AI: The Emerging Power Struggle Over Energy Consumption
The rapid expansion of artificial intelligence is driving an unprecedented surge in global energy consumption. As tech companies race to deploy ever more powerful AI models, concerns are mounting over the sustainability and environmental impact of this technological revolution.
Tl;dr
A New Kind of Race: AI and Its Unseen Costs
In recent months, the dramatic progress of generative artificial intelligence has ignited both enthusiasm and deep concern. While users marvel at the capabilities of systems like ChatGPT-4, a crucial dimension often goes unaddressed: the significant environmental impact these advances entail. A recent analysis from French carbon accounting specialist Greenly brings this oversight into sharp relief, focusing not just on technical supremacy but on the overlooked question of sustainability.
The DeepSeek Approach: Rethinking Efficiency
Challenging established giants in the field, Chinese newcomer DeepSeek is introducing a distinctive path forward. Rather than relying on brute force, DeepSeek employs an innovative « Mixture-of-Experts » (MoE) architecture. This means only the essential sub-models are activated for each task—a marked departure from monolithic approaches. The practical results are striking: engineers managed to train their system using just 2,000 NVIDIA H800 chips, far fewer than the estimated 25,000 typically required by ChatGPT-4. According to data gathered by Greenly, this equates to up to ten times fewer GPU hours consumed compared to American or European leaders.
Three main factors drive this efficiency:
The Western Giants: Power Comes at a Price
On the other hand, Western frontrunners continue escalating scale at considerable ecological cost. With its staggering 1.8 trillion parameters, ChatGPT-4‘s power surge comes with an equally explosive energy appetite—twenty times higher than its predecessor. According to Greenly, running the model intensively for automated responses (say, one million emails monthly) leads to annual emissions of nearly 7,138 tonnes of CO₂e. To put it in perspective, that’s roughly comparable to more than 4,000 Paris–New York round-trip flights.
Routine usage only amplifies these figures. For example, generating a single text prompt already uses about 16% of a full smartphone charge; switching to image generation tools like DALL-E inflates this impact up to sixtyfold per request. And when trends go viral on social media—think « starter packs Ghibli »—the environmental toll worsens: each image can require as much as 3.5 liters of water, in addition to substantial electricity.
The Challenge Ahead: Is Sustainable AI Possible?
As demand for generative AI balloons worldwide, experts and industry insiders are exploring possible solutions—from adopting specialized chips such as Google’s TPUs to leveraging renewable energy or advocating for ambitious legislative standards. Yet uncertainty lingers. In the words of Alexis Normand, CEO of Greenly: «With DeepSeek entering the fray, AI competition now hinges not just on performance but also energy efficiency… The question remains unresolved.»
Ultimately, whether these green initiatives will suffice—or whether relentless pursuit of power will continue overshadowing climate concerns—remains an open and pressing issue for tomorrow’s AI landscape.