On the 10th January, Chinese startup company DeepSeek launched its first AI chatbot, DeepSeek-R1. Western stock markets reacted with panic: almost $1tn was wiped off the Nasdaq Composite index in one day, while Nvidia recorded the largest drop in share price in US history. But, why did the launch of yet another AI chatbot in a field already crowded with products by OpenAI, Google, and Microsoft make waves?

“It was developed at a fraction of the cost: it took $6m to train, compared to over $100m for ChatGPT”

Part of the answer lies in the geopolitical context. The US has tried to prevent China from accessing the most powerful computer chips, in order to curtail China’s growing technology sector. Chinese companies have not yet mastered the technology to manufacture these intricate chips, which have been used extensively in Western AI models such as ChatGPT. Many people thought that these powerful devices were a requirement for any successful AI large-language model (LLM). However, the emergence of DeepSeek blew this assumption out of the water.

DeepSeek is based on the slower Nvidia H800 chipsets which were still allowed to be exported to China, rather than the more powerful H100 chips used in Western LLMs. However, it still performs as well as, or even better than models such as ChatGPT. To put the cherry on the cake, it was developed at a fraction of the cost: it took $6m to train, compared to over $100m for ChatGPT.

“This is a clear point of failure for the US policy”

Both Western governments and companies are worried, but for slightly different reasons. This is a clear point of failure for the US policy of preventing the development of the Chinese high-tech industry. It may even have been counter-productive. Faced with the lack of powerful chipsets, DeepSeek was forced to innovate their model to require less memory use. For the current big hitters in the field such as OpenAI, not only is the prospect of increased competition from Chinese companies worrying, but they have been blindsided by the low training cost for the DeepSeek model.

Although this may sound like a good thing, this does suggest that about 95% of the money spent on training ChatGPT was a waste. Investors will no longer keep funding AI models that cannot justify their exorbitant cost, hence they will be forced to reduce their costs or else get left in the dust. Another concern is that because of this, in a rush to innovate, AI safety will be compromised in favour of developing faster and more powerful models. However, the increased competition in the sector might turn out to be a positive for consumers like us, with lower costs for AI models allowing their benefits to become more accessible.

The future is uncertain. If Western AI makers manage to reduce their costs while still improving performance, this may just be a blip in their control of the market. However, if this proves difficult, we may be leaving the era of Silicon Valley dominance behind.

Want to share your thoughts on this article? Send us a letter to letters@varsity.co.uk or by using this form.