AI

The Artificial Intelligence/Sustainability Paradox

IMC University of Applied Sciences Krems, Austria
Roman Mesicek

SAG Part 3 Basics

Full reference list available at GitHub

The Paradox We Face

One ChatGPT query uses electricity to light a bulb for 20 minutes (de Vries, 2023).

vs.

AI could reduce global emissions by 5 billion tons annually (BCG, 2024).


Which number matters more?

Photo by Yumu on Unsplash

Our Journey Today

Act I: What Is

The Growing Environmental and Social Cost of AI

Act II: What Could Be

AI as Societal Solution Catalyst

Act III: New Bliss

Navigating the Paradox Responsibly

Sources: Duarte, 2010

Act I: What Is

The Growing Societal Costs of AI

Energy • Carbon • Resources • Society

AI's Energy Reality

Current State (2025)
Data Centers 3-4% global electricity
AI Workloads 10-15% of data center energy
Growth Rate 30% Compound Annual Growth Rate (CAGR) through 2030
2030 Projection 415 → 945 TWh

By 2030: AI = Argentina's entire electricity consumption

Sources: IEA, 2025; Andrae & Edler, 2015

The Cost of Intelligence

Training Energy Consumption

Model Energy CO₂ Emissions Equivalent
GPT-3 1,287 MWh 552 tons 120 cars/year
GPT-4 ~51,800 MWh 13,000 tons 2,800 cars/year
BLOOM 433 MWh 50.5 tons 11 cars/year
Llama 3 2,700 MWh 1,090 tons 240 cars/year

GPT-4 training = Annual consumption of 13,000-17,000 EU households
Sources: Patterson et al., 2022; Luccioni et al., 2022; Besiroglu et al., 2024

Every Query Counts

Inference at Scale

Daily Reality:

  • Google Search: 0.3 Wh
  • ChatGPT Query: 2-3 Wh (10x increase)
  • ChatGPT Daily: 564 MWh (200M users)

Water Cost per Conversation:

  • 20-50 questions = 500ml water
  • 300M weekly users = swimming pool daily

Hidden Multiplier: Each response generates 3-5x more

Sources: de Vries, 2023; Li et al., 2023

Corporate Carbon Reality

Emissions Growth Despite Pledges

Company Emissions Change Cause
Google +48% since 2019 14.3M tons total
Microsoft +29% since 2020 "AI acceleration"
Meta +66% since 2019 Data center expansion

"The exponential growth in AI is straining our net-zero commitments" (Microsoft, 2024)

Sources: Google, 2024; Microsoft, 2024; Meta, 2024

The Water Crisis Nobody Discusses

AI's Hidden Thirst

Training Requirements:

  • GPT-3: 700,000 liters cooling water
  • Microsoft Azure: 6.4 million m³ annually (+34% YoY)
  • Google: 6.1 billion gallons in 2023

Geographic Conflict:

  • 160+ data centers in water-stressed regions
  • Arizona, Singapore, Northern Virginia hotspots
  • Competition with agriculture and communities

Per Query: 16.9ml water evaporated

Sources: Li et al., 2023; Mytton, 2021; Microsoft, 2024

Hardware's Dark Side

The Physical Foundation

Critical Materials:

  • 17 rare earth elements per GPU
  • 80% supply from China
  • <1% recycling rate for AI hardware

E-Waste Trajectory:

  • 2025: 54 million tons globally
  • 2030: 1.2-5M tons from AI alone
  • GPU replacement: every 1-3 years

Mining Impact:

  • Dysprosium, Terbium, Cobalt extraction
  • Child labor in mining, (e.g. proven cases in Cobalt mines Democratic Republic of the Congo)
  • Water pollution in rare earth processing
Sources: Crawford & Joler, 2018; USGS, 2024; Forti et al., 2020

The Rebound Effect in AI

Why Efficiency Gains Fail

Historical Pattern:

  • Steam engines: 100x efficiency → 1000x coal use
  • Cars: 30% fuel efficiency → 50% more driving

AI's Paradox (2017-2023):

  • 44x efficiency improvement per query
  • 100x increase in total AI workloads
  • Net result: 2x total energy consumption

The Trap: Lower costs → More use cases → Higher total impact

Sources: Koomey et al., 2024; Jevons, 1865

The Social Cost of AI

Digital Inequality Deepens

Global AI Divide:

  • 90% of AI researchers in 10 countries
  • 2.6 billion people remain offline
  • 72% of data centers in high-income nations

Labor Disruption:

  • 47% of jobs at high automation risk
  • 450,000 fossil fuel workers displaced
  • 87% lack re-training access

Environmental & Social Injustice:

  • Data centers in 65% minority communities (USA)
  • 70% e-waste exported to Global South
  • Low-income pay 3x more per kWh for AI services
Sources: World Bank, 2024; ILO, 2024; Brewster et al., 2023

Act II: What Could Be

AI as Solution Catalyst

Modeling • Energy • Agriculture • Innovation

Climate Modeling Revolution

From Months to Hours

Speed Transformation:

  • Traditional: 90 days for 1000-year simulation
  • AI-Enhanced: 12 hours
  • Impact: Test 100x more scenarios in the same time

Breakthrough Applications:

  • Hurricane paths: 85% accuracy at 7 days
  • Regional downscaling: 10km → 1km resolution
  • Tipping point detection: advance prediction with scenarios

Estimated Market Growth: €266M (2024) → €1.2B (2029)

Sources: Bi et al., 2023; University of Washington, 2024

Smart Grid Transformation

Real Impact at Scale

Deployment Reality:

  • 74% of utilities using AI (2024)
  • 20% error reduction in demand forecasting
  • 15% peak load reduction

Case Studies:

  • DeepMind + UK Grid: 10% efficiency gain
  • Duke Energy + AWS: 30% outage reduction
  • Austria Verbund AG: 18% efficiency gain

Renewable Integration:

  • Wind prediction: 20% value improvement
  • Solar forecasting: 92% accuracy
  • Battery optimization: 40% lifetime extension
Sources: IEA, 2024; DeepMind, 2024; Austrian Research Promotion Agency, 2024

Agricultural Revolution

Beyond Yield Optimization

Water & Chemical Reduction:

  • Irrigation: 40% water savings (Israel)
  • Pesticides: 97% reduction potential (ecoRobotix)
  • Fertilizer: 30% reduction with same yield

Real Deployments:

  • John Deere: 300,000 connected machines
  • Full Nature Farms: Serving 40% of US tomatoes
  • India: 7 million farmers on AI advisory

Food Waste: AI reduces 20% of 1.3B tons annual waste

Sources: BCG, 2024; ecoRobotix, 2024; FAO, 2024

Transportation & Logistics

Immediate Carbon Wins

Google's Fleet Management Impact (2024):

  • Fuel-efficient routing: 2.7M tons CO₂ saved
  • Green Light (traffic): 30% fewer stops
  • Contrails reduction: 54% with American Airlines

Logistics Optimization:

  • UPS: 100,000 tons CO₂/year
  • DHL: 15% fuel reduction
  • Maersk: 20% empty container reduction

Urban Impact: 700+ cities alreadyusing AI traffic management

Sources: Google, 2024; UPS, 2024; C40 Cities, 2024

Material Science Breakthroughs

Accelerating Green Innovation

Discovery Speed:

  • Traditional: 10-20 years per material
  • AI-Assisted: Weeks to months

Recent Breakthroughs:

  • DeepMind GNoME: 2.2M new crystals (45x previous knowledge)
  • Microsoft + PNNL: New battery in 80 hours
  • MIT: Concrete that stores energy

Carbon Capture Materials:

  • 1000+ new MOFs identified
  • 10x improvement in CO₂ absorption
  • Cost reduction: 70% projected
Sources: Merchant et al., 2023; Microsoft Research, 2024; MIT, 2024

Biodiversity & Conservation

AI as Nature's Guardian

Wildlife Protection:

  • WWF Kenya: 96% reduction in elephant poaching
  • Rainforest Connection: Illegal logging down 70%
  • Ocean monitoring: 75% of illegal fishing detected

Species Discovery:

  • BioCLIP: 300+ new species identified
  • iNaturalist: 150M observations, 400K species
  • eBird: 1 billion bird observations analyzed

Ecosystem Monitoring:

  • Amazon: Daily deforestation alerts
  • Coral reefs: Bleaching prediction 2 weeks advance
  • Migration patterns: Climate adaptation tracking
Sources: WWF, 2024; Global Fishing Watch, 2024; Tuia et al., 2024

The Net Impact Equation

Quantifying the Balance

Potential by 2035:

Application Area Annual CO₂ Reduction
Energy Systems 1.2 GtCO₂
Transportation 0.8 GtCO₂
Manufacturing 0.6 GtCO₂
Buildings 1.0 GtCO₂
Agriculture 0.4-1.4 GtCO₂
Total Potential 4.0-5.0 GtCO₂

**AI's Own Footprint:** 0.4-1.6 GtCO₂

Net Benefit: 2.4-4.6 GtCO₂ (3-8x positive)

Sources: BCG, 2024; PwC, 2024; WEF, 2024

Act III: New Bliss

Solutions • Governance • Future Pathways

Technical Solutions Today

Efficiency Without Rebound

Architecture Innovation:

  • Model pruning: 90% size reduction
  • Quantization: 75% memory savings
  • Knowledge distillation: 96% inference energy cut
  • DeepSeek R1: Outperforms at 1/10th size

Hardware Evolution:

