Training AI: The Real Carbon Footprint Could AI Be the Next Big Energy Consumer After Bitcoin?

The adoption of artificial intelligence (AI) is growing continuously with the aid of innovations, automation, robots, machine learning, and countless other advanced technologies, we can perform tasks with greater efficiency and precision. However, amidst the buzz of AI’s current and potential achievements lies the perilous issue of AI’s increasing energy consumption and sustainable growth. Should we format AI as an added energy predator like we do with Bitcoin?  

The Increase in Electricity Consumption Eagered by AI Models  

The revolution in AI technology is underpinned by deep-learning-based language models, such as GPT series developed by OpenAI and Gemini by Google DeepMind. Such models require overwhelming computational capabilities for training and execution. During training, billions of parameters (the frameworks of an AI’s “understanding”) must be processed, utilizing powerful GPU clusters available in an expansive network of data centers.  

One research suggested that training an AI model consumes energy equivalent to a household’s annual consumption. In the AI industry’s context, running on idle capping machines in different data center locations translates to millions of dollars in costs throughout the year: Maintaining Energy Hungry AI Models’ infrastructure. In this regard, OpenAI claimed GPT-3 consumed approximately 1,287 MWh of electricity (equivalent to about 550 tons of carbon dioxide emissions, depending on the energy source). This worrying singular model trend spells doom as demand-oriented deployments bolster the industry’s energy appetite.

From Bitcoin to AI: A Looming Energy Battle?

The unique nature of verifying transactions through the proof-of-work (PoW) algorithm in Bitcoin mining poses a massive contradiction as a decentralized and competitive practice. This has culminated in the excessive criticism directed towards Bitcoin mining for its massive energy consumption which is estimated to be 120-150 TWh terawatt hours.’ A good analogy for this would be to equate energy consumption to entire countries such as the Netherlands and Argentina.

Bitcoin, however, has some competition in the form of AI, and as it seems it is catching up at an astonishing rate. It is estimated that virtually by 2025 the global energy demand in training and deploying AI systems could rival or surpass Bitcoin’s current consumption levels. Unlike Bitcoin, AI’s energy requirements are not limited to training; they also encompass inference engines for real-time applications. These include chatbots, recommendation systems, search engines, etc.

The Environmental Impact

The trend AI adoption and development follows is deeply concerning. Data centers that enable the deployment and training of AI models account for roughly 1-2% of global electricity usage, and this figure will drastically rise in light of the accelerated adoption of AI across the globe. Should these data centers be powered by fossil fuels, the carbon emissions would further deteriorate the climate and environment well-being, all the while making attempts to urge people to cut down on their carbon footprint.

In the case of large companies, some of these problems are being addressed by investing in renewable energies and improving the efficacy of the hardware being used. For example, NVIDIA’s new AI chips are built to do more calculations for every unit of energy used. Still, the advances in these technologies and the scale of the construction of AI systems remain extraordinarily difficult.

The Future: Moderating the Rate of Developments and Sustainability

What is the answer then? Energy usage needs to be approached in a most effective manner. From the very beginning, it should be a concern in hardware and software structuring. There are some attempts utilizing machine learning on pruning and quantization methodologies which limit the resources which are used to perform computations while retaining sufficiently high levels of performance. A considerable amount of attention is also required from the states and the IT circles to clean energy to sustain the newly designed AI infrastructure.

In AIXCircle, we are of the opinion that AI can change the world in many domains such as health care, finance, etc. As these opportunities are enticing, one cannot stay back and tried grabbing them. However, it’s also essential that AI should not become a technology like Bitcoin which offers a lot in terms of capabilities but fails to address the environmental challenges it creates.

Conclusion

The sharp rise in AI’s energy consumption serves as a reminder that model training should be approached with a balanced perspective toward sustainable resource use. As the competition to create environmentally responsible, resource-efficient AI systems heats up, it is critical to align innovations with ethical responsibility.