Controversies in AI in light of DeepSeek
Below is Chautauqua’s latest take on how the entrance of DeepSeek has disrupted the AI market and how the team is thinking about it in relation to their investment process.
DeepSeek is a Chinese Artificial Intelligence (“AI”) company that released its latest model, called R1, in late January. R1’s capabilities are comparable to those of advanced models from the leading U.S. AI companies such as OpenAI and Google. What is most remarkable is that DeepSeek accomplished these results by building better efficiencies into its model out of necessity—it was constrained with computing resources due to U.S. export controls. U.S. companies are not beholden to these same constraints, so their path for AI involved aggressively scaling up computing clusters rather than being hyper-focused on extracting model efficiencies. This implies that DeepSeek spent significantly less than U.S. peers to produce a model that performs similarly.
Cost of the cheapest Large Language Models1
Algorithmic improvements and optimizations decrease cost and increase capability
Source: SemiAnalysis, OpenAI, Together.ai; https://semianalysis.com/2025/01/31/deepseek-debates/
DeepSeek’s success, in spite of its use of lower-spec graphics processing units (“GPUs”), challenges the paradigm set by U.S. companies that larger computing clusters built with leading-edge GPUs are needed to train the best AI models. Therefore, the markets have become concerned over whether datacenter capital expenditures—which has been an arms race, especially over the last two years, amongst the large AI, internet, and cloud infrastructure companies—would continue to grow unabated, or whether DeepSeek’s efficiency breakthrough means that we are already at or near peak datacenter capital expenditures.
From an investment perspective, the semiconductor supply chain is at the center of this controversy. The datacenter arms race has undergirded the recent growth of several of our portfolio companies, including Nvidia (GPUs), TSMC (leading-edge manufacturing), ASML (equipment for leading-edge manufacturing), and Micron (memory). DeepSeek demonstrated that heavy optimization can produce remarkable results on weaker hardware. But an extension of this thinking could also mean that combining DeepSeek’s model efficiency with the very best hardware would produce even better results.
In the near term, DeepSeek’s breakthrough has cast some doubt on the most optimistic version of the datacenter capex growth story—that growth simply continues uninterrupted—and maybe has heightened the risk of an air pocket or digestion period. But even still, datacenter capex comments from this recent earnings season portend high double-digit growth in 2025.
In the long term, and as history has shown, advances in computing power drive more capable and ubiquitous software applications, which reinforce the virtuous cycle of long-term structural growth for both hardware and software. Never in the history of the computer has the demand for computing power declined or plateaued. Looking forward, more computing power is likely to be used to drive ever-higher AI model capabilities and usage. We believe our portfolio companies are in the most advantageous positions in their parts of the semiconductor supply chain to benefit from these trends.
Computation used to train notable artificial intelligence systems2
Computing power needed to train AI continues to increase, consistently requiring more computing power to train AI as the number of task domains using AI has grown exponentially in the last 10-15 years.
Data source: OurWorldinData.org and Epoch (2024)
Separately, it is also noteworthy that U.S. export controls on semiconductors have not only failed with their key objective, but in a way have backfired. Export controls may have delayed Chinese progress in AI by a few years, but they also induced some major innovations. DeepSeek appears to be just as far down the road with AI as the leading U.S. companies. Much of the shock created by DeepSeek is owed to its provenance. But rather than re-thinking policies, the more likely response is that the U.S. will double down—export controls are more important than ever now that China has leading AI models of its own.
Disclosures:
Google had a 0.00%, Nvidia 0.00%, Taiwan Semiconductor 5.14%, ASML 3.08%, and Micron 0.00% weighting in the International Fund as of 12/31/2024. Nvidia had a 3.33%, Taiwan Semiconductor 4.13%, ASML 2.39%, and Micron 1.99% weighting in the Global Fund as of 12/31/2024.
1SemiAnalysis, OpenAI, Together.ai; https://semianalysis.com/2025/01/31/deepseek-debates/. Cheapest Large Language Models (LLM) above certain Massive Multitask Language Understanding (MMLU) cost per 1 million tokens.
2Epoch AI, ‘Parameter, Compute and Data Trends in Machine Learning’. Published online at epochai.org. Retrieved from: https://epoch.ai/data/epochdb/visualization. Computation is measured in total petaFLOP, which is 10¹⁵ floating-point operations. Estimated from AI literature, albeit with some uncertainty. Estimates are expected to be accurate within a factor of 2, or a factor of 5 for recent undisclosed models like GPT-4. A floating-point operation (FLOP) is a type of computer operation. One FLOP represents a single arithmetic operation involving floating-point numbers, such as addition, subtraction, multiplication, or division. Note: The Executive Order on AI refers to a directive issued by President Biden on October 30, 2023, aimed at establishing guidelines and standards for the responsible development and use of artificial intelligence within the United States.
Investors should consider the investment objectives, risks, charges and expense of each fund carefully before investing. This and other information is found in the prospectus and summary prospectus. For a prospectus or summary prospectus, visit bairdfunds.com. Please read the prospectus or summary prospectus carefully before investing.
The Funds may hold fewer securities than other diversified funds, which increases the risk and volatility because each investment has a greater effect on the overall performance. The Fund invests inequity securities of foreign companies. Foreign investments involve additional risks such as currency rate fluctuations and the potential for political and economic instability, and different and sometimes less strict financial reporting standards and regulations.
Information in this document regarding market or economic trends, or the factors influencing historical or future performance, reflects the opinions of management as of the date of this document. These statements should not be relied upon for any other purpose. Holdings are subject to risk and can change without notice. The performance of any single fund holding is no indication of the performance of other holdings of the Baird Chautauqua Funds. References to individual securities should not be construed as a recommendation to buy, hold, or sell a security, and it should not be assumed that holdings have been or will be profitable.
Baird Funds are offered through Robert W. Baird & Co. Incorporated, a registered broker/dealer, member NYSE and SIPC. Robert W. Baird & Co. Incorporated also serves as investment advisor for the Fund and receives compensation for these services as disclosed in the current prospectus.
©2025 Robert W. Baird & Co. Incorporated.
First Use: 02/2025