27 January 2025
If you are covering DeepSeek AI's success and its impact on tech stocks, please see the following comment from Ben Barringer, technology analyst at Quilter Cheviot:
“Since its launch on 20 January, DeepSeek’s AI app has climbed to the top of the download charts and appears to be delivering performance on a par with the latest version of OpenAI’s ChatGPT.
“DeepSeek claims to have delivered this performance using somewhat dated technology, however there are questions around whether this is really the case given the stringent Chinese rules around what technology can be used when training AI models.
“It has also been reported that DeepSeek spent $6 million on this training, which is just a fraction of the cost of training a conventional model. However, it appears to be building on the results that have already been developed by other models and seems to utilise Meta’s open-source AI model, Llama 3.1. Given DeepSeek has not kicked off from a standing start, there is an element of uncertainty around the validity of comparisons. Nonetheless, there is no denying that it produces good results, and it has been developed to use the latest techniques such as chain of thought.
“DeepSeek’s success will serve to intensify the US/China AI war, particularly following the recent announcement of the Stargate project in the US. It also provides a wake up call and somewhat of a question mark on how much needs to be spent in order to build a model, and whether quite the level of CapEx we have been seeing is really required. However, bringing the price of model training down is no bad thing as it will help to lower the entry point and this price elasticity could help drive usage and volume.”
