The estimated computing cost of training Google’s Gemini 1.0 Ultra was approximately $192 million which is roughly ten times the annual public-university research budget of my home country, Bangladesh. This stark disparity highlights the financial barriers that researchers in resource-constrained countries face when developing advanced AI systems. The potential for scientific breakthroughs should never be limited by computing budgets; the brilliance of ideas should matter more than the resources behind them.

But money is only part of the story. The environmental burden is equally concerning. According to the International Energy Agency, data centers consumed approximately 1.5% of the world’s electricity in 2024, a share projected to nearly double by 2030.

Meanwhile, a United Nations University study warns that AI-related water consumption could equal the basic annual domestic needs of 1.3 billion people by the end of the decade. With the rapid growth of AI adoption, increasing demand for computing resources could place additional pressure on electricity grids, water supplies, and the environment.

Can we make AI research accessible to everyone without placing an ever-growing burden on the environment? One promising answer lies in efficient AI. Instead of relying solely on larger and more expensive computers, we should focus on developing AI systems that accomplish more with fewer resources.

Efficiency can take several forms: reducing computational costs, lowering memory requirements so models can run on affordable hardware, and minimizing energy consumption. Together, these improvements could enable more researchers to participate in AI development while reducing its environmental footprint.

Just as efficiency has multiple dimensions, researchers are exploring different approaches to achieving it. Stanford’s 2025 AI Index reports that an AI model 142 times smaller than one from 2022 achieved the same performance threshold on a knowledge benchmark in 2024.

Consider the financial implications: an NVIDIA B200 graphics processor can cost around $30,000–$40,000, while a consumer RTX 5070 launched at $549. If efficiency improvements reduce a model’s memory requirements from 100 GB to 10 GB, some workloads could move from expensive data-center hardware to affordable consumer equipment, provided it offers sufficient computing performance. This could open new opportunities for researchers with limited budgets.

Green AI has also become a growing focus among researchers. In a 2025 study, researchers at Microsoft demonstrated that their BitNet model could generate text using substantially less estimated energy than selected conventional AI models of similar size, while maintaining competitive performance. Their results suggest that making AI more energy-efficient does not necessarily require sacrificing its capabilities.

Cheaper hardware and lower computational requirements can encourage greater AI adoption, potentially offsetting environmental savings. Therefore, efficiency alone cannot guarantee sustainability; it must be accompanied by responsible resource management. Together, these efforts can make AI more accessible while reducing its environmental footprint.

Without the combined efforts of all stakeholders, this goal cannot become a reality. Universities should encourage research that reduces computing costs, memory requirements, and energy consumption. Funding agencies should prioritize projects that improve resource efficiency alongside model performance and support researchers working with limited computational resources.

Technology companies should invest in efficient AI systems and transparently report their computing costs, electricity consumption, and water usage. Policymakers should establish environmental standards for data centers and ensure that new facilities are developed with due consideration for local energy and water resources. Together, these measures would encourage responsible growth without unnecessarily restricting innovation.

The future of AI should be built on smarter ideas, not simply bigger computers and greater resource consumption. Our goal should be to make AI innovation both inclusive and sustainable.