AI and Sustainability

• Total corporate AI investment: $252.3B in 2024, up 44.5% YoY (Stanford HAI)
• VC funding to AI: $131.5B in 2024, capturing ~36% of all global VC (PitchBook)
• U.S. dominates: 57% of global private AI investment ($109.1B), 12x China ($9.3B) (Stanford HAI)
• McKinsey 2025: 88% of organizations regularly use AI (up from 55% in 2023)
• GenAI adoption: 65–71% of organizations in 2024, up from 33% in 2023
• OECD: 20.2% of firms reported using AI in 2025, up from 8.7% in 2023
• IDC: AI spending more than doubling to $632B by 2028
• Global AI spending: Projected $1.5T in 2025, $2T+ by 2026 (WEF)
• GenAI diffusion: 16.3% of world population in H2 2025 (Microsoft)
• Global North advantage: Adoption growing nearly twice as fast as Global South

• WEF Future of Jobs 2025: 170M new jobs created, 92M displaced, net +78M by 2030; 22% of jobs disrupted
• IMF exposure: ~40% of global employment exposed to AI; ~60% in advanced economies
• Goldman Sachs (Aug 2025): AI could displace 6-7% of U.S. workforce in baseline scenario
• Eloundou et al. (Science, 2024): ~80% of U.S. workforce could have ≥10% of tasks affected by LLMs
• ILO: Impact likely "transformational" (job redesign) rather than purely eliminative; female-dominated occupations more exposed in high-income contexts
• NY Fed: Few firms reported AI-induced layoffs; retraining is most common response
• WEF reskilling: 59% of global workforce will require reskilling by 2030

• Data center electricity: ~415-460 TWh in 2024 (~1.5% of global electricity) → 945-1,000+ TWh by 2030 (IEA)
• U.S. data centers: ~4% of total electricity in 2024, expected to double+ by 2030 (Pew)
• AI-specific servers: Growing ~30% annually, nearly half of net increase in data center electricity
• Training GPT-3: 1,287 MWh electricity, ~700,000 liters water, 550+ tons CO₂
• Google water: 6.1 billion gallons in 2023 (+88% since 2019)
• Water stress locations: Over 70% of new U.S. data centers since 2022 built in high water stress areas (Bloomberg/WRI)
• AI carbon footprint: Estimated up to 80 million tonnes CO₂e in 2025
• AI water footprint: Projected 312-765 billion liters in 2025

• Google AI reductions: Claims AI-powered products enabled ~26M metric tons GHG reductions (fuel-efficient routing, Nest, traffic optimization)
• IEA potential: AI could reduce energy-related emissions by ~4-5% by 2035
• Fossil fuel discovery: Shell/SparkCognition used AI to find fossil fuel resources in 9 days vs 9 months
• Oil unlock potential: WoodMackenzie: AI could help unlock additional 1 trillion barrels of oil through 2050
• Infrastructure lock-in: Louisiana's Entergy building 2.3 GW of new natural gas plants (30-year lifetimes) specifically for Meta's data center
• Key insight: "AI for sustainability" must be evaluated on system boundaries — locally efficient products can increase total consumption

• Training: One-time energy cost, builds model parameters/weights. GPT-3 (175B parameters) required ~1,287 MWh. GPT-4 estimated ~50 GWh (40x more).
• Inference: Running queries, dominates lifecycle — 80-90% of all AI computing. Billions of daily queries mean ChatGPT alone may consume 391,000-463,000 MWh annually.
• Scaling laws (Kaplan et al., 2020): Performance improves as power-law of size, data, compute — drove "bigger is better" paradigm
• Parameter scaling: Migrating from 7B to 70B parameters increases per-token energy cost by 100x
• Query efficiency: AI queries use ~5-10x more energy than standard Google searches

