DeepSeek’s AI disruption – what it means for Investors

Large learning models (LLMs) are a type of AI program that can recognise and generate text. An LLM is essentially a computer program that uses machine learning, a practice that involves feeding a program an enormous volume of data that enables it to recognise and interpret human languages or other types of complex data without human intervention [i].

LLM’s are currently being used in the following broad functions:

• Sentiment analysis
• DNA research
• Customer service (call centres)
• Chatbots
• Online searches (as seen with Chat GPT)

Despite the fanfare surrounding AI disruption, the use cases for LLM models within business operations remain low, primarily due to their early-stage development. However, a shift by developers to create smaller, task-centric reasoning models is expected to rapidly expand their implementation over the next 5 years.

On 20th January 2025, a relatively unknown Chinese AI software developer called DeepSeek released their latest AI reasoning model called R1. It quickly became the most downloaded app on the Apple Store, after independent assessments of R1’s capabilities showed results like the much larger LLM models like Chat-GPT4. [ii]

Within a week of DeepSeek’s R1 launch, the US Nasdaq market fell by 3.5% on January 28th, while some of the world’s leading technology hardware companies, including Nvidia, Broadcom, SK Hynix and ARM Holdings, fell by 10-20% in a single trading session.

 What changed overnight to trigger this sharp selloff?

The selloff of technology hardware companies, like NVIDIA, Broadcom and ARM, resulted from information contained in DeepSeek’s R1 model white paper, released on 20th January. In this report [iii], DeepSeek stated that R1 had the following capabilities;

Lower Training Cost

• Deep Seek spent just US$6m training their R1 model, a fraction of the more than $100m OpenAI CEO Sam Altman estimated was spent on their ChatGPT-4 model.

Lower Processing Requirements

• R1 required approximately 1/10th of the processing hardware compared to Chat-GPT (using 2,048 Nvidia GPUs for training, compared with the more than 20,000 used to build Chat-GPT.)

Equivalent Performance to Market Leaders

• R1’s reasoning capabilities were like Open AI’s Chat GPT in both language and mathematics.

This white paper suggested that DeepSeek had reinvented AI training at a substantially lower cost, both in terms of computer hardware requirements and model development [iv][v].

Did this mean that demand for specialised chips like NVIDIA GPUs would be substantially lower than current expectations? Let us come back to this question later. 

Who is DeepSeek?

DeepSeek was started in 2021 by High-Flyer, a mainland Chinese hedge fund that was an early adopter of AI in its trading strategies. Over the past 4 years, High-Flyer has invested US$1.6bn, acquiring up to 60,000 NVIDIA GPU chips and other computer hardware, with an additional US$1bn of operating costs incurred over this period [vi].

Now that the dust has settled after the market collapse on January 28th, substantial evidence has emerged that raises serious doubts about the validity of DeepSeek’s claimed achievements.

It turns out the $6m R1 development cost does not include the substantial operating and computer hardware expense requirements in its calculation. The figure appears to be just related to direct model training costs.

Furthermore, given R1’s performance relative to OpenAI and High-Flyer’s capabilities (60,000 GPUs), it also seems likely that they used a much larger number of NVIDIA GPUs to train R1, refuting their lower cost and lower hardware requirement claims.

If we consider performance, DeepSeek’s R1 achieving comparable reasoning outcomes as the much larger OpenAI model (GPT-4 and o-1) is most likely a function of the rapid improvements being achieved with newer AI reasoning algorithms undergoing model training. Dario Amodei, the CEO of AI research company Anthropic that built Claude AI (a competing LLM model), has suggested that algorithm improvements can deliver 10-fold annual improvements in efficiency (90% reduction in compute capacity each year for the same outcome) while AI is still in its infancy [vii].

With DeepSeek released 5 months later than Open AI’s latest model (o1), its ability to compete with the more established models is most likely explained by the algorithm efficiency gains outlined by Dario Amodei.

This is supported by the results of “Humanity’s Last Exam”, a newly created test designed by an international team of AI experts to create a standard benchmark to measure AI model reasoning performance. OpenAI’s latest model, o3-mini (launched February 5th) achieved the highest score of 13% versus 9.4% for DeepSeek R1. OpenAI’s ‘Deep Research’ tool scored an even more impressive 26.6% (though as the tool has functions not included by the other model, this is a less relevant comparison) [viii].