AI threatens the natural resources of billions of people

Introduction

The rapid development of artificial intelligence is reshaping the global economic and social landscape at an unprecedented speed, but the resource consumption issues hidden behind it are often overlooked. In 2025, the United Nations University Institute for Water, Environment and Health released a report titled "Environmental Costs of AI Energy Consumption: Carbon, Water and Land Footprints," which for the first time systematically quantified the enormous global demand for energy, water resources and land from data centers supporting AI operations. The report predicts that by 2030, global AI related data centers will consume 945 TWh of electricity annually, equivalent to the basic living needs of 1.3 billion people per year, and occupy over 14500 square kilometers of land. These numbers reveal a harsh reality: while AI brings efficiency and intelligence, it is posing a substantial threat to the natural resources that billions of people rely on for survival.

1、 Energy consumption: AI driven power black hole

The report states that by 2025, global data centers will have consumed 448 TWh of electricity. If viewed as a country, this number makes it the 11th largest electricity consumer in the world, second only to France and higher than Saudi Arabia. By 2030, this number is expected to double to 945 TWh, accounting for over 3% of the world's total electricity generation. Taking single model training as an example, GPT-5 training is expected to require about 100 GWh of electricity - equivalent to the residential electricity consumption of about 770000 people a year in sub Saharan Africa. Considering that there are still about 800 million people worldwide who lack basic electricity supply, the contradiction between AI's energy demand and human basic needs is becoming increasingly prominent. More noteworthy is that every 1 kWh of electricity consumption corresponds to three environmental footprints: carbon footprint from fossil fuel power generation, freshwater resource consumption from cooling processes, and land occupation for infrastructure construction. Currently, about 60% of global data centers still rely on fossil fuels for power supply, which means that for every 1 TWh increase in AI computing power, it is equivalent to emitting about 400-600 tons of carbon dioxide into the atmosphere. If the transition to clean energy is not accelerated, AI will become a new engine for global carbon emissions growth.

2、 Water resource consumption: the debate between cooling demand and domestic water use

The cooling system of AI data centers is the main source of water consumption. The report estimates that by 2030, the annual water consumption of AI related data centers will reach approximately 10 billion cubic meters, equivalent to the basic domestic water demand of 1.3 billion people per year (calculated at 70 liters per person per day). This number has exceeded the total annual water supply of many arid countries or regions. The water consumption for training GPT-5 is about 1 billion liters, enough to fill 400 standard Olympic swimming pools. The pressure of water resources is particularly uneven in regional distribution. About 40% of global data centers are located in water scarce areas such as the southwestern United States, India, and northern China. These regions are already facing fierce competition for agricultural and residential water use, and the additional demand for AI facilities may exacerbate the local water resource crisis. In addition, cooling water often carries heat and chemical additives after use, and even if discharged after treatment, it can cause thermal and chemical pollution to the aquatic ecosystem. With the explosive growth of inference demand in AI applications (inference processes account for 80% -90% of total AI energy consumption), water resource consumption will show an exponential upward trend.

3、 Land Occupation: From Data Centers to Energy Infrastructure

The report predicts that by 2030, the land occupation of AI data centers and related infrastructure will exceed 14500 square kilometers, equivalent to approximately 2 million football fields. Although the land occupation of 1.5 square kilometers for GPT-5 training seems small, this is a single model training phase; The data center cluster, supporting substations, renewable energy power plants (such as solar and wind energy), and resource extraction facilities (such as rare earth, copper, and lithium mines) in the actual inference stage will lead to a sharp expansion of total land demand. The ecological impacts caused by land occupation include: firstly, destruction of natural vegetation and loss of biodiversity; Secondly, the replacement of farmland and green spaces on the outskirts of cities; Thirdly, the changes in surface runoff and soil erosion caused by large-scale infrastructure construction. Especially in developing countries, a large number of data center locations often prioritize areas with low land prices but ecological sensitivity, leading to an increasing number of environmental and social conflicts.

4、 Inference stage: underestimated sustained consumption

The report emphasizes that although the training process is the focus of media attention, it is the reasoning process that truly drives the continuous growth of AI resource consumption. The inference process accounts for 80% -90% of the total energy consumption of AI - once the model is deployed, it needs to constantly respond to user requests, generate text, images, or analyze results, and each query consumes electricity, water, and land. Taking ChatGPT as an example, the computational resources required for a single query are about 10-30 times that of traditional Google search. If the daily active users worldwide reach hundreds of millions, the cumulative energy consumption will far exceed any training session. What is even more alarming is that the scale of AI models is still growing rapidly. The number of parameters ranges from 175 billion for GPT-3 to 1 trillion for GPT-4, and even rumored to reach the level of 10 trillion for GPT-5. There is no non-linear relationship between model complexity and resource consumption, but rather an approximate exponential growth. Even if the energy efficiency of chips improves by 30% -50% annually, it still cannot offset the explosive expansion of user numbers and application scenarios. This means that by 2030, AI inference energy consumption may account for over 90% of the total global data center energy consumption.

5、 Challenge and Balance Path

The above data reveals a fundamental dilemma: there is a structural tension between the development logic of AI and resource and environmental constraints. On the one hand, AI has irreplaceable potential for reducing emissions and increasing efficiency in fields such as healthcare, climate prediction, and energy management; On the other hand, without intervention, its own resource consumption may offset these benefits and even threaten global sustainable development goals. Solving this problem requires a multi pronged approach: firstly, promoting the transformation of data centers towards clean energy, requiring new data centers to be 100% matched with green electricity; Secondly, developing more efficient computing architectures and cooling technologies, such as liquid cooling and immersion cooling, can reduce water consumption by more than 50%; Thirdly, establish quantitative disclosure standards for AI resource consumption, incorporating carbon footprint, water footprint, and land footprint into corporate ESG reports to guide market choices; Fourthly, at the policy level, an upper limit on AI energy consumption intensity should be set, and environmental impact assessments should be implemented for training large-scale models. In addition, users can also reduce resource requirements by reducing unnecessary AI calls, optimizing model compression and edge computing.

Conclusion

The report by the United Nations University for the first time systematically reveals the astonishing resource consumption panorama behind the development of AI: by 2030, AI data centers will consume nearly 1000 TWh of electricity, approximately 10 billion cubic meters of fresh water, and occupy over 14500 square kilometers of land annually, with water usage equivalent to the basic needs of 1.3 billion people per year. These numbers are not distant predictions, but inevitable results of the current growth trend. Although AI technology is undoubtedly a powerful tool for solving major human challenges, its own resource consumption is no longer an "implicit cost", but directly threatens the natural resource foundation that billions of people rely on for survival. While pursuing intelligent progress, we must be aware that no planet can withstand infinite growth in computing. Only through the triple transformation of technological innovation, institutional design, and social consensus can a true balance be found between the AI dividend and the environmental bottom line.