• Strubell, Ganesh, McCallum (2019): Training a large Transformer emitted ~284 tons CO₂ — 5x lifetime emissions of average American car. Launched the field of AI energy accounting.
• Patterson et al. (2021, 2022) at Google: Argued efficiency gains could make ML training emissions "plateau, then shrink"
• Luccioni, Strubell, Crawford (ACM FAccT, Jan 2025): Challenged this with Jevons paradox — more efficient AI → cheaper to run → more deployment → greater total consumption
• Evidence: Google delivers 6x more computing power per unit of electricity vs 5 years ago, yet total energy consumption rose 27% in 2024
• IEA projections: Data center CO₂ could peak at ~320M tons by 2030 before declining with renewables. Morgan Stanley projects 600M tons by 2030.
• Rebound effect: Economy-wide models show 30-60% of gross energy savings offset in transport, manufacturing, data processing
• MIT Net Climate Impact Score: Novel framework balancing immediate emissions against future benefits, accounting for time value of carbon

Daron Acemoglu (MIT, 2024 Nobel laureate)

Most cautious. AI yields ≤0.66-0.71% TFP increase over 10 years (GDP +1.1-1.6%). Key concept: "so-so automation" — tech that replaces workers without meaningful productivity gains. Warns AI currently widens capital-labor gap. Notes manipulative AI (targeted advertising) may add 2% GDP with -0.72% welfare impact.

Erik Brynjolfsson (Stanford)

"Productivity J-curve" — GPTs require 2-3 decades of complementary investments before benefits materialize. Landmark 2025 QJE study: AI assistance increased customer support productivity by 15% on average, with less experienced workers improving 30%+ while top performers saw minimal gains. Contradicts skill-biased technical change hypothesis.

David Autor (MIT)

AI could reverse decades of labor market polarization by enabling broader workers to perform expert tasks. Warning about "devaluation of expertise."

• Noy and Zhang (Science, 2023): ChatGPT exposure increased job satisfaction and self-efficacy
• German SOEP longitudinal study (2000-2020): No negative mental health effects from actual AI exposure — anticipated disruption may be more damaging than reality
• Algorithmic management: Cornell/McMaster study found greater automated management intensity → lower well-being, job satisfaction
• Monitoring prevalence: 40-50% of warehouse/telecom/retail/healthcare workers report electronic productivity monitoring
• Worker sentiment (Pew): 52% of U.S. workers worried about future AI; only 6% expect more opportunities
• Manager perspectives: OECD survey of 6,000+ mid-level managers across 6 countries: algorithmic tools increasingly present in managing workers

• Target 8.2 (productivity through innovation): AI projected $2.6-4.4T annual value creation (McKinsey). Indicator 8.2.1 (GDP per employed person) as macro proxy for "AI productivity dividend."
• Target 8.5 (full employment, decent work, equal pay): OECD projects 14% of jobs at high risk. Indicators 8.5.1 (hourly earnings) and 8.5.2 (unemployment) capture distributional impacts.
• Target 8.6 (reduce youth NEET): Relevant if entry-level pathways shrink. 77% of new AI roles require master's degrees (PwC).
• Target 8.8 (labor rights, safe environments): Algorithmic management and surveillance concerns.

• Target 9.4 (upgrade infrastructure for sustainability): Green data centers, efficiency standards. Indicator 9.4.1 (CO₂ per unit value added).
• Target 9.5 (enhance scientific research): AI-driven R&D acceleration. AlphaFold won 2024 Nobel Prize in Chemistry.
• Target 9.b (domestic tech development): Infrastructure divide — U.S. $109.1B vs Africa <1 kWh per capita data center electricity vs U.S. ~540 kWh.
• Target 9.c (universal internet access): AI enablement and inclusion gaps.

