January 2025

DeepSeek, an innovative artificial intelligence (AI) model developer, has captured investor attention over the last few days and exerted significant price action on public AI infrastructure companies due to perceived supply and demand implications.

WHAT WE KNOW

• DeepSeek is an AI subsidiary of a Chinese hedge fund that has developed and published several open-source large language models (LLM) exhibiting unique innovations.
• These innovations are expected to drive stronger efficiency for future LLMs, while significantly democratizing the cost of LLM inference (human-like responses to prompts).
• Many credible AI leaders have praised the advancements DeepSeek has made, including Satya Nadella (CEO, Microsoft), Marc Andreessen (General Partner, Andreessen Horowitz), and Andrej Karpathy (Co-Founder, OpenAI).

WHAT WE DON’T KNOW

• DeepSeek has claimed their training costs were 90% lower to deliver the same performance of leading-edge LLMs, this claim has not been externally verified and may exclude significant costs common to other LLM deployments.
• Censorship and data privacy concerns may exist for model users due to broad terms of service and large-scale malicious attacks DeepSeek has reported on its services.
• Broader geopolitical context, including any regulatory response from the US, given the uniform importance of AI superiority as a national priority for many developed countries.

BAHL & GAYNOR VIEW

• DeepSeek’s innovations are an important reminder that general purpose technologies tend to become commoditized. In this case, LLMs could end up like air travel or electricity – wonderful technologies for society, but unable to earn monopoly rents.
• The deflationary trends of LLM training and inference costs, and the innovations DeepSeek and other organizations have brought to the table, will likely continue to accelerate AI development (known as Jevon’s Paradox).
• DeepSeek’s accomplishments are a validation of the open source approach to AI model development, the use of multiple LLMs to arrive at the best solution, and the use of “distillation” (extracting understanding from an existing “teacher” model to train the “student” model) to accelerate development.
• If details of DeepSeek’s approach from a hardware perspective are validated, it would also be a validation of the custom silicon approach to compute clusters and that programming unique optimizations on custom silicon is possible and indeed, preferable.

BOTTOM LINE

Much of the focus on AI over the past two years has been around voracious demand for best-in-class AI infrastructure, and the benefits of this focus have been extremely concentrated among a select few providers. The ongoing efficiency gains that human ingenuity can unlock in this and other technologies should not be underestimated. Disruption is a fact of life in technology, requiring constant vigilance in investing to address risks and capture opportunities.

Published on 1/29/2025.

The information provided herein is for informational purposes only and does not constitute an offer or solicitation to buy or sell any securities. The views expressed reflect the opinions of Bahl & Gaynor as of the date of this communication and are subject to change. Bahl & Gaynor assumes no liability for the interpretation or use of this report