China’s DeepSeek launch, a generative AI chatbot (to rival ChatGPT) at a fraction of the cost, has caused turmoil in the financial markets and raised questions in the tech sector about the challenger’s true capabilities.
Daniel Casali, Chief Investment Strategist, at Evelyn Partners, the UK wealth manager, said, “The recent release of China’s DeepSeek Artificial Intelligence (AI) model has caused turmoil in the markets, stirring concerns among AI investors and raising uncertainty about chip demand. This new model, which appears to be cost-effective, could impact the broader AI landscape by making AI models more affordable and efficient. However, it has yet to be fully verified, leaving questions about its true capabilities and comparison with OpenAI’s ChatGPT models.
“In the short term, this has led to share price volatility, especially among companies most exposed to the AI theme. Nvidia fell -17%, Broadcom is down -17.4%, Cadence Design Systems is off -9.5%, and Alphabet, Microsoft and Tesla shares slid -8.4%, -4.2% and -3.2% respectively. We see this as more of a rotation within the market as per our ‘broadening out’ theme, than a risk-off move away from equities, with most stocks in the S&P 500 in positive territory.
“The US stock market remains strong, supported by robust US GDP which is growing at a solid pace. With strong earnings growth from top US companies and continued demand for micro-chips, the AI sector remains a key point of focus amidst the broader economic landscape and shifting geopolitical dynamics. The ongoing US-China tech rivalry is also a factor, as China seeks to challenge US leadership in AI development.
What this means for alternative investments?
Luke Hunter, Investment Manager and Sector Specialist at Evelyn Partners, said, “Chinese AI model, R1, from the company DeepSeek, utilises technologies like chain-of-thought prompting and reinforcement learning, similar to OpenAI’s models, but at a much lower cost—up to 95% cheaper. To reiterate there are question marks over how cheap it is and, when asked, the model sometimes says itself that it uses ChatGPT to operate which could undermine the story, if true.
“The new approach constructs a model which is broadly comparable in overall size to existing models but only utilises a fraction of the parameters of the model at any one time, reducing the compute intensity significantly and bringing costs down. The performance of R1 and GPT’s 01 are similar for coding and maths, but GPT still has a general knowledge advantage.
“In our view, what DeepSeek have demonstrated is completely normal and is a natural evolution of this sort of technology. The drop in AI inference costs (measured $ per million tokens) isn’t surprising, as they have been falling consistently – down 90% in 2024 and though we don’t model these ourselves we expect them to fall further. We do not think that cheaper models will necessarily render frontier models, like OpenAI’s GPT, obsolete. While efficiency gains are valuable, frontier models are still necessary for pushing the boundaries of AI development.
“Ultimately, we do not believe cheaper models will lead to a reduced demand for computing power. Historically, with all technology platforms efficiency gains usually drive greater volume growth, rather than a decline in computing needs. So, we believe that the recent sell-off of AI related stocks like TSMC and Nvidia is based on overblown fears and that the development of AI is proceeding as expected, with both cost-effective models and frontier models able to co-exist. If anything, increased efficiency is likely to be offset by increased adoption.”
Leave a Comment