Energy gluttony in the AI age
The dangerous appetite of large language models
Large language models (LLMs) eat electricity at an alarming rate. As these popular AI models continue to grow in size and popularity, their energy consumption threatens to outpace our ability to power them sustainably. This presents a looming crisis at the intersection of technological progress and environmental responsibility.
Can we do anything to stop it?
The scale of the problem
LLMs require massive computing power, both for training and inference. A single training run for a model like GPT-3 consumed an estimated 1,287 MWh of electricity - equivalent to the carbon dioxide emissions from 550 round-trip flights between New York and San Francisco. Inference, the process of actually using these models to generate responses, compounds the issue. ChatGPT's monthly energy usage for inference alone may range from 1 to 23 million kWh.
Proponents of LLMs often tout improvements in energy efficiency as a solution. They point to techniques like quantization and knowledge distillation that can reduce model size and power consumption. However, these incremental gains pale in comparison to the exponential growth in model size and usage.
This thinking is a fallacy. While smaller models like Guanaco can achieve 97% of ChatGPT's performance with less computational resources, it fails to address the fundamental issue: the relentless push for ever-larger models and wider deployment.
The demand for LLMs has sparked a GPU arms race among tech giants. Companies scramble to secure vast quantities of high-performance GPUs, driving up costs and energy consumption. This frenzy exacerbates the environmental impact of AI development. The Wall Street Journal notes that the "AI boom is fueling an insatiable appetite for electricity." While the article overestimates the immediate impact of AI on energy demand, it correctly identifies the shift towards fossil fuels to meet this growing need.
Consume, consume, consume!
As demand for AI computation grows, tech companies increasingly turn to fossil fuels, particularly natural gas, to power their data centers. This reverses years of progress in transitioning to renewable energy sources. The rush to build and expand data centers outpaces the development of clean energy infrastructure. The timeline for building a data center spans just a few years, while adding renewable energy to the grid faces longer delays due to permitting issues. This mismatch incentivizes the use of readily available fossil fuel power sources.
Beyond electricity consumption, LLMs place a significant burden on water resources. Data centers require enormous amounts of water for cooling. A recent study found that ChatGPT uses approximately 500 milliliters of water for every 50-100 queries - roughly equivalent to a standard water bottle. The study cites figures released by Google and Microsoft showing spikes in their water usage attributed to AI work — as LLMs proliferate, their water consumption could strain local resources, particularly in drought-prone areas.
The energy consumption of LLMs bears a disturbing resemblance to the cryptocurrency mining crisis. Both technologies rely on computationally intensive processes that scale rapidly and consume vast amounts of electricity. However, unlike cryptocurrencies, which offer dubious societal value, LLMs promise tangible benefits in various fields.
Lux Research draws this parallel, noting that Bitcoin alone consumes more than 140 TWh annually. While LLMs haven't yet reached this level, their potential for rapid growth and widespread adoption could lead to similar energy consumption patterns.
The transparency problem
A major obstacle in addressing the energy impact of LLMs lies in the lack of transparency from major AI companies. Many closed-source models, including industry leaders like GPT-4o and Gemini, provide little-to-no information about their energy consumption or environmental impact. Many of these providers fall short in meeting the energy criteria set by the EU AI Act. This opacity hinders effective regulation and public accountability.
Current regulations fail to adequately address the energy consumption of AI systems. While initiatives like the EU AI Act propose disclosure and harmonization of AI energy data, implementation remains in early stages. The lack of standardized metrics for measuring and comparing LLM energy efficiency further complicates regulatory efforts. There is a growing need for advocation of policies that facilitate the addition of renewable energy to the grid, including permitting reform in the U.S. Without such measures, the AI industry will likely continue to rely on fossil fuels to meet its growing energy needs.
Efficiency vs ethicality
Proponents of LLMs often cite historical trends in data center energy efficiency as evidence that AI's energy consumption will naturally plateau. They point to the fact that despite massive growth in computing power and digital services over the past decade, data center energy usage has remained relatively flat.
However, this argument ignores the unique characteristics of LLMs. Unlike traditional data center workloads, LLMs have the potential for exponential growth in both model size and usage. An observation of potential energy consumption for AI-enhanced Google searches demonstrates how even modest adoption of LLMs could double a major tech company's energy usage.
While LLMs may offer benefits in various fields, their current trajectory threatens to divert resources from more immediate and impactful solutions. The focus on building ever-larger language models may come at the expense of targeted AI applications that could directly address environmental and social issues. The energy, water, and computational power consumed by these models could be directed towards solving pressing global challenges, including climate change mitigation and adaptation.
The energy consumption of LLMs raises ethical questions about the distribution of resources and environmental burdens. As tech companies in wealthy nations push the boundaries of AI capabilities, the environmental consequences may disproportionately affect vulnerable communities around the world. The water usage of data centers, for example, could exacerbate water scarcity in regions already struggling with drought. Similarly, the increased demand for fossil fuel energy to power AI infrastructure may accelerate climate change, with its most severe impacts felt by developing nations.
The path forward
Addressing the energy gluttony of LLMs requires a multi-faceted approach:
Transparency and standardization. Tech companies must provide detailed, standardized information about the energy consumption of their AI models. This includes both training and inference costs.
Regulatory framework. Governments need to establish clear guidelines and limits on the energy usage of AI systems, similar to energy efficiency standards for consumer appliances.
Renewable infrastructure. Massive investment in renewable energy infrastructure is necessary to meet the growing demand from AI without relying on fossil fuels.
Efficiency research. Continued research into energy-efficient AI architectures and training methods is crucial. This must be coupled with a willingness to limit model size and complexity when the marginal benefits fail to justify the energy costs.
Purposeful development. The AI community should prioritize the development of smaller, task-specific models that offer concrete benefits while minimizing energy consumption.
Water conservation. Data center designs must prioritize water conservation, exploring alternative cooling methods and locating facilities in areas with abundant water resources.
Carbon pricing. Implementing carbon pricing mechanisms would incentivize tech companies to fully account for the environmental costs of their AI systems.
Public awareness. Educating the public about the energy costs of AI is important for building support for necessary regulatory and infrastructure changes.
The unchecked growth of large language models poses a serious threat to global efforts to reduce energy consumption and combat climate change. While these AI systems offer potential benefits, their current trajectory is unsustainable. The tech industry's pursuit of ever-larger models and wider deployment must be tempered by a realistic assessment of the environmental costs.