European semiconductor shares with artificial intelligence exposure—including ASML Holding ASML, ASM International ASM, BE Semiconductor BESI, and Arm Holdings ARM—dropped in the 7%-10% range on Jan.27 following news about Chinese company DeepSeek’s AI models.
Questions have been raised about the true cost of AI model training and inference. We are maintaining our fair value estimates for these firms, as this development could have mixed effects, but our base case still assumes healthy AI demand in the long term.
Until now, the main market hypothesis for the AI supply chain tied increased spending to better AI model performance. Tech firms like Microsoft MSFT, Alphabet GOOGL, Amazon AMZN, and Meta Platforms META have collectively invested hundreds of billions to purchase GPUs from Nvidia NVDA to satisfy the insatiable demand for AI applications.
DeepSeek is a challenge to this hypothesis, as its V3 model rivals GPT-4o and Llama 3.1 across most benchmarks, and it was reportedly trained on a $5.5 million budget, significantly lower than competitors. However, that number only includes the cost of GPU training and excludes other overheads, making comparisons difficult and raising questions about the truthfulness of DeepSeek’s claims.
Semiconductor Stocks - Jan.27 Performance
DeepSeek: The Bearish Scenario
If a bearish scenario materializes and hyperscalers cut their capital expenditures, this would flow upstream in the semiconductor supply chain, lowering near-term growth rates around the industry and impacting European semiconductor equipment makers and Arm. The duration of any potential deceleration in growth would also be uncertain.
The fact this breakthrough comes from a Chinese firm in the midst of tighter semiconductor export controls and higher geopolitical tension could be a trigger for the Donald Trump administration to launch newer export controls across the supply chain, affecting firms like ASML.
ASML’s management is already assuming a sharp normalization of its Chinese business in 2025, but a new batch of export controls would result in even lower demand, potentially hurting 2025 guidance. We will remain attentive for any news on potential new export controls, with ASML’s next earnings happening on Jan. 29.
Can OpenAI Learn From DeepSeek?
DeepSeek has showed that there are alternative routes for AI model improvements than just deploying more capital. Its V3 model employs a fundamentally different architecture than most AI models. While OpenAI or Llama models rely on a dense architecture, where the entire neural network is used for every input, DeepSeek uses a mixture of experts architecture, which only activates certain parts of the network (experts) for every given input.
Thanks to this, inferencing cost can be reduced by 10 times or more, depending on the use case. We expect leading AI models like OpenAI will incorporate any useful learnings from DeepSeek into their models soon, in order to reduce unitary costs users for developers.
The author or authors do not own shares in any securities mentioned in this article. Find out about Morningstar's editorial policies.