• Target 10.1 (income growth for bottom 40%): Acemoglu shows AI threatens to depress labor share of GDP
• Target 10.2 (social/economic inclusion): ML enabling alternative credit scoring for unbanked; AI-powered assistive tech for disabilities
• Target 10.3 (equal opportunity): AI bias in hiring, lending, criminal justice. Trust varies: 72% China vs 32% U.S. (Edelman 2025)
• Target 10.4 (wage/social protection policies): Tax-benefit and labor market institutions that cushion displacement

• Target 12.2 (efficient resource use): Predictive analytics matching supply/demand, reducing overproduction
• Target 12.5 (reduce waste): Greyparrot (111 waste categories), AMP Robotics (300+ sorting systems, >90% recovery)
• Target 12.6 (sustainability reporting): AI ESG platforms; but hardware cycles create e-waste
• Tension: AI recommendation engines and targeted advertising drive overconsumption

• Target 13.1 (resilience): AI-enhanced forecasting (GraphCast), disaster risk management
• Target 13.2 (integrate climate measures): Carbon-aware compute standards
• AI emission reduction potential: 5.4 GtCO₂e annually by 2035 (LSE/Systemiq); 4-5% of energy-related emissions (IEA)
• Energy sourcing challenge: Natural gas currently supplies >40% of data center power globally

• Industrial Revolution: English real wages stagnated for ~40 years despite rising productivity ("Engels' Pause," Robert Allen, 2009)
• 1810s-1830s: Threshing machines → Swing Riots (machine-breaking, social unrest)
• Computerization (1980s-2000s): Eliminated 400K+ typist/bookkeeper positions, 1.8M office clerk jobs, 1.9M production worker positions — but PC/internet ultimately created ~15.8M net new jobs
• Key difference: AI targets non-routine cognitive tasks (writing, analysis, decision-making), not just manual/clerical. Digital = near-zero marginal cost = potentially faster diffusion. Affects many sectors simultaneously.
• Brynjolfsson's J-curve: Productivity gains may require 2-3 decades of complementary investments

• Environmental costs unpriced: No AI-specific carbon pricing anywhere. EU ETS covers electricity generation but not data centers directly.
• Labor displacement externality: Diffused across many workers/communities. Companies capture productivity gains; workers/communities bear retraining and disruption costs
• Information asymmetry: Buyers can't observe lifecycle footprint; <3% of LLMs on Hugging Face report emissions
• Tax code bias (U.S.): IRC section 168(k) allows immediate full expensing of AI servers/hardware; human capital investments face restrictive amortization — state-sponsored bias toward replacing rather than upskilling workers
• Pilot purgatory: 70-95% of enterprise AI pilots fail to scale or show ROI, yet $644B was spent in 2025 (market not self-correcting because externalities not priced)

• Cooling: Traditional air cooling physically failing at AI densities. Liquid cooling capital-intensive to retrofit.
• Unchecked parameter growth: Diminishing returns for intelligence while exponentially increasing energy
• Critical gap: No standard methodology for attributing environmental footprint to specific AI workloads/models/applications — single most critical gap

• Real bottleneck: Organizational redesign — firms adopt AI without reworking processes → limited productivity gains and surveillance harms
• Shadow AI economy: Frustrated employees using unsanctioned AI tools because internal solutions too rigid → security/compliance risks
• Maturity crisis: Only 1% of executives consider their organizations highly mature in AI deployment (McKinsey)
• Scaling struggles: 74% of companies struggle to scale AI value; 70% of required resources are people- and process-related (BCG)

• Governance lag: EU AI Act took 5-6 years from proposal to full application; AI model releases happen in months
• Executive reversals: U.S. executive action can be reversed in hours (Biden→Trump transition)
• Moore's Law for compute: Computing power doubling every ~3.4 months (OpenAI analysis) vs any regulatory timeline
• Jurisdictional fragmentation: EU regulates stringently, U.S. deregulates, China controls → no coherent global standard
• Federalism battles: U.S. states vs federal government preemption

• Opacity: Footprint data often opaque (model-level energy/water, supply chain)
• Accounting methods: Market-based vs location-based Scope 2 accounting yields different results (GHG Protocol)
• SDG data: Fewer than half of 193 countries have comparable SDG data since 2015
• Disclosure standards: No widely enforced standard for model-level disclosure

