How DeepSeek's Breakthrough Is Changing The Landscape For Multi-Industrials

BofA Securities analyst Andrew Obin has provided his view about the implication of DeepSeek on multi-industrial stocks and reiterated a Buy rating on the shares of Vertiv Holdings Co VRT, GE Vernova Inc GEV and Eaton Corporation PLC ETN with a price forecast of $165, $485 and $410 respectively.

In December, Chinese startup DeepSeek introduced its latest AI model, DeepSeek-V3, which has shown comparable performance to OpenAI’s GPT-4 across various benchmark tests, noted the analyst.

According to the analyst, Investors are wary that this technological breakthrough could result in reduced spending on AI infrastructure and power generation.

Also Read: Jensen Huang Loses $20B In Wealth: How DeepSeek Hit Nvidia Stock And World’s Richest People

DeepSeek’s inference costs ($/million tokens) are approximately 70-90% lower than those of other AI models. The critical factor to monitor will be the capital expenditure plans of cloud service providers for 2025-2026, noted the analyst.

The technical report states that V3 utilized just 2.8 million GPU hours. At a rental rate of $2 per hour, this would total $5.6 million, per the analyst.

However, the analyst opined that DeepSeek-V3 was “distilled” using the previously launched DeepSeek-R1 model. Some media reports suggest that DeepSeek may have also used other open-source models, such as Meta Platforms Llama, implying that the actual training cost could be significantly higher than reported.

Venture capital investor Marc Andreessen referred to DeepSeek-R1 as a “Sputnik moment” for AI progress. However, the analyst interprets the Sputnik comparison as favorable for AI infrastructure companies.

Following the launch of Sputnik (October 4, 1957), the U.S. Federal space R&D budget grew significantly, from $0.5 billion annually to over $10.5 billion in 1958.

When a new technology boosts efficiency, it typically reduces demand. However, it often leads to increased consumption, a phenomenon known as Jevons Paradox.

A similar situation occurred in U.S. steel production after the Bessemer process was introduced: between 1875 and 1900, steel prices dropped by around 90%, while production surged from 0.4 million tons per year to 60 million tons per year.

The analyst notes this will apply to the AI sector as well. Reduced training costs accelerate model improvements, and enhanced models create more use cases, resulting in greater inference demand.

For instance, the cost of GPT-4 output tokens has dropped by approximately 80% in under a year since its launch.

Cloud service providers are still experiencing growth in non-AI revenue, with Microsoft Corp MSFT Azure non-AI revenue rising by 22% year-over-year in the quarter ended September.

Colocation companies, which make up about 50% of global data centers, continue to benefit from favorable market conditions, concluded the analyst.

Read Next: