Introduction: The Urgency of Sustainability in the Age of AI

Green AI represents one of the most crucial paradigms of 2025, emerging as a necessary response to the explosive growth of artificial intelligence and its environmental impact. Green AI is more environmentally friendly and inclusive than conventional AI, as it not only produces accurate results without increasing computational costs, but also ensures that technological innovation goes hand in hand with environmental responsibility.

‍

The urgency of this approach is demonstrated by the most recent data: according to MIT News, data center energy requirements in North America have increased from 2,688 megawatts at the end of 2022 to 5,341 megawatts at the end of 2023, driven in part by the demands of generative AI. More significantly, according to MIT Technology Review, 4.4 percent of all energy in the United States now goes to data centers, with the carbon intensity of electricity used by data centers being 48 percent higher than the U.S. average (according to a study by the Harvard T.H. Chan School of Public Health).

‍

The Environmental Impact of AI: An Evolving Crisis.

Explosive Energy Consumption

The growth of AI has led to a dramatic change in the global energy landscape. According to the MIT Technology Review, as of 2018, data centers now account for 4.4 percent of total demand, up from 1.9 percent in 2018. Future projections are even more alarming: according to a report by the International Energy Agency, data center electricity consumption is projected to more than double by 2030.

‍

Generative AI models are pushing these numbers upward. As highlighted by MIT News, training a generative AI cluster could consume seven or eight times more energy than a typical computational workload. To put this in context, training GPT-3 consumed 1,287 megawatt hours of electricity (enough to power about 120 average American homes for a year), generating about 552 tons of carbon dioxide.

‍

The Escalation of Hardware Power

The race toward more and more powerful models has led to an escalation in hardware power. According to Deloitte, GPUs for AI were running at 400 watts until 2022, while state-of-the-art GPUs for generative AI in 2023 are running at 700 watts, and next-generation chips in 2024 are expected to run at 1,200 watts. This represents an exponential increase that puts a strain on the global energy infrastructure.

‍

Hardware Solutions for Energy Efficiency

Specialized Chips: The AI Hardware Revolution

The hardware industry's response to the AI power crisis is materializing through increasingly specialized and efficient chips:

Tensor Processing Units (TPUs): According to TechTarget, TPUs are ASICs designed for high computation volume at low precision with multiple input/output operations per joule. TPU v6e is the newest Trillium chip, released in October 2024, with 4.7 times higher peak computational performance per chip than TPU v5e.

‍

Field-Programmable Gate Arrays (FPGAs): As highlighted by IBM, FPGAs are also well suited for tasks that value energy efficiency over processing speeds and offer the flexibility to accommodate the rapid evolution of AI algorithms.

‍

Application-Specific Integrated Circuits (ASICs): According to Geniatech, ASICs offer the advantages of low power consumption, speed and a small footprint, representing the most efficient solution for specific, high-volume AI workloads.

The Emergence of Edge AI

A crucial trend for sustainability is the movement toward edge computing. According to Geniatech, Gartner predicts that by 2025, edge computing will process 75 percent of the data generated by all use cases, significantly reducing the need for data transmission to centralized data centers and the associated energy.

‍

Progress in Energy Efficiency

Dramatic Improvements in Performance per Watt.

The industry is making significant progress in energy efficiency. According to NVIDIA, it has achieved a 10,000x efficiency gain in AI training and inference from 2016 to 2025, demonstrating the potential for dramatic improvements.

‍

However, the reality is more complex. As reported by David Mytton in his DeVSustainability blog, dual-socket servers now consume between 600-750 W, up from 365 W in 2007-2023, indicating that while efficiency per operation improves, the total power of systems continues to grow.

Software and Architectural Optimizations

Software strategies are emerging as a crucial complement to hardware improvements:

Model Optimization: Quantum models tend to require significantly fewer parameters for training than their classical counterparts, suggesting alternative approaches to reduce computational complexity.

‍

Intelligent Energy Management: According to MIT Sloan, limiting usage to 150 or 250 watts (about 60 percent to 80 percent of their total power) depending on the processor used not only reduces the overall power consumption of workloads; it also reduces operating temperatures.

‍

Corporate Initiatives and Environmental Commitments

Microsoft: Leadership in Carbon Negativity

Microsoft has made one of the most ambitious commitments in the industry. As stated in Microsoft's official 2020 blog, "By 2025, we will move to 100 percent renewable energy supply, which means we will have power purchase agreements for contracted green power for 100 percent of the carbon-emitting electricity consumed by all of our data centers, buildings, and campuses."

‍

The company has also established a new $1 billion climate innovation fund to accelerate the global development of carbon reduction, capture and removal technologies.