• Skills translation lag: IBM estimates 3 years to upskill for AI; 40% of global workforce needs reskilling
• Talent shortage: AI job postings outnumber qualified candidates 3.5:1
• Geographic mismatch: SF Bay Area captured $90B in AI funding 2024; data centers cluster in some regions while high-wage AI jobs cluster elsewhere
• Education requirement creep: 77% of new AI roles require master's degrees (PwC); PwC projects 1/3 reduction in entry-level hires by 2028
• Inequality amplifier: Workers most likely to transition successfully already possess education, mobility, and resources — exacerbating inequality

• Schwartz, Dodge, Smith, Etzioni (Communications of the ACM, 2020): Make computational efficiency an evaluation criterion alongside accuracy
• Knowledge distillation: DistilBERT delivers ~60% faster inference with 40% fewer parameters, retaining 97% baseline performance
• Quantization: Up to 50% energy reductions
• UNESCO/UCL (2025): Smaller, task-specific models cut energy by up to 90% without losing performance
• Mixture of Experts: DeepSeek MoE uses 16B total parameters but activates only 2.8B per token
• Federated learning: Reduces data movement, concentrates compute at edge; energy-privacy tradeoffs

• DeepMind cooling: 40% reduction in cooling energy at Google data centers, 15% overall PUE improvement
• UPS ORION: Saves 38M liters of fuel annually, prevents ~100K metric tons CO₂/year
• DeepMind GNoME: 2.2M new crystal structures, 380,000 stable materials (~800 years of human knowledge), 528 potential lithium-ion conductors (25x previous studies)
• DeepMind GraphCast (Science): Outperforms ECMWF forecasts on 90% of evaluated targets
• Precision agriculture: 30-50% water savings, 20-30% yield improvements
• Carbon capture: AI boosting efficiency by 16.7%, reducing energy consumption by 36%+
• Carbon capture cost reduction: McKinsey: AI process optimization can reduce capture costs 15-25%
• Google demand-response: Agreements with utilities for data center load management
• NVIDIA Earth-2 (Jan 2026): Climate simulations 500x faster, 90% compute reduction

Neuromorphic Computing

Brain-inspired, event-driven processing. Human brain operates on ~20 watts. UC Santa Cruz adapted LLM to Intel Loihi 2 at half GPU energy (April 2025). Separately demonstrated billion-parameter LLM on 13 watts (power of a lightbulb) — 50-fold improvement. Potential >90% energy reduction.

Photonic Computing

Light instead of electrical currents, up to 30% lower power

Quantum-Classical Hybrid

Google Willow performs certain computations in <5 minutes that would take classical supercomputers 10 septillion years. Practical advantage for broad AI tasks 5-10+ years away. Quantinuum/NVIDIA hybrid architectures via NVQLink.

• Pew (Aug 2024, 5,410 U.S. adults): 51% more concerned than excited about AI (up from 37% in 2021); only 11% more excited
• Edelman Trust Barometer 2025: 72% trust in China vs 32% in U.S. In U.S., 3x more reject (49%) than embrace (17%); in China 5.5x more embrace than reject
• KPMG (48,000 people, 47 countries): 72% accept AI in daily lives but 54% remain wary/distrustful. Sharp distinction between functional trust (ability) and ethical trust (safety, bias)
• Pew (Sept 2025): 57% rate AI societal risks as high vs 25% who see high benefits
• Consumer trust in AI companies: Only 21% inherently trust AI companies (KPMG)

• Gender gap: Men more likely than women to foresee personal benefit (31% vs 18%, Pew 2024)
• Income effect: Higher-income more likely to increase spending on AI devices (34% vs 24%, Deloitte 2024)
• Education gap: Workers with bachelor's: nearly 2x likely to use AI at work vs less education (28% vs 16%, Pew 2025)
• Age effect: Younger workers more likely to report excitement
• Global divide: Trust higher in emerging economies (57%) than advanced economies (39%) (KPMG)
• Development enthusiasm: Developing markets (Brazil, China, India) substantially more enthusiastic

