AI summary by TwinMind
How much energy AI really needs. And why that's not its main problem.
Sabine Hossenfelder
8m 6s
AI Energy Consumption Overview
- AI consumes significant energy during training and operation
- Sam Altman (OpenAI CEO) stated need for small modular nuclear reactors to power AI systems
- Must distinguish between training energy vs operational energy requirements
Training Energy Requirements
- GPT-3 training consumed at least 1,300 megawatt hours (2022 estimate)
- Equivalent to powering 130 US homes for entire year
- Large language models have grown significantly larger since then
- GPT-4 training estimated cost: ~$100 million or more
- No exact numbers available from companies
OpenAI-Musk Lawsuit Context
- Elon Musk suing OpenAI over original nonprofit agreement
- OpenAI blog post (early 2017): realized AGI requires "vast quantities of compute"
- Estimated need: "billions of dollars per year" for compute
- Timeline:
- OpenAI originally nonprofit structure insufficient for funding
- Restructured to for-profit model
- Musk wanted OpenAI to become part of Tesla
- OpenAI declined, Musk left 2018
- Microsoft invested 2019
- OpenAI became huge success 2022
- Musk now suing for breach of original contract
Operational Energy Use Study
- December paper from Hugging Face and Carnegie Mellon University
- Preprint (not yet peer-reviewed)
- Tested 88 different AI models across various tasks
- Text generation: few milliwatt hours per task (few watt hours per 1,000 tasks)
- Image generation: ~1,000x higher energy requirement than text
- Few watt hours per image
- Comparable to smartphone charging energy
Global Energy Impact
- Data centers currently: 1-2% of global electricity use
- ~400 terawatt hours per year (equivalent to entire UK consumption)
- International Energy Agency projection by 2026: could double current usage
- Factors: increased AI use + continued crypto mining
Energy Efficiency Efforts
- Dedicated hardware development for AI tasks
- AI used to optimize energy efficiency (feedback loop)
- Example: DeepMind system helped cool Google data centers more efficiently
- Overall expectation: energy use will still increase despite efficiency gains
Future Concerns: Cost and Access Inequality
- Primary concern: cost barrier rather than energy consumption itself
- Building/maintaining large AI systems extremely expensive
- Prediction: few global AI systems owned by:
- Large companies
- Wealthy governments (not wanting corporate dependence)
- Most people will need AI service subscriptions
- Access levels determined by payment capacity
- Potential scenarios requiring enormous compute:
- Theory of everything discovery
- Cancer cure research
- Complex optimization tasks
- Risk: wealth inequality amplification (individual and national levels)
- Rich get richer, poor get poorer due to AI access barriers
