Key Takeaways
Since its release on Monday, Jan. 20, DeepSeek R1 has upended assumptions about American technological supremacy, casting doubt over the United States’ continued AI leadership and rattling global markets.
The Chinese AI model has achieved a level of performance that competes with the best American tech. But crucially, DeepSeek R1 requires just a fraction of the resources to run.
After years of American sanctions that limit Chinese firms’ access to the high-end GPUs used in machine learning, DeepSeek’s AI breakthrough proves that necessity is the mother of innovation.
The more efficient use of processing power helped the Chinese startup overcome U.S. export controls and deliver impressive performance despite hardware limitations. But it also makes R1 especially attractive to users looking to cut costs.
Access to the model via DeepSeek’s API costs $0.55 per million input tokens, $0.14 for cached inputs, and $2.19 per million output tokens. Meanwhile, OpenAI o1 costs $15 per million input tokens, $7.50 for cached inputs, and $60 per million output tokens.
In a further blow to the established AI hierarchy, DeepSeek has released a series of smaller, distilled versions of the new model under open-source licenses.
To be clear, R1 doesn’t quite match o1’s performance on key benchmarks. But it comes close enough to beg the question: is OpenAI’s razor-thin performance margin worth paying 30 times more for?
While AI efficiency can be thought of in terms of cost per token, the underlying factor that determines price is hardware usage, i.e., how many FLOPS are needed for a given output.
Until now, the assumption has been that demand for AI and demand for GPU compute will increase more or less linearly.
This assumption has been a major growth driver for Nvidia, which currently has a near-monopoly on the most powerful GPUs favored by data centers.
However, if DeepSeek ushers in a new era of dramatically more efficient AI models, demand for GPU capacity may not grow as rapidly as anticipated.
After a week in which DeepSeek dominated headlines and attracted the attention of major AI customers, Nvidia stock dropped sharply in premarket trading on Monday.
Ultimately, the latest Chinese AI model is unlikely to seriously dent Nvidia’s sales on its own, but it does underscore the sector’s potential for disruption.
It also raises questions about the effectiveness of American export controls, which may have lit a fire under Chinese innovation.
The same forces that drove DeepSeek to develop more efficient AI models are also inspiring breakthroughs in hardware and firmware.
While these aren’t going to render Nvidia GPUs obsolete overnight, they could dramatically alter the course of AI development and infrastructure investment in the long run.