DeepSeek’s recently released open-source reasoning, large language model (LLM) V3-R1 has benchmarked closely to OpenAI’s o1 framework. The Chinese company drastically lowered the cost of running its LLM via several novel methods. Firstly, DeepSeek drastically improved reasoning using stronger reinforcement learning techniques that reduce precision while maintaining accuracy. They also employ a mixture-of-experts architecture, which combines many focused models that only call on individual components when necessary, rather than relying on one large model that requires vast amounts of memory. These techniques allowed for high process performance with limited hardware capabilities.
On top of achieving similar performance to OpenAI’s o1, DeepSeek R1 is believed to have been trained on chips comparable to Nvidia’s A100 chip, which is two generations behind the latest Blackwell chip. This drastic change in model optimization has shifted industry leaders’ focus to refining their models to improve efficiency and reduce power consumption. We do not expect this to change near-term data center growth, but it raises questions about longer-term capital allocations for computing hub expansion.
Research Highlights:
- Class VI Update 4Q24 – The Wave Is Coming – In this new quarterly report series, Enverus Intelligence® Research provides an overview of recent changes and additions to the growing list of Class VI wells associated with CCS project in the U.S. Leveraging the Enverus FOUNDATIONS® — Carbon Innovation Wells database, this report series covers new Class VI applications, changes to permit status, permit approvals and newly disclosed project details. In this inaugural report, we cover updates in 4Q24 as well as highlights from 2024.
- Delayed Exit – Rising Demand Keeps Power Plants Online – How have thermal power plant retirements shifted in response to rising demand and the growing need for baseload generation? Which natural gas plants are likely to have their retirements delayed based on plant efficiency and projected power demand growth?
- DeepSeek R1 Breakthrough – Putting the Bull Case on Notice – This report addresses the release of DeepSeek’s R1 AI model, the impacted markets and the bull case for energy.
About Enverus Intelligence®| Research
Enverus Intelligence® | Research, Inc. (EIR) is a subsidiary of Enverus that publishes energy-sector research focused on the oil, natural gas, power and renewable industries. EIR publishes reports including asset and company valuations, resource assessments, technical evaluations, and macro-economic forecasts and helps make intelligent connections for energy industry participants, service companies, and capital providers worldwide. See additional disclosures here.
