Balancing Innovation and Sustainability: AI’s Environmental Dilemma

Key Takeaways

• Training an AI model can produce around 626,155 pounds of CO₂ — nearly five times the lifetime emissions of an average car.
• At the current rate, generative AI’s e-waste could reach 2.5 tons per year by 2030.
• By 2027, the global demand for water due to AI-related activities could reach 6.6 billion m³.
• Of the tech leaders analyzed, Samsung consumes the highest amount of resources. In 2023, its AI-related activities required 29.9 million MWh and 97,482,000 m³ of water.
• Green AI practices could lead to significant reductions in greenhouse gas emissions, potentially mitigating them by 5%-10% globally by 2030.

As AI advances, we face a growing dilemma: how can we harness the power of AI while also protecting our planet?

We at vpnMentor decided to explore the environmental impact of AI, while also shedding light on how it can be used to benefit the environment. By examining this tension, we can better understand how to innovate responsibly and create a sustainable future for all.

The Rising Environmental Cost of AI

While AI has the potential to enhance energy efficiency and improve sustainability efforts as a whole, its impact on the environment is a growing concern. AI requires high energy consumption and relies heavily on fossil fuels, which are the biggest contributors to global warming.

Training generative AI is particularly energy-intensive. A study conducted by the University of Massachusetts found that training an AI model can produce around 626,155 pounds of CO2. To put it into perspective, that’s nearly five times the lifetime emissions of an average car or the equivalent of someone taking 300 round-trip flights between New York and San Francisco.

AI also requires the mining of natural resources (like cobalt, silicon, and gold), leading to soil erosion and pollution. These natural resources are then used to produce the required hardware devices to meet AI’s need for servers and data centers.

To complicate matters further, improper disposal of electronics can lead to the rapid growth of electronic waste. E-waste is particularly dangerous as it contains several toxic substances (such as lead, mercury, and arsenic), which can pose serious risks to both human health and the ecosystem.

A 2024 study published in Nature Computational Science found that generative AI’s e-waste could reach 2.5 tons per year by 2030 if no waste-reduction measures are implemented. The study only took into account large language models, as they are the most computationally intensive.

Finally, the development of AI technology has also led to excessive water consumption by tech giants like Apple and Google. AI data centers require substantial amounts of water to maintain optimal temperatures for the servers that handle complex computations.

Researchers at UC Riverside found that by 2027, the global demand for water from AI-related activities could reach up to 6.6 billion m³ — roughly equivalent to the yearly water consumption of half of the United Kingdom.

Tech Giants’ Energy and Water Usage Trends

To see how the advancement of AI technology correlates with resource consumption, we studied the energy and water usage trends across Meta, Apple, Google, and Samsung. Each of these companies has made groundbreaking contributions to AI, making them ideal for analyzing the intersection of sustainability and AI development.

Meta

In recent years, AI development has become a crucial part of Meta's operations — from natural language processing to generative AI and metaverse technologies. This has led to a significant increase in its water and electricity consumption.

By 2022, its electricity usage reached 11,508,131 MWh — a 60% increase from 2020. The cooling systems used for Meta’s data centers have also contributed to a surge in water consumption — an astounding 214% increase from 2017 to 2022.

Having said that, Meta has also made significant water restoration efforts. In 2022, 2,351,562 m³ of water were restored, which is a huge increase compared to the 132,000 m³ restored only four years prior. The company claims its goal is to be water positive by 2030.

Furthermore, since 2020, Meta has transitioned to 100% renewable energy, largely mitigating the environmental impact of its growing energy demands. The company primarily uses solar and wind energy, and it recently partnered with Sage Geosystems to expand the use of geothermal energy in the United States.

Apple

Over the past decade, Apple has made significant advancements in AI, from Siri to Vision Pro (Apple’s mixed reality headset). These advancements are closely associated with a constant increase in the company’s electricity and water consumption.

For instance, its corporate electricity usage increased by 35% from 2020 to 2023, reaching 3,487,000 MWh. Similarly, water usage at corporate facilities rose from 4,872,474 m³ in 2020 to 6,094,513 m³ in 2023, correlating with the cooling needs of AI activities and expanded infrastructure.

Like Meta, Apple has maintained 100% renewable energy use in corporate facilities since 2018. It has also made strides in water conservation, recycling over 568,000 m³ in 2023 and saving significant amounts of freshwater in its supply chain. This shows Apple’s commitment to reducing the environmental impact of its AI-related operations.

Google

Google has also made significant advancements in the AI industry, especially through training large-scale models like BERT and Gemini. These innovations have led to a considerable expansion of the company’s resource demands.

