By Shibasis Rath
Bioscience and technology enthusiast, specializing in life sciences and innovative tech solutions.

2/Feb/2025

Last month, DeepSeek made waves by asserting its AI model consumes roughly one-tenth the computing power of Meta’s Llama 3.1—a claim that challenges long-held assumptions about the energy and resources required to develop advanced AI.

The environmental benefits, on face value, are huge. With tech giants competing to build behemoth data centers that might be comparable to the electricity demand of small towns, concerns are mounting about how AI contributes to climate change: pollution and the pressure on the power grid. That makes any saving in the energy required for training and running generative AI models a boon.

It has been attracting notice since the beginning of December when it launched its V3 model. According to the company's technical report, the final training run cost $5.6 million and consumed 2.78 million GPU hours on Nvidia's older H800 chips. In contrast, Meta's Llama 3.1 405B model reportedly required around 30.8 million GPU hours on newer, more efficient H100 chips—with costs estimated between $60 million and $1 billion for similar models.

The momentum rolled into last week with the release of DeepSeek's R1 model, hailed by venture capitalist Marc Andreessen as "a profound gift to the world." Only days after topping both Apple's and Google's app stores, the danger to competitors' stock prices helped illustrate the industry-shaking potential of a lower-cost, energy-efficient AI alternative. Notably, shares of Nvidia sank sharply after discovering that DeepSeek's V3 required only 2,000 chips for training, compared with 16,000 or more by its competitors.

DeepSeek attributes its efficiency gains to innovative training methods. The company says it reduces energy use by orders of magnitude through an auxiliary-loss-free strategy—selectively training only parts of the model at a time, much like choosing the right experts from a customer service team. Further savings are realized during inference via key value caching and compression, akin to consulting summarized index cards rather than rereading full reports.

In conclusion, energy experts like Madalsa Singh, who is a postdoctoral research fellow at the University of California, Santa Barbara, are positive but cautiously optimistic. "It just shows that AI doesn't have to be an energy hog," she states, pointing out that sustainable AI is an excellent alternative. Doubts, though, still surround it. Carlos Torres Diaz, head of power research at Rystad Energy, noted that there is no concrete data on DeepSeek's energy consumption, while Philip Krein of the University of Illinois Urbana-Champaign cautions that increased efficiency could spur a surge in AI deployment—a phenomenon known as Jevons paradox.

Despite the uncertainty, the implications are profound. If DeepSeek's claims hold true, then reduction in energy demand could ease the environmental footprint of data centers while freeing up renewable energy resources for other sectors. However, long-term effects will depend on how major players respond and whether more sustainable infrastructure planning becomes the norm amid escalating AI-driven power consumption.

Ultimately, DeepSeek's developments will mark a pivotal juncture for the AI industry: continue down the brute-force path to energy consumption or embrace innovative approaches that could alter the balance of technological advancement versus environmental stewardship.

ADVERTISEMENTS

ADVERTISEMENTS

ADVERTISEMENTS

ADVERTISEMENTS