• Authenticity penalty: Consumers penalize perceived authenticity when content/products revealed as AI-generated
• Algorithm aversion: Consumers devalue AI outputs especially in subjective domains
• Appreciation for search: Algorithm appreciation for search-based products; human guidance preferred for experiential ones

• Design tension: AI recommendation engines designed to maximize engagement/purchases — direct tension with sustainable consumption
• Luccioni et al. (FAccT 2025): "AI-driven targeted advertising can increase superfluous consumer purchases"
• Autonomy risk: Algorithms narrow consumer options, undermine autonomy (Laitinen and Sahlgren, 2021)
• Household impact: Household consumption accounts for 72% of global emissions (Energy Research & Social Science)

• Pilot programs: 40+ pilot programs globally, consistently positive well-being findings, no mass work withdrawal
• Finland (2017-18): 2,000 participants, €560/month. Higher life satisfaction (7.3 vs 6.8/10), lower stress, no negative employment effect
• Stockton SEED (2019-21): 125 participants, $500/month. Full-time employment increased 28% → 40%
• Kenya GiveDirectly: 20,000+ recipients, 12-year study. Reduced food insecurity, improved health
• Limitation: All time-limited, too small for macroeconomic effects. National-scale financing unresolved.
• Financing requirement: Sustainable 11%-of-GDP UBI requires AI productivity 5-6x current levels (arXiv analysis)
• Social warning: Sociologists warn UBI could entrench "symbolic violence" — permanent stratification between AI owners and passive recipients
• Equity path: Raising public revenue share of AI capital to 33% could halve required productivity threshold

• South Korea (Aug 2017): Reduced automation investment tax deductions by up to 2 percentage points — not direct tax. New installations did decrease, causation debated.
• EU rejection: Parliament rejected robot tax proposal (2017)
• MIT economic analysis (2022): Optimal robot tax would be "pretty small"
• Definitional challenge: "What counts as a robot?"

• Singapore SkillsFuture (2015): All citizens 25+ get S$500 training credit; enhanced subsidies up to 90% for mid-career
• France Compte Personnel de Formation: €500/year (€800 disadvantaged), cap €5,000, portable across jobs
• OECD finding: Disproportionately benefit already highly-qualified workers
• Uptake reality: Only 0.1% of France's 1.5M trainees chose green skills courses

• EU Just Transition Funds: €17.5B (2021-2027) for regions most affected by climate transition
• Germany coal regions: €40B through 2038
• AI-specific gap: No equivalent "AI Transition Fund" exists in U.S.
• Portable benefits: Proposals for benefits that travel across employers and gig roles, addressing freelance nature of algorithmic work

• Carbon pricing gap: No AI-specific carbon pricing anywhere
• EU Energy Efficiency Directive (recast 2023): Member states must promote energy-efficient data centers; reporting from May 2024
• EU AI Act Article 40: GPAI providers report energy/resource use (but enforcement specifics undeveloped)
• U.S. legislative effort: AI Environmental Impacts Act of 2024: would have directed EPA study + NIST standards — never enacted
• Integrated strategy potential: Nature Sustainability (2025): coordinated strategies (smart siting + grid decarbonization + efficiency) could reduce AI data center carbon ~73% and water ~86%
• Germany EnEFG: PUE 1.2 by 2026, 10% energy reuse, 100% renewable by 2027

1. AI-Powered Climate Intelligence

Examples: ClimateAI (agriculture/supply chain risk), Tomorrow.io (weather intelligence, proprietary satellites), Jupiter Intelligence (real estate/insurance analytics)

Model: B2B SaaS: assess climate risk exposure, optimize supply chains, underwrite insurance premiums

2. Circular Economy AI

Examples: Greyparrot (waste analytics, 111 categories, WEF Tech Pioneer), AMP Robotics (300+ sorting systems, >90% recovery rates)

Model: Hardware + software: automated waste identification, robotics coordination, closed-loop material recovery