Its electricity consumption grew by 43% between 2020 and 2022, reaching 21,776,200 MWh. As a result of the cooling needs of data centers housing the AI infrastructure, Google’s water intake increased significantly, too, surpassing 28 million m³ in 2022.

The company has made various sustainability efforts; for instance, it’s been matching all its electricity consumption with 100% renewable energy since 2017. Google also committed to replenishing 120% of the water it withdraws by 2030. Like Meta, Google’s water stewardship strategy is all about improving watershed health in the communities where it operates. In 2023, Google restored 18% of the water it used.

Samsung

Samsung is another tech giant that reports a consistent rise in its resource usage, which is closely related to AI advancements. In early 2024, the company presented its “AI for All” vision, which aims to make AI available and beneficial to everyone.

In 2023, Samsung’s electricity consumption reached 29.9 million MWh. By that point, only 31% of its electricity usage was derived from renewable sources — a small increase from roughly 21% in 2021.

The company’s water consumption also went up, rising from 84.74 million m³ in 2021 to 97.48 million m³ in 2023, which is the highest compared to the other companies analyzed.

Comparing Tech Giants’ Consumption Trends

Overall, Samsung consumed the highest amount of electricity out of the four companies we analyzed, reaching 29.9 million MWh in 2023. Moreover, Samsung’s water consumption was the highest by far, reaching 97.5 million m³ in 2023. Google comes at a distant second, consuming 32.7 million m³ of water that same year.

While all four companies seem to be committed towards minimizing the impact they have on the environment, some have achieved more progress than others. By 2023, Samsung had only managed to match 31% of the electricity it consumed with renewable energy, while the three others have been matching 100% of their electricity consumption for years.

Now, we must keep in mind that being 100% renewable and being carbon-free are not the same thing. The first means that a company is still relying on non-renewable energy some of the time to keep operating, but that it is also buying the equivalent in clean energy so that others can use it. Being carbon-free means always operating on renewable energy, generating no carbon emissions at all — and none of the four companies are there just yet.

Like with electricity, these tech giants have all been making some efforts towards water restoration with varying degrees of success. Meta restored nearly 50% of the water it withdrew in 2022, while Google only managed to restore 18% in 2023. That said, Google’s 18% represents over 5 million m³ of water — more than double the amount that Meta restored.

Impact of GPU Usage in Data Centers

To provide an estimate of the carbon emissions of AI and machine learning, we used the Machine Learning CO₂ Impact tool. This tool allows users to calculate the approximate carbon emissions associated with using machine learning technology on different hardware from various cloud providers.

It also helps assess the environmental impact of machine learning processes by providing insights into the carbon footprint generated during model training and evaluation. The calculator takes into account whether the provider fully offsets its total emissions or not.

Servers without carbon offset don’t compensate for their greenhouse gas emissions through any form of carbon credit purchase. They also typically rely on non-renewable energy sources for their operation, resulting in a higher carbon footprint compared to those that engage in carbon offsetting practices.

When carbon offset is available, the data indicates the two servers with the highest environmental impact are located in Pretoria and Stellenboch, South Africa. The majority of the other highly polluting servers are in India and Australia.

As for servers that lack a carbon offset, those in Australia and India still stand out predominantly, though the majority of the remaining highly polluting servers are located in East Asia.

When analyzing the carbon emissions associated with using machine learning technology on different hardware, we found that the A100 SXM4 80GB hardware emits the most pollution due to its 400 TDP watts consumption.

Combining this with the list of servers that lack a carbon offset, we found that the most polluting server is located in Mumbai, India, emitting 36.80 kg of CO₂. This is equivalent to driving a conventional internal combustion car for about 150 km. The next most polluting servers are located in Sydney (32.08 kg of CO₂) and Hong Kong (28.02 of CO₂).

For comparison, one of the servers with the lowest impact is located in Beauharnois, Canada, emitting only 0.80 kg of CO₂. The significant difference in the amount of carbon emitted by these servers lies in several factors, including the method of energy generation (e.g., hydrocarbons or renewable sources), the effectiveness of the service provider in generating energy, and the efficiency of the data centers in using that energy.

Sustainability Opportunities With AI

Despite its growing resource demands, AI technology can also help minimize environmental degradation and offer innovative solutions to environmental issues.

