The DeepSeek disruption

DeepSeek released R1 (a Reasoning Large Language Model) on 20 January to little initial fanfare, but over the next seven days discussion spread before culminating in becoming the number one downloaded free app on the App Store and a sharp market reaction on 27 January, with the heavily tech-weighted Nasdaq 100 falling -2.9%.

Who is DeepSeek?
DeepSeek is a Chinese AI startup founded by Liang Wenfeng, a co-founder of quant hedge fund High-Flyer. Similar to peers OpenAI and Anthropic, DeepSeek produces Large Language models. In contrast to most US firms DeepSeek has been much more transparent in their methods, publishing research papers explaining the detailed work behind the R1 model whereas peers mostly tightly guard their intellectual property.

Why does this model matter?
The model raised two important questions.

1. Has there been an overspend in AI infrastructure investment?

2. Has the US’s AI technology lead versus China been eroded?

Addressing the first, R1 achieves similar performance to leading reasoning models such as OpenAI’s o1 at a fraction of the cost. Whilst leading edge models are rumoured to cost well in excess of $1bn+ to train, DeepSeek claim to have trained R1 at a fraction of the cost. There is much debate on the exact cost and the veracity of DeepSeek’s claims, however it would appear even if they are understated this is still an order of magnitude improvement in terms of efficiency of compute. The AI race in the US has been dominated by “who can invest more?” and the belief in Scaling Laws (that AI model performance improves with larger AI clusters). Now a smaller, infrastructure constrained startup has shown there are significant performance gains possible through innovative software methods.

In terms of geopolitical implications, US policy has explicitly targeted slowing China’s progress in AI through semiconductor export controls first introduced in October 2022 and subsequently expanded.¹ DeepSeek’s achievements raise questions around the effectiveness of these policies. There are claims China has found ways around these rules to gain access to leading edge chips, however it’s also entirely possible DeepSeek was able to create a frontier model on less performant hardware in full compliance with regulations.

Tech market reaction
The implications of cheaper models as well a greater diversity of leading edge models are hotly debated, the initial market reaction has generally been:

- Positive for companies who benefit from cheaper AI inference costs.
Under the assumption that the cost to integrate AI functionality (e.g. software) will sharply reduce.

- Negative for those exposed to continued increases in AI datacentre investments.
Including semiconductors, networking and power companies.

Responses from CEOs overseeing tens of billions of dollars in annual investment into AI has been that DeepSeek has shown technological progress which we hope to adopt and that reducing the cost of a technology only helps drive further AI adoption.

Mark Zuckerberg on Meta’s earnings call 29/01/25:
“I can start on the DeepSeek question. I think that there's a number of novel things that they did that I think we're still digesting. And there are a number of advances that we will hope to implement in our systems. And that's part of the nature of how this works, whether it's a Chinese competitor or not. I kind of expect that every new company that has an advance – that has a launch is going to have some new advances that the rest of the field learns from, and that's sort of how the technology industry goes.”

Satya Nadella on Microsoft’s earnings call 29/01/25:
“In some sense, what's happening with AI is no different than what was happening with the regular compute cycle. It's always about bending the curve and then putting more points up the curve. So, there's Moore's Law that's working in hyperdrive. Then on top of that, there is the AI scaling laws, both the pre-training and the inference time compute that compound, and that's all software. You should think of what I said in my remarks, which we have observed for a while, which is 10x on improvements per cycle just because of all the software optimizations on inference. And so, that's what you see.”

It’s still too early to say whether this will be remembered as just another technological advancement lowering the cost to serve a still nascent technology or something with greater impact both geopolitically and for the companies investing so much into AI.

¹ https://www.csis.org/analysis/choking-chinas-access-future-ai

Important Information

The value of investments and the income derived from them can fall as well as rise, and past performance is no indicator of future performance. Investment products may be subject to investment risks involving, but not limited to, possible loss of all or part of the principal invested.

This document does not constitute and shall not be construed as a prospectus, advertisement, public offering or placement of, nor a recommendation to buy, sell, hold or solicit, any investment, security, other financial instrument or other product or service. It is not intended to be a final representation of the terms and conditions of any investment, security, other financial instrument or other product or service. This document is for general information only and is not intended as investment advice or any other specific recommendation as to any particular course of action or inaction. The information in this document does not take into account the specific investment objectives, financial situation or particular needs of the recipient. You should seek your own professional advice suitable to your particular circumstances prior to making any investment or if you are in doubt as to the information in this document.

Although information in this document has been obtained from sources believed to be reliable, no member of the EFG group represents or warrants its accuracy, and such information may be incomplete or condensed. Any opinions in this document are subject to change without notice. This document may contain personal opinions which do not necessarily reflect the position of any member of the EFG group. To the fullest extent permissible by law, no member of the EFG group shall be responsible for the consequences of any errors or omissions herein, or reliance upon any opinion or statement contained herein, and each member of the EFG group expressly disclaims any liability, including (without limitation) liability for incidental or consequential damages, arising from the same or resulting from any action or inaction on the part of the recipient in reliance on this document.
The availability of this document in any jurisdiction or country may be contrary to local law or regulation and persons who come into possession of this document should inform themselves of and observe any restrictions. This document may not be reproduced, disclosed or distributed (in whole or in part) to any other person without prior written permission from an authorised member of the EFG group.

This document has been produced by EFG Asset Management (UK) Limited for use by the EFG group and the worldwide subsidiaries and affiliates within the EFG group. EFG Asset Management (UK) Limited is authorised and regulated by the UK Financial Conduct Authority, registered no. 7389746. Registered address: EFG Asset Management (UK) Limited, Park House, 116 Park Street, London W1K 6AP, United Kingdom, telephone +44 (0)20 7491 9111.