
According to analysis by The New York Times, most of the world’s companies that create large AI-chatbot models (Meta, OpenAI, Microsoft, etc) use over 16,000 chips from Nvidia during the processing phase and cost nearly a hundred million dollars in processing power to produce. DeepSeek required only 2,000 chips and cost six million dollars in processing power.
DeepSeek essentially cut the cost to develop a model by a factor of 10 while retaining the same (and sometimes higher) level of accuracy compared to models like ChatGPT o1 while reducing the reliance on chip manufacturers. This provides an enormous benefit to consumers. Cheaper to develop models will end up being more cost effective for users, but it does take business away from companies like Nvidia.
DeepSeek’s R1 model proves that AI models don’t have to be computationally expensive in order to reach high levels of performance, potentially reducing energy costs and reliance on massive cooling systems that have detrimental climate impacts. It also means that U.S.-based companies like OpenAI and Meta aren’t at the forefront of AI anymore. Foreign companies have caught up, and they’re doing so more efficiently and economically. It’ll be interesting to track how the U.S. responds to DeepSeek since it’s a Chinese company, and whether upcoming tariffs announced by the Trump administration will be imposed in a way to keep U.S. artificial intelligence companies competitive.
For now, Meta and OpenAI have been dethroned and DeepSeek reigns supreme.