3. Carbon Management Platforms

Examples: Persefoni (8,000+ organizations), Watershed (Airbnb, Shopify clients), Climatiq (API, 70,000+ emission factors, $11.7M Series A July 2025)

Market growth: $13B in 2026 → $68B by 2033 (22% CAGR)

Model: Enterprise GHG accounting, Scope 1-3 reporting, ESG compliance automation

4. Green Cloud / Carbon-Aware Compute

Google: 24/7 carbon-free energy by 2030, contracted 8 GW clean energy in 2024, invested in SMRs (500 MW Kairos Power, 1.8 GW Elementl Power)

Microsoft: Climate Innovation Fund ($1B), Three Mile Island restart, Cloud for Sustainability

AWS: 100% renewable matching 2023 (7 years early), extended server lifetimes 5→6 years

Crusoe Energy: Stranded energy for AI compute

Model: Infrastructure arbitrage: renewable-only data centers, workload shifting, demand-response partnerships

5. AI Footprint Measurement

Foundation: Green Software Foundation SCI and SCI-for-AI spec → standardized measurement/audit services

Emerging market: B2B "AI ESG accounting" — measure, report, reduce energy/water footprint per workload

6. Sustainable Agriculture AI

Applications: Precision irrigation (30-50% water savings), precision fertilizer (20% reduction), OctaPulse (aquaculture, 5-min→30-sec inspections)

7. Environmental Compliance

Example: Rimba — parse regulatory PDFs + SCADA data for industrial compliance

8. Renewable Energy Siting

Example: Paces — geospatial due diligence automation

9. Workforce Reskilling

Example: Coursera/Udemy merging into $2.5B platform (AI/data science focus)

Alternative: Multiverse apprenticeships (87% retention, 50% promoted in 6 months)

10. Healthcare Sustainability and AI

Healthcare represents one of AI's highest-impact application domains — AI diagnostics improving accuracy by 15–30%, treatment optimization reducing hospital readmissions — yet the energy footprint of medical AI systems creates a "central contradiction" requiring explicit lifecycle benefit-cost frameworks.

• Energy intensity of medical AI: A single ChatGPT query consumes roughly 10× the electricity of a standard Google search (IEA/UNEP, 2024). Scaled globally for medical diagnostics, pathology imaging analysis, and personalized treatment planning, medical AI servers could consume 85–134 TWh annually — comparable to current healthcare sector electricity use in developed nations.
• The Lancet critique: The Lancet Digital Health (2024) characterizes the tension between AI's clinical benefits and environmental costs as "a central contradiction demanding explicit lifecycle cost-benefit frameworks" — i.e., clinicians and hospital administrators need tools to weigh improved patient outcomes against carbon footprint and freshwater use of model inference.
• WHO governance guidance: WHO's Global Initiative on AI for Health provides governance standards and implementation guidance specifically tailored to low- and middle-income countries, addressing both quality assurance and equitable access (npj Digital Medicine, 2025). Framework includes environmental impact assessment.
• Ambient AI clinical trials: NEJM AI (2024) proposes a pragmatic-trial framework for monitoring "ambient AI" — AI systems operating continuously in clinical settings — with integrated environmental impact tracking alongside clinical outcomes.
• Market opportunity: Integrating lifecycle assessment into clinical AI purchasing decisions could differentiate vendors and drive efficiency innovation. Early movers in "green medical AI" may capture premium pricing in European and Scandinavian markets.

• Data access/quality: Environmental data fragmented, inconsistent, unaudited
• Compute costs: 55% of AI-for-SDG grants ≤$250K; 60% of all AI investment goes to infrastructure
• Talent scarcity: 74% of companies struggle to scale (BCG) due to people- and process-related resource constraints
• Enterprise pilot purgatory: 70-95% of AI pilots fail to scale; enterprise transitions take 9+ months vs 90 days for mid-market
• Regulatory uncertainty: Jurisdictions moving independently with conflicting standards