According to Yuan Yao, associate professor of industrial ecology and sustainable systems at the Yale School of the Environment, the application of AI can have numerous environmental benefits, including enhanced energy efficiency and reduced energy usage. Here are a few examples:

• Data Analysis: AI can analyze large amounts of data, helping companies make smarter decisions about how to use energy more efficiently. One real-world example of this is Google’s DeepMind AI system, which has achieved a 40% reduction in cooling costs for data centers by predicting cooling needs and optimizing operations. These savings are impactful, considering data centers are responsible for 1%-2% of global energy consumption.
• Energy Efficiency in Buildings: AI systems can automatically control heating, cooling, and lighting in buildings based on occupancy and weather conditions. This leads to significant energy savings without sacrificing comfort. For example, an AI system from BrainBox AI was implemented at 45 Broadway, a 32-story office building in Manhattan. This resulted in a 15.8% reduction in HVAC-related energy consumption, saving over $42,000 within 11 months.
• Smart Grid Management: AI enhances smart grids, which are advanced electrical systems that use real-time data to manage energy distribution. For example, a recent initiative in the UK uses smart meter data to help consumers manage their energy use better. The project aims to optimize network loads and reduce carbon emissions.
• Integration of Renewable Energy: AI assists in incorporating renewable energy sources like solar and wind into the power grid. It predicts their output based on weather conditions, ensuring a stable energy supply while maximizing the use of clean energy.
• Environmental Monitoring and Management: AI can also keep track of soil, water, and air quality. For instance, in cities like Shanghai, the implementation of AI for monitoring industrial emissions has significantly improved air quality by enabling more effective pollution-control measures.

Proposed Solutions for Green AI

Green AI is an emerging movement that focuses on minimizing AI’s environmental impact while also leveraging AI advancements to solve environmental challenges.

It emphasizes energy efficiency, minimal carbon emissions, and responsible resource use throughout the AI lifecycle. It also aims to integrate sustainability into every stage of AI development — from research and model training to deployment and maintenance.

A report by Boston Consulting Group and Google found that the adoption of Green AI practices could lead to significant reductions in greenhouse gas emissions — potentially mitigating them by 5%-10% globally by 2030 if used wisely.

Some of Green AI’s proposed solutions include:

• Lightweight Models: Developing algorithms that require less computational power can significantly lower energy consumption. Techniques like pruning, quantization, and model distillation help in creating smaller models that maintain performance while using fewer resources. A good example of this is DeepSeek, an advanced AI platform developed by a Chinese startup that was reportedly built using NVIDIA’s less advanced H-800 chips.
• Energy-Efficient Hardware: Utilizing specialized hardware such as GPUs and Tensor Processing Units (TPUs) that provide more computations per watt can reduce overall energy usage. Additionally, edge computing allows for data processing closer to its source, minimizing energy-intensive data transfers.
• Renewable Energy Sources: Transitioning data centers to use renewable energy sources is crucial. Companies like Google and Microsoft are leading initiatives to power their operations with clean energy, which significantly decreases the carbon footprint of AI technologies.
• Greening Intelligence: Designing algorithms with sustainability in mind ensures that AI applications inherently prioritize energy efficiency and minimal carbon emissions throughout their lifecycle.

Policy Interventions: EU’s AI Act

Although over 60 countries around the world have presented AI policy initiatives, just a few of those policies address environmental concerns. Perhaps one of the most notable regulatory frameworks passed so far is the EU AI Act.

Established by the European Union to govern AI technologies, the Act is meant to encourage safety, transparency, and ethical standards in AI deployment. In regards to environmental concerns, the EU AI Act encourages companies to take the following steps:

• Energy Consumption Reporting: Providers of general-purpose AI models, such as large language models, are required to track and report their energy consumption during training and operation. This aims to establish benchmarks for energy efficiency and promote accountability within the industry,
• Lifecycle Assessment: High-risk AI systems must log energy consumption and assess their environmental impact from design through deployment. This includes adherence to standards that improve energy efficiency and reduce resource use.
• Regulatory Sandboxes: The Act encourages the creation of regulatory sandboxes for testing AI systems, allowing for innovation while ensuring that environmental protection measures are considered, particularly concerning biodiversity and climate change mitigation.

Despite these positive steps, critics argue that the EU AI Act could do more to enforce mandatory environmental standards rather than relying on voluntary compliance. Some provisions have been described as insufficiently robust, potentially limiting their effectiveness in significantly reducing AI's environmental footprint.

Conclusion

As we navigate the complexities of AI's environmental impact, it becomes clear that innovation must go hand in hand with sustainability. While AI holds the promise of solving pressing global challenges, its development and operation can lead to significant carbon emissions and electronic waste.

To create a sustainable future, we must prioritize eco-friendly practices in AI development, such as using renewable energy and efficient algorithms. By making conscious choices now, we can harness the power of AI while protecting our planet for generations to come.