  • TPUs: 80% less energy than GPUs
  • Neuromorphic chips: 100x efficiency
  • Optical computing: 1000x potential

Edge Computing: 60% reduction in transmission energy

Sources: Liu et al., 2025; Google, 2024; Intel, 2024

Green Infrastructure

Available Now, Underdeployed

Cooling Revolution:

  • Liquid cooling: 40% energy reduction
  • Stockholm: Heats 25,000 homes with waste heat
  • Free cooling: 65% of year in Austria

Renewable Integration:

  • Google: 64% carbon-free energy
  • Microsoft: 100% renewable by 2025
  • 12 GW clean energy PPAs contracted

Location Matters:

  • Iceland: 100% renewable, natural cooling
  • Austria: 78% renewable grid advantage
Sources: Uptime Institute, 2024; RE100, 2024; Austrian Energy Agency, 2024

Circular AI Economy

From Linear to Regenerative

Linear Model (Current):
Extract → Manufacture → Use (2-3 years) → Dispose

Circular Model (Achievable):
Design for longevity → Modular upgrades → Reuse → Refurbish → Recycle

Impact Potential:

  • 70% material reduction
  • 7-10 year hardware lifecycle
  • 90% recovery rate possible
  • 60% cost savings long-term
Sources: Ellen MacArthur Foundation, 2024; EU Commission, 2024

EU Regulatory Framework

From Voluntary to Mandatory

Corporate Sustainability Reporting Directive (CSRD):

  • 50,000 companies affected
  • Scope 3 emissions (includes AI services)
  • First reports: 2026 for 2025 data

EU AI Act Environmental Provisions:

  • High-risk AI: Mandatory impact assessment
  • Energy consumption disclosure required
  • A-G efficiency rating system

Digital Product Passport (2027):

  • Full lifecycle tracking
  • 90% recyclability target
  • Right to repair provisions
Sources: European Commission, 2024; EU AI Act, 2024; EU, 2024

The Doughnut Perspective

Safe and Just Operating Space

Ecological Ceiling - AI Impact:

  • Climate: 0.5-1% of global emissions
  • Freshwater: 0.2% of consumption
  • Chemical pollution: 5% annual growth in e-waste

Social Foundation - AI Contribution:

  • Energy access: 200M people via optimization
  • Education: 500M students with AI tutors
  • Health: 1.2B served without doctors

Current Status: Pushing BOTH boundaries simultaneously

Alternative Path: Selective deployment for maximum social benefit within ecological limits

Sources: Raworth, 2017; Stockholm Resilience Centre, 2024; Future Earth, 2024

Just Transition Imperatives

Who Wins, Who Loses?

Current Winners:

  • Tech companies: €2.3T market cap increase
  • Skilled workers: 45% wage premium
  • Early adopters: 30% productivity gains

Current Losers:

  • Displaced workers: 2.3M in routine jobs
  • Global South: 78% of e-waste burden
  • Future generations: Degraded systems inherited

Transition Finance Needed:

  • Reskilling: €450B globally
  • Infrastructure: €1.2T for sustainable data centers
  • Compensation: €230B for affected communities
Sources: ILO, 2024; UNCTAD, 2024; Just Transition Fund, 2024

Three Pathways Forward

The Choice to Make

Path 1: Business-as-Usual
→ 2% global emissions by 2030
→ Deepening inequality
→ Ecological overshoot

Path 2: Efficiency-First
→ Carbon neutral with offsets
→ Rebound effects unchecked
→ Social dimensions ignored

Path 3: Sufficiency & Justice
→ Selective AI deployment
→ Net negative emissions possible
→ Within Doughnut boundaries

Sources: IPCC, 2023; Future Earth, 2024

Your Role as Responsible AI Professionals

Five Principles for Action

1. Question Necessity
"Do we need AI for this?" before "How can AI do this?"

2. Design for Longevity
Build systems that last 10 years, not 2

3. Measure Full Impact
Carbon + Water + Materials + Social costs

4. Ensure Fair Distribution
Benefits to those who bear the costs

5. Challenge Power Structures
Democratic governance of AI deployment

Sources: IEEE, 2024; AI Now Institute, 2024

Resolving the Paradox

"Green AI isn't about choosing between innovation and sustainability. It's about ensuring every joule we invest in intelligence returns dividends in planetary health."

Speaker notes: - Image suggestion: Split image - data center on left, flourishing forest on right - Opening question for audience: "How many of you used AI today? Now estimate the energy cost." - Build tension between innovation promise and environmental reality

Speaker notes: - Pause after revealing each statement for impact - This paradox frames our entire discussion - Audience reflection: "What's your initial instinct?" - We'll resolve this tension by session end

Speaker notes: - Three-act structure borrowed from narrative theory (Duarte sparkline) - Each act challenges the previous understanding - Interactive element: Vote now - is AI net positive or negative for environment? - We'll vote again at the end

Speaker notes: - Transition: "Let's start with uncomfortable truths" - This section may challenge tech optimism - Prepare for cognitive dissonance

Speaker notes: - Image suggestion: Exponential growth curve chart - Context: This is BEFORE mass AI adoption - Question: "Is exponential growth sustainable in finite systems?"

