Training large language models requires massive computational resources. Recent studies have quantified these costs:

• GPT-3 (175B parameters): Estimated 1,287 MWh during training, equivalent to the annual energy consumption of ~130 US homes
• BLOOM (176B parameters): Consumed approximately 433 MWh for training
• PaLM (540B parameters): Training consumed an estimated 2,500 MWh of energy
• Meta's LLaMA-2 70B: Required approximately 1,720 MWh for training

The energy costs scale non-linearly with model size, with each doubling of parameters requiring significantly more than double the energy.

Data centers require substantial water for cooling systems:

• GPT-3 training is estimated to have consumed ~700,000 liters of water for cooling
• A single data center can consume 1-5 million liters of water daily
• Microsoft reported 34% increase in water consumption from 2021-2022, largely attributed to AI infrastructure
• Google's water usage increased by 20% in 2022, with AI workloads being a significant factor

• GPT-3: ~3,640 petaflop-days of compute
• GPT-4: Estimated 50,000-100,000 petaflop-days (rumored, not officially disclosed)
• Training duration: From several weeks to several months depending on model size and infrastructure
• Hardware costs: NVIDIA A100 GPU clusters can cost $10M-$100M+ for training facilities

Running inference on LLMs incurs costs for every token generated:

• GPT-4: ~$0.03-0.06 per 1K tokens (input), $0.06-0.12 per 1K tokens (output)
• GPT-3.5: ~$0.0015-0.002 per 1K tokens
• Claude 3 Opus: ~$0.015 per 1K input tokens, $0.075 per 1K output tokens
• Open-source models (self-hosted): $0.001-0.01 per 1K tokens depending on infrastructure

• Single ChatGPT query: ~0.002-0.003 kWh (approximately 2-3 Wh)
• Daily operational cost for ChatGPT estimated at $100,000-700,000 (energy + infrastructure)
• Scaling estimates: 10 billion queries/day would require ~20-30 MW continuous power

• Training GPT-3: Estimated 552 metric tons CO2e (equivalent to 120 cars driven for a year)
• BLOOM training: ~25 metric tons CO2e (significantly lower due to French nuclear power grid)
• Full lifecycle emissions (including hardware manufacturing): 2-5x higher than training alone

• Training one large model: Equivalent to 5x the lifetime emissions of an average car
• Daily ChatGPT operations: Comparable to a small town's electricity consumption
• Projected AI sector emissions by 2030: 0.5-1.5% of global greenhouse gas emissions

Key research findings:

1. Strubell et al. (2019): "Energy and Policy Considerations for Deep Learning in NLP"
   • Demonstrated that training a single large transformer model can emit as much carbon as five cars in their lifetimes
   • Highlighted the environmental cost of neural architecture search (NAS)
2. Patterson et al. (2021): "Carbon Emissions and Large Neural Network Training"
   • Showed that carbon footprint varies dramatically based on energy grid composition
   • Proposed using carbon-aware computing to reduce emissions by 100-1000x
3. Luccioni et al. (2023): "Power Hungry Processing: Watts Driving the Cost of AI Deployment?"
   • First comprehensive study of inference costs across multiple model families
   • Found that image generation models have significantly higher per-query costs than text models

• Model distillation: Reduce parameters by 10-100x while maintaining 95%+ performance
• Quantization: 4-bit and 8-bit models reduce memory and compute by 2-4x
• Sparse models: Mixture-of-Experts (MoE) architectures activate only 10-20% of parameters per query
• Pruning: Remove redundant weights to reduce model size by 30-70%

• Carbon-aware scheduling: Run training jobs when renewable energy is available
• Geographic optimization: Locate data centers in regions with clean energy grids
• Liquid cooling: Reduce water consumption by 20-30% compared to traditional cooling
• Custom accelerators: Google TPUs, AWS Trainium offer 2-5x better price/performance

• Batch processing: Amortize fixed costs across multiple requests
• Caching: Store and reuse common completions
• Prompt optimization: Reduce token counts through efficient prompt engineering
• Model selection: Use smallest model sufficient for task (GPT-3.5 vs GPT-4)
• Local deployment: Self-host smaller models for privacy and cost reduction

• Retrieval-Augmented Generation (RAG): Reduce model size requirements
• Parameter-efficient fine-tuning (PEFT): LoRA, QLoRA reduce training costs by 100x
• Edge deployment: Move inference to user devices
• Specialized models: Task-specific smaller models replacing general-purpose large ones

• Green AI movement: Developing energy-efficient architectures
• Renewable energy commitments: Major providers targeting 100% renewable energy
• Carbon offset programs: Companies purchasing carbon credits for AI operations
• Efficiency reporting: Standardized metrics for comparing model environmental costs