Brief: Knock-On Effects of Open-Source AI

Mobius Intel Brief:

DeepSeek captured the internet’s attention this week with the open-source release of its newest R1 reasoning model, which boasts performance comparable to leading Large Language Models (LLMs)— such as OpenAI’s o1-mini—at (purportedly) one-thirtieth of the training cost.

• ICYMI: DeepSeek’s ~$6 million input for an advanced LLM model fueled nervous claims of an ‘AI bubble’ and raised questions about the $billions pumped into household-name AI companies like OpenAI, Anthropic, Meta, and Nvidia.

While information continues to emerge, available details point us to a different (and much more optimistic) conclusion than the market’s initial reaction to DeepSeek’s announcement:

1. Novel Training Methods: DeepSeek’s reduced training costs stem from novel training methods like reinforcement learning (RL) and predictive modeling to reduce training parameters (see attached for details). These techniques will quickly be adopted by the broader industry and deployed at a much larger scale.

2. Implications for Chip Demand: Contrary to prevailing headline narratives, an advanced open-source model like DeepSeek’s R1 will further support chip demand. Technical consumers have already started procuring hardware capable of running the model locally, and only three major chip manufacturers (Apple, Nvidia, and AMD) currently produce hardware capable of running the full R1 model.

3. Open-Source Tailwinds: Technical and hardware barriers to entry will recede, enabling more consumers/businesses to operate resource-intensive open-source models at scale. The rise of “distributed” LLMs on relatively less efficient hardware will contribute to growing AI-related electricity demand. Quoting former Google CEO Eric Schmidt, “…the fact of the matter is the demand is infinite…”

Takeaway:

• Large Language Models are advancing at non-linear speeds, reinforcing our takeaway that price and reliability risks (driven by regulators’ lagging demand forecasts) skew to the upside.

• Compute remains a material bottleneck for AI advancement. Natural gas will be the primary beneficiary of this constraint.

This commentary contains our views and opinions and is based on information from sources we believe are reliable. This commentary is for informational purposes, should not be considered investment advice, and is not intended as an offer or solicitation with respect to the purchase and sale of commodity interests or to serve as the basis for one to decide to execute derivatives or other transactions. This commentary is exclusively intended for Mobius clients and is not considered promotional material.