Speaker notes: - GPT-4 is 40x larger than GPT-3 - Training cost: $100 million in compute alone - Discussion: "Should we require energy labels for AI models?" - Note efficiency paradox: Better models require MORE energy

Speaker notes: - Image suggestion: Infographic comparing search vs AI query - Calculation: Your daily AI use = ? - Most users unaware of resource cost - Question: "Should AI interfaces show environmental cost?"

Speaker notes: - Contradiction: Net-zero promises vs AI investments - Image suggestion: Chart showing pledge vs reality divergence - Discussion: "Can companies be carbon neutral while scaling AI?"

Speaker notes: - Image suggestion: Map overlay of water stress and data centers - Meta's Goodyear facility: 56M gallons/year - Question: "Should data centers be in deserts?" - Hidden: Most cooling water evaporates, doesn't return

Speaker notes: - Image suggestion: Breakdown of materials in a single GPU - Salar de Uyuni: 1/3 of world's lithium - Question: "Is 'clean' tech really clean?" - E-waste: 54M tons = 10.8M elephants

Speaker notes: - This is THE critical concept many miss - Efficiency ≠ Sustainability - Discussion: "Can we have sustainable growth?" - Connect to Session 1: Planetary boundaries are absolute

Speaker notes: - Image suggestion: World map showing AI capability vs GDP - Digital colonialism: Extract data, export waste - Question: "Who benefits from AI advancement?" - Connect to Doughnut's social foundation

Speaker notes: - Transition: "Now the other side of the paradox" - Same technology, different application - Focus on empirical evidence, not promises

Speaker notes: - Image suggestion: Side-by-side simulation comparison - This changes policy response time dramatically - Question: "What decisions need faster climate data?" - Note: AI doesn't replace physics, enhances it

Speaker notes: - This is happening NOW, not future - Austria's 78% renewable grid = perfect testbed - Discussion: "Why isn't this bigger news?" - Economics driving adoption

Speaker notes: - Image suggestion: Before/after field comparison - 97% pesticide reduction = game changer - Question: "Why do we still use blanket spraying?" - Note: Technology exists, adoption is barrier

Speaker notes: - These are REALIZED savings, not projections - Simple algorithm changes = massive impact - Discussion: "Should fuel-efficient routing be mandatory?" - Every saved mile counts at scale

Speaker notes: - This is chemistry's "AlphaGo moment" - New materials enable entire green economy - Question: "What material would change everything?" - Connect to circular economy potential

Speaker notes: - Image suggestion: Thermal camera elephant detection - AI democratizes conservation science - Discussion: "Can technology save what technology destroyed?" - Every species matters for ecosystem stability

Speaker notes: - BUT: This is NOT automatic - Requires intentional deployment - Question: "Why aren't we achieving this?" - Currently <10% of AI focused on climate

Speaker notes: - Synthesis: Both sides are true - The path forward requires nuance - Your generation will decide the outcome

Speaker notes: - Efficiency IS possible without growth - DeepSeek: Chinese breakthrough changes everything - Discussion: "Should we regulate model sizes?" - Note: Mixture of Experts architecture

Speaker notes: - Image suggestion: Stockholm district heating diagram - Austria uniquely positioned for green AI - Question: "Why build data centers in deserts?" - Policy could incentivize right locations

Speaker notes: - Fairphone model for servers is possible - Image suggestion: Circular flow diagram - Discussion: "Why do we accept planned obsolescence?" - Connect to R-strategies from Session 2

Speaker notes: - Austria must transpose by June 2025 - Penalties: Up to 5% global turnover - Question: "Is regulation the answer?" - Your careers will implement these

Speaker notes: - Image suggestion: Doughnut with AI impacts mapped - We're overshooting ceiling AND failing foundation - Discussion: "Can AI help us thrive in the safe space?" - Key metric: Social Return on Carbon Investment

Speaker notes: - Transition happens regardless - just or unjust? - Austrian example: Styrian coal region transition - Question: "Who should pay for transition?" - Connect to CSR from Session 2

Speaker notes: - No predetermined outcome - Your generation decides - Discussion: "Which path is Austria taking?" - Individual AND systemic action needed

Speaker notes: - These are YOUR tools for change - Every line of code is a choice - Question: "Which principle resonates most?" - Connect to personal agency

Speaker notes: - Return to opening vote - has it changed? - The paradox is false - we need AND thinking - Your homework: Calculate your AI carbon footprint - Final question: "What will you do differently tomorrow?" - Image suggestion: Balanced scale with AI and Earth