AI's environmental impact is far larger than most companies report, because there's no standard way to measure it. A systematic review of 62 peer-reviewed studies published in March 2026 found that the carbon footprint of artificial intelligence systems is fragmented across algorithms, hardware, manufacturing, and deployment, yet researchers rarely account for all of these layers together. This measurement gap means organizations optimizing AI for energy efficiency at the software level may still be creating massive emissions elsewhere in the system .

The stakes are enormous. Training GPT-3, a large language model with 175 billion parameters, consumed 1,287 megawatt-hours of electricity and generated 502 metric tons of carbon dioxide emissions . More broadly, a recent inventory of 369 large generative AI models built between 2018 and 2024 estimated their combined annual energy consumption between 25 and 41 terawatt-hours, producing 10.7 to 18.6 million tons of carbon dioxide per year . To put this in perspective, global data centers consumed 415 terawatt-hours in 2024, roughly 1.5 percent of all world electricity, and the International Energy Agency projects that demand will double to 945 terawatt-hours by 2030 as AI drives server expansion .

Why Are We Measuring AI's Carbon Footprint So Inconsistently?

The review, conducted according to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) principles, identified 784 unique academic citations and extracted 62 studies that met strict inclusion criteria. Researchers organized these studies into four domains: algorithms for energy efficiency, hardware and accelerators, life-cycle assessment, and operational deployment strategies . The findings reveal a fragmented research landscape where different teams use different methodologies, making it nearly impossible to compare results or understand the true environmental impact of AI systems.

One critical problem is that most organizations focus narrowly on operational energy, the electricity consumed while training or running AI models. However, this approach ignores embodied carbon, the emissions generated during semiconductor fabrication and hardware manufacturing. As companies refresh their equipment more frequently or deploy larger fleets of servers, embodied carbon becomes an increasingly significant portion of total emissions . Without accounting for this, companies report artificially low carbon footprints.

What Are the Four Layers of AI's Carbon Footprint?

• Algorithmic Efficiency: Techniques like pruning (removing unnecessary neural network connections), quantization (reducing numerical precision), and sparsity (using fewer computations) can reduce computational burden, but only deliver real carbon savings when integrated with hardware and data center infrastructure.
• Hardware and Accelerators: The choice of processors, graphics processing units (GPUs), and specialized chips significantly impacts energy consumption, yet manufacturers rarely provide complete lifecycle emission data for their products.
• Life-Cycle Assessment: This includes raw material extraction, chip manufacturing, model development, deployment, operation, and end-of-life hardware disposal, yet most studies examine only one or two stages rather than the complete picture.
• Operational Deployment: Decisions about data center location, carbon-aware scheduling of computations, and placement of workloads between cloud and edge devices create far more variance in real-world emissions than optimizations at the model level.

The review emphasizes that algorithmic improvements alone cannot solve the problem. A model that runs 20 percent faster due to software optimization may still generate massive emissions if it runs on hardware manufactured with high-carbon processes or in a data center powered by fossil fuels .

How to Build Carbon-Aware AI Systems: A Research Agenda

• Standardize Carbon Reporting: Establish consistent methodologies for measuring and reporting carbon footprints across the entire lifecycle, enabling meaningful comparisons between AI systems and organizations.
• Integrate Hardware-Software Co-Optimization: Design algorithms and hardware together from the start, rather than optimizing each independently, to achieve genuine reductions in both energy and embodied emissions.
• Improve Embodied Emission Data: Require semiconductor manufacturers and hardware vendors to publish detailed lifecycle emission data for their products, making it possible for AI developers to make informed purchasing decisions.
• Embed Carbon in AutoML and Scheduling: Incorporate carbon metrics directly into automated machine learning systems and workload scheduling algorithms, so carbon efficiency becomes a primary optimization goal alongside accuracy and speed.

The U.S. Department of Energy estimates that American data centers are on track to consume between 325 and 580 terawatt-hours per year by 2028, with approximately 22 percent of that consumption driven by AI workloads . One recent study suggests that U.S. AI data centers alone could emit between 24 and 44 million metric tons of carbon dioxide per year between 2024 and 2030, a figure significantly higher than many current estimates . These projections make it clear that without immediate action, AI's energy explosion will make it difficult for organizations and governments to meet climate commitments such as the Paris Agreement's 2030 targets.

The review identifies a critical structural gap: research on green AI is scattered across academia and industry, with different teams working on isolated pieces of the puzzle without an overarching framework. Current tools and dashboards for measuring AI carbon footprint do not use consistent methodologies and commonly lack full lifecycle impact assessment . Benchmarking efforts like MLPerf and the AI Energy Score are beginning to tackle energy measurement, but no comprehensive framework exists yet.

The path forward requires collaboration across multiple disciplines. Hardware manufacturers must provide better data on embodied emissions. Software researchers must design algorithms with carbon constraints in mind. Data center operators must implement carbon-aware scheduling that shifts computations to times and locations where renewable energy is available. And organizations deploying AI must demand transparency from vendors and measure their own systems comprehensively. Until these pieces come together, the true environmental cost of artificial intelligence will remain hidden behind incomplete measurements and fragmented reporting standards.