Our expert's take on DeepSeek – AI's Sputnik moment?

Shares in Nvidia and ASML have plummeted on fears that a new AI model from the Chinese startup DeepSeek could disrupt the current AI landscape. DeepSeek's R1 model, which was released on 20 January, appears to perform similarly to the US's top big language models, but was developed at a fraction of the cost. The development also suggests that Chinese AI engineers have found a way to circumvent US-led export bans on advanced semiconductor technology.

Technological advances show that much higher cost efficiencies are possible, but this may not yet lead to a slowdown in investment.

The astonishing performance of DeepSeek's model is mainly due to a combination of architectural optimizations such as a Mixture-of-Experts design and custom communication schemes between chips. This technological advancement shows that much higher cost efficiency is possible in AI development. However, this seems unlikely to cause a slowdown in overall AI investments just yet, as the race towards Artificial General Intelligence (AGI) is in full swing. Meta recently announced a 50 per cent increase in investment in AI projects by 2025, and “Project Stargate” by OpenAI, Oracle and Softbank, revealed at the World Economic Forum, is expected to be worth $500 billion.

Further export tightening likely as sanctions appear ineffective

Despite the US's efforts to stall China's progress in AI through export restrictions, these sanctions appear to be ineffective. DeepSeek states it used Nvidia H800 chips for training, a down-graded version of the H100 for the Chinese market. But it's unclear if H100 chips bought prior to the restrictions were involved, or if the company leased more high-end chips from data centers in South East Asia. Further tightening of export restrictions by the US seems likely.

Investment implications: AI arms race to continue as cost effectiveness increases

The investment implications of these developments are significant. The strong Q4 performance in tech hardware was largely driven by demand for lower-cost custom AI chips, as demonstrated by Amazon. The DeepSeek revelations support the expectation of a continued drive for cost efficiency. With increased scrutiny on the sustainability of large investments in AI infrastructure, there may be an initial hit to hardware stocks. For some, this could be seen as a buying opportunity, as improved cost efficiency should ultimately drive adoption of AI applications and benefit a wider technology supply chain.

In conclusion, Chinese companies' ability to succeed in AI development despite sanctions is supported by the news. Lower training and inference costs should generally support faster adoption and monetization of AI solutions. Investors may begin to see value in diversifying their portfolios beyond US-centric stocks, as Chinese companies, for instance, are making significant progress in sectors such as AI despite regulatory challenges and barriers from the US. This might mark a shift in investor sentiment, with emerging markets once again becoming increasingly competitive and attractive for investment.

Even the smallest undertaking starts with a conversation