DeepSeek rattles markets – is it overblown or concerning? A short analysis

A Chinese AI start-up called DeepSeek generated a lot of noise over the weekend after mainstream news outlets picked up on their new ‘R1’ model, which appears to offer equivalent performance to some of the leading large language models at a fraction of the cost. The NASDAQ opened down 3.5%, Microsoft down 4.7% and NVIDIA down 13% We think it is likely that the noise is overblown:

• DeepSeek has not used revolutionary new methods to train its models. The V3 model, which purportedly cost $5m to train, uses a technique known as ‘MOE’ (mixture of experts) to lower the cost. This has already been used by the major US players.
• We do not have cost details for the R1 model, which offers equivalent performance to OpenAI’s leading o1 model. We also know that OpenAI’s new o3 model is significantly better than o1 – so while DeepSeek may have a lead in efficiency, it does not lead in performance.
• Most importantly, efficiency gains are critical to AI development. If these techniques enable further scaling, it should be positive in the long-term as lower cost will drive higher volumes. This is an example of ‘Jevons paradox’ and has been seen many times in the history of computing.

Turning to our stocks, this development is largely positive for big tech in the long-term:

• For Microsoft and Amazon, higher efficiency is likely to drive some combination of lower capex and higher volumes. This should be good news for their cloud businesses.
• Similarly for Apple, if models become smaller it is likely that AI on edge devices (like phones and laptops) becomes more important, and Apple is well placed here.
• Meta is maybe the biggest winner of all as cheaper AI undergirds so many of their initiatives: from better digital advertising to better targeting to more content for its nascent Metaverse offerings.
• Alphabet (Google) is perhaps worst-placed given their investment into dedicated hardware and LLMs. Even here though, higher efficiency is likely to mean lower cost and higher usage.

The setup for NVIDIA and the rest of the semiconductor supply chain is more complex: it is hard to argue with the idea that higher efficiency might lead to lower GPU sales in future. It also appears that DeepSeek circumvented some parts of NVIDIA’s software stack (known as CUDA), which is a key part of what locks clients in. Even here though we note that DeepSeek still used NVIDIA GPUs. We also note that Meta recently raised their capex guide from $50bn to $60-65bn, and OpenAI just announced a $500bn capex initiative called StarGate. This does not seem to indicate a weaker environment for AI-capex in the near-term,  In the longer-term, for NVIDIA et al, the key aspect is Jevons paradox: it is likely that lower cost of compute drives higher adoption of AI and that, in turn, is positive for NVIDIA. As the Bernstein analyst succinctly put it: ‘these guys are trying to build God… they need a lot of compute!’. We continue to monitor developments and remain nimble: we will look to take advantage of any irrational share price movements where we can.

Disclaimer:

This message is provided for information purposes and should not be construed as a solicitation or offer to buy or sell any securities or related financial instruments, nor does the information constitute advice or an expression of our view as to whether a particular financial product is appropriate for you. Please note that we do not provide any tax or legal advice and clients must seek their own tax advice independently.

VAR Capital is an independent financial services firm offering asset management, lending and family office services. It was founded by individuals with extensive experience from Banking, Asset Management and Family Offices. Based in Mayfair, London, VAR Capital Ltd is authorised and regulated by the Financial Conduct Authority (FCA).

Source: VAR Capital

Media Contact: Vikash Gupta, vikash@varcapital.co.uk