However, reality presents challenges. According to GeekWire, Microsoft emitted more than 15.4 million metric tons of carbon dioxide equivalent last year, with scope 3 emissions accounting for more than 96 percent of its carbon footprint.

Google: Pioneer in Renewable Energy

Google has achieved significant milestones in sustainability. According to the official Google Sustainability website, "In 2017, Google committed to equalizing 100 percent of its energy consumption with renewable energy. This goal was successfully achieved in 2020."

‍

The company continues to invest heavily: according to the Google 2025 sustainability report, "In 2024, we contracted an additional 19 GW of new renewable energy in 16 countries, and expanded into nuclear power with the signing of our first large-scale nuclear power purchase agreement."

Amazon Web Services: The Challenge of Scale

AWS, the world's largest cloud provider, has set ambitious goals but faces criticism for transparency. According to Climatiq, "Amazon announced 18 new wind and solar power projects in AWS regions in the United States, Finland, Germany, Italy and the United Kingdom, totaling 5.6 GW of new renewable energy capacity."

‍

However, according to the same source, "criticisms of AWS's carbon footprint reporting can be summarized as not granular, not transparent, and not useful for technology teams seeking to optimize their use."

Frameworks and Tools for Green AI.

Emission Monitoring Tools

Industry has developed several tools to monitor and reduce the environmental impact of AI:

CarbonTracker and CodeCarbon: According to Carbon Credits, "Some of the tools used to estimate the carbon footprint of AI technologies are CarbonTracker, CodeCarbon, Green algorithms and PowerTop."

eco2AI: As described in Doklady Mathematics, "eco2AI is an open-source library capable of tracking carbon-equivalent emissions during the training or inference of Python-based AI models, taking into account the energy consumption of CPU, GPU, RAM devices."

Green AI Implementation Framework

According to a study published in Industry Science, "A sophisticated Python-based tool tailored to track and manage the carbon footprint of training machine learning models and other computational tasks represents the evolution toward more sophisticated tools for AI sustainability management."

‍

Regulation and Public Policy

The EU AI Act: A Global Model.

The European Union has taken the lead in regulating sustainable AI. According to the European Parliament, "In June 2024, the EU adopted the world's first rules on AI. The Artificial Intelligence Act will be fully applicable 24 months after it enters into force."

‍

Significantly, according to the Green Software Foundation, "the EU AI Act gains further significance when viewed in the context that the EU is currently one of the world's leaders in environmental and climate policies."

Regulatory Gaps and Recommendations

Despite progress, significant gaps persist. As highlighted in a paper published in arXiv, "Current proposals for AI regulation, in the EU and beyond, aim to stimulate reliable (e.g., AI Act) and responsible (e.g., AI Liability) AI. What is missing, however, is a robust regulatory discourse and roadmap for making AI, and technology more generally, environmentally sustainable."

‍

Experts propose concrete solutions: according to a report by the Tony Blair Institute for Global Change, "Establish and adopt best-practice metrics for data center energy consumption and carbon emissions while isolating AI-related information."

‍

Future Perspectives and Challenges

Energy Growth Projections

Projections for the future are simultaneously worrisome and full of opportunity. According to IDC, "Global data center electricity consumption will more than double between 2023 and 2028 with a five-year CAGR of 19.5 percent and reaching 857 Terawatt hours (TWh) in 2028."

‍

More specifically for AI, according to the same IDC report, "AI data center energy consumption is expected to grow at a CAGR of 44.7 percent, reaching 146.2 Terawatt hours (TWh) by 2027."

Innovative Approaches to Sustainability

Innovative approaches are emerging such as Sustain AI, described in an MDPI paper as "a multi-modal deep learning framework that integrates Convolutional Neural Networks (CNNs) for defect detection, Recurrent Neural Networks (RNNs) for predictive modeling of energy consumption, and Reinforcement Learning (RL) for dynamic energy optimization."

‍

Green-in AI vs Green-by AI: Two Complementary Paradigms.

Green-in AI: Intrinsic Optimization.

The "Green-in AI" paradigm focuses on designing inherently more efficient algorithms and models. According to a review published in ScienceDirect, these are "strategies for designing more energy-efficient machine learning algorithms and models by focusing on optimizing hardware and software."

‍

Green-by AI: AI for Sustainability

The "Green-by AI" paradigm leverages AI to improve sustainability in other sectors. According to the same ScienceDirect review, it represents "AI approaches to improve eco-friendly practices in other sectors by using artificial intelligence to optimize energy efficiency in outdoor applications."

‍

Conclusions: Toward a Sustainable Future of AI

Green AI represents a fundamental transformation in the way we conceive and implement artificial intelligence. The 2025 data show that we are at a tipping point: The growth in electricity demand for AI and data centers is a test case for how society will respond to the demands and challenges of broader electrification.

‍

Emerging solutions-from specialized hardware to quantum computing, from monitoring frameworks to regulatory policies-offer a path to sustainability. However, success will depend on the industry's ability to balance innovation with environmental responsibility, ensuring that AI can become the key driving force behind the global effort to achieve carbon neutrality.

‍

The year 2025 looms as a crucial year in which decisions made today will determine whether AI will be part of the climate problem or part of its solution. Green AI is no longer an option, but an imperative necessity for a technologically advanced and environmentally sustainable future.

‍

Frequently asked questions

What is Green AI?

Green AI is a technological paradigm that aims to make artificial intelligence more environmentally friendly and sustainable. It focuses on producing accurate results without increasing computational costs, ensuring that technological innovation goes hand in hand with environmental responsibility.

Why is Green AI so urgent in 2025?

The urgency stems from the explosive environmental impact of AI. Data from 2025 show that:

• Data center energy requirements in North America increased from 2,688 megawatts (end 2022) to 5,341 megawatts (end 2023)
• 4.4 percent of all energy in the United States goes to data centers
• Carbon intensity of data center electricity is 48% higher than the U.S. average

How much energy does AI actually consume?

The energy impact of AI is dramatic:

• Since 2018, data centers have increased from 1.9% to 4.4% of total energy demand
• Training a generative AI cluster can consume 7-8 times more energy than a typical computational load
• GPT-3 training consumed 1,287 megawatt hours (enough for 120 U.S. homes for one year)
• Data center consumption is expected to double by 2030

How is the hardware evolving to be more efficient?

The industry is developing specialized chips:

Tensor Processing Units (TPU): TPU v6e offers 4.7 times higher performance than v5e

Field-Programmable Gate Arrays (FPGAs): Optimized for energy efficiency and flexibility

Application-Specific Integrated Circuits (ASICs): Offer low power consumption, speed and small footprint

However, total power continues to grow: GPUs have increased from 400W (2022) to 700W (2023), with forecasts of 1,200W by 2024.

‍

What is Edge AI and why is it important for sustainability?

Edge AI processes data locally instead of sending it to centralized data centers. Gartner predicts that by 2025, edge computing will process 75 percent of the data generated, significantly reducing the energy consumption associated with data transmission.

‍

What progress has been made in energy efficiency?

NVIDIA has achieved a 10,000x efficiency improvement in AI training and inference from 2016 to 2025. However, modern servers consume 600-750W compared to 365W from 2007-2023, showing that while efficiency per operation improves, total power continues to increase.

‍

What are big tech companies doing for sustainability?

Microsoft: Committed to 100% renewable energy by 2025 and 1 billion fund for climate innovation. However, emitted 15.4 million tons of CO2 equivalent in 2024.

Google: Achieved 100% renewable energy in 2020 and contracted 19 GW of new renewable energy in 16 countries by 2024.

Amazon AWS: Announced 18 new renewable projects for 5.6 GW, but is criticized for lack of transparency in reporting.

‍

Are there tools to monitor the environmental impact of AI?

Yes, several tools are available:

• CarbonTracker and CodeCarbon: To estimate the carbon footprint.
• eco2AI: Open-source library for tracking emissions during training and inference
• Green algorithms and PowerTop: More specialized monitoring tools.

How is Green AI regulated?

The EU took the lead with theEU AI Act, adopted in June 2024-the world's first AI rules that will be fully applicable after 24 months. However, experts point to gaps in the regulatory discourse to make AI environmentally sustainable.

What are the future projections for AI energy consumption?

The forecast is alarming:

• Global data center consumption will double between 2023 and 2028 (CAGR 19.5%)
• It will reach 857 TWh in 2028
• AI-specific energy consumption will grow with CAGR of 44.7%, reaching 146.2 TWh by 2027

What is the difference between green-in AI and green-by AI?

Green-in AI: Focuses on designing inherently more energy-efficient algorithms and models by optimizing hardware and software.

Green-by AI: Use AI to improve sustainability in other areas by leveraging artificial intelligence to optimize energy efficiency in outdoor applications.

‍

Why is 2025 considered a crucial year for Green AI?

The year 2025 represents a turning point where decisions made today will determine whether AI will be part of the climate problem or part of its solution. The growth of electric demand for AI is a test for how society will respond to the challenges of broader electrification. Green AI is no longer an option, but an imperative necessity for a technologically advanced and environmentally sustainable future.

‍

‍