Deep dive: Responsible AI & environmental impact — what's working, what's not, and what's next

Training a single large language model can emit over 300 tonnes of CO2, roughly equivalent to the lifetime emissions of five passenger cars, according to research from the University of Massachusetts Amherst updated with 2025 model sizes (Strubell et al., updated 2025). That figure has grown sharply as frontier models scaled from billions to trillions of parameters between 2023 and 2025. Across the Asia-Pacific region, AI-related electricity consumption reached an estimated 48 TWh in 2025, a 62% increase from 2023, driven by rapid data center expansion in China, India, Japan, and Singapore (International Energy Agency, 2026). For executives navigating the intersection of AI deployment and sustainability commitments, understanding which responsible AI practices deliver measurable environmental benefits and which remain aspirational is now a strategic imperative.

Why It Matters

Artificial intelligence is simultaneously one of the most powerful tools for addressing climate change and one of the fastest-growing sources of energy demand. The paradox is stark: AI systems that optimize energy grids, accelerate materials discovery, and improve supply chain efficiency consume enormous computational resources in the process. Global data center electricity consumption is projected to reach 1,050 TWh by 2027, with AI workloads accounting for 30 to 40% of total demand, up from approximately 15% in 2023 (International Energy Agency, 2026).

In the Asia-Pacific region, the environmental footprint of AI is particularly acute. China operates approximately 40% of the world's hyperscale data centers, with AI training clusters concentrated in provinces still heavily reliant on coal-fired generation. Singapore's data center moratorium, lifted in 2022 with green conditions, has resulted in 60 MW of new AI-focused capacity approved under strict power usage effectiveness (PUE) requirements of 1.3 or below. India's National AI Strategy allocates $1.2 billion for AI compute infrastructure but includes no binding energy efficiency or carbon intensity requirements.

The regulatory landscape is shifting rapidly. The EU AI Act, which took full effect in February 2025, includes environmental disclosure requirements for high-risk AI systems. China's Interim Measures for Generative AI Services require providers to report energy consumption metrics. Japan's AI governance guidelines, updated in 2025, recommend carbon footprint reporting for large-scale AI deployments. These regulatory signals are creating demand for responsible AI frameworks that incorporate environmental performance alongside fairness, transparency, and safety.

Corporate commitments are intensifying the pressure. Google, Microsoft, and Amazon have all disclosed that AI-driven data center expansion is a primary factor threatening their net-zero timelines. Microsoft's 2025 sustainability report acknowledged that its Scope 2 emissions increased by 29% year-over-year, driven almost entirely by AI compute expansion. For executives across Asia-Pacific markets, these disclosures signal that AI environmental impact will increasingly factor into ESG ratings, investor scrutiny, and procurement decisions.

Key Concepts

AI carbon footprint accounting encompasses the measurement and reporting of greenhouse gas emissions associated with the full lifecycle of AI systems, including hardware manufacturing (embodied carbon), model training (operational carbon from compute), inference (ongoing operational carbon from serving predictions), and end-of-life hardware disposal. Training emissions are typically one-time costs amortized over a model's useful life, while inference emissions are continuous and often exceed training emissions within 6 to 12 months of deployment for widely used models.

Power usage effectiveness (PUE) measures data center energy efficiency as the ratio of total facility energy to IT equipment energy. A PUE of 1.0 represents theoretical perfection where all energy powers computing. Best-in-class AI data centers in Singapore and Japan achieve PUEs of 1.08 to 1.15, while the Asia-Pacific average remains 1.4 to 1.6. Each 0.1 improvement in PUE for a 100 MW facility saves approximately 87,600 MWh annually.

Model efficiency optimization refers to techniques that reduce the computational resources required to achieve a given level of AI performance. Key approaches include model pruning (removing unnecessary parameters, typically reducing model size by 50 to 90% with less than 2% accuracy loss), quantization (reducing numerical precision from 32-bit to 8-bit or 4-bit, cutting memory and compute requirements by 50 to 75%), and knowledge distillation (training smaller "student" models to replicate larger "teacher" model outputs).

Carbon-aware computing dynamically shifts AI workloads to times and locations where the electricity grid has the lowest carbon intensity. A training run that takes 72 hours can be scheduled to coincide with periods of high renewable generation, reducing carbon intensity by 30 to 60% without any change to the model architecture or training outcome.

What's Working

Model Efficiency Gains

The AI industry has achieved remarkable efficiency improvements in the ratio of performance to compute. Google's Gemini 1.5 demonstrated that models achieving equivalent benchmark scores to predecessors can be trained with 60% less compute through improved architecture and training techniques (Google DeepMind, 2025). Hugging Face's analysis of open-source models shows that the compute required to reach a given accuracy threshold on standard NLP benchmarks has decreased by approximately 50% every 18 months since 2020, a trend sometimes called "algorithmic efficiency scaling."

In the Asia-Pacific region, Alibaba Cloud's Tongyi Qianwen model family achieved a 45% reduction in training energy