The Dual Nature of Artificial Intelligence

• Operational carbon splits into training and inference emissions.
• Training a model like GPT-3 emitted over 550 metric tons of CO2e ^[2][3].
• Newer models likely exceed these figures by orders of magnitude.

• Inference constitutes the majority of lifecycle emissions for deployed models ^[1][2].
• Energy-intensive models consume over 29 Wh per long prompt ^[1].
• Efficient models can operate at approximately 0.4 Wh per prompt ^[1].

• Median Gemini App text prompt consumes 0.24 Wh of energy ^[4][5].
• Emissions per prompt are roughly 0.03 grams of CO2e ^[4][5].
• Comparable to running a 60W light bulb for about 14 seconds.

• Data centers require massive cooling systems to maintain operation.
• Early estimates: GPT-3 training consumed ~700,000 liters of freshwater ^[1][6].
• Global AI water demand could reach 6.6 billion cubic meters by 2027 ^[1][6].

• Studies debate the metric of "bottles of water per conversation" ^[7][8].
• Google reports 0.26 mL (five drops) per median query due to efficiency ^[4][5].
• Efficiency gains fight against the sheer scale of deployment.

• "Embodied footprint" includes extraction, manufacturing, and disposal.
• Manufacturing AI accelerators is chemically intensive and energy-demanding ^[11][12].
• 2025 LCA shows efficiency gains can improve Compute Carbon Intensity by 3x ^[11].

• AI hardware becomes obsolete faster than general-purpose servers ^[13][15].
• Global e-waste reached 62 million tonnes in 2022 ^[16].
• Specialized chips contain hazardous materials and rare earth elements ^[16][17].

• Shift toward liquid cooling: Direct-to-chip or immersion cooling ^[5].
• DeepMind-style optimization reduces cooling energy by up to 40% ^[18].
• Balancing Water Usage Effectiveness (WUE) against Power Usage Effectiveness (PUE).

• GraphCast predicts weather 10 days in advance using Graph Neural Networks ^[19].
• Outperforms traditional Numerical Weather Prediction (NWP) in speed and energy efficiency.

• NVIDIA's FourCastNet v3 uses Spherical Fourier Neural Operators ^[20][21].
• Enables rapid ensemble forecasting to predict extreme weather probabilities ^[21].
• Democratizes forecasting: Runs on a few GPUs instead of supercomputers ^[22].

• Integrating variable sources like wind and solar into the grid.
• AI predicts solar irradiance and wind speeds with high precision ^[23].
• Reinforcement Learning manages Hybrid Energy Storage Systems (HESS) ^[23][24].

• Maximizing yield while minimizing chemical inputs and water usage.
• Analysis of satellite imagery and soil sensors for Variable Rate Technology ^[26][28].
• Smart irrigation systems optimize water delivery in real-time ^[27].

Contrasting development philosophies

• Red AI: Buying performance with massive computational cost ^[29][30].
• Green AI: Treating carbon efficiency as a primary evaluation metric.
• Decoupling AI progress from exponential resource consumption.

Techniques to reduce model size and energy consumption

• Quantization: Reducing precision (e.g., 32-bit to 4-bit) to save energy.
• Pruning: Removing redundant weights to create sparse models.
• Knowledge Distillation: Training small student models from large teachers ^[31][32].

• The "Small is Sufficient" trend: Using task-specific models ^[34].
• Adoption could save roughly 28% of global AI electricity by 2025 ^[34].
• Avoiding the use of massive generalist models for simple queries.

• Hardware inspired by the human brain's architecture.
• Spiking Neural Networks (SNNs): Neurons only consume energy when active ^[35][36].
• Eliminates the energy cost of moving data between memory and processors.

• Intel Loihi 2 and BrainChip Akida showing commercial viability ^[36][37].
• Potential for 100x to 1000x efficiency gains in specific tasks ^[36][38].
• Ideal for edge AI and sensory processing applications.

• Passive Acoustic Monitoring (PAM) tracks species via soundscapes ^[44][45].
• Computer vision identifies species in real-time via camera traps ^[44][46].
• Non-invasive methods for tracking elusive or endangered species.

• Mandates transparency regarding "AI systems resource performance" ^[48][49].
• General-Purpose AI providers must publish energy consumption data ^[50][51].
• Applies to models trained with more than 10^23 FLOPs ^[50].

• ISO/IEC 42001: Management system standard including sustainability ^[53].
• ISO/IEC TR 20226: Focuses on energy, water, and e-waste aspects ^[54][55].
• Moving from voluntary "greenwashing" to rigorous compliance.

• Software tools like CodeCarbon and Green Algorithms estimate emissions ^[3][56].
• Compute Carbon Intensity (CCI) quantifies carbon per unit of work ^[11].
• Tools may underestimate true usage by missing overheads like cooling ^[56].

• Integrate carbon tracking into CI/CD pipelines.
• Default to Small Language Models (SLMs) where possible.
• Schedule training runs for times with low grid carbon intensity.

• Require Life Cycle Assessments (LCA) from hardware vendors.
• Prioritize data centers with water-free or recycled water cooling.
• Align reporting with EU AI Act and ISO standards immediately.

• Standardizing water reporting to prevent "water washing".
• Hardware-software co-design for neuromorphic architectures.
• Mitigating Jevons Paradox: Efficiency leading to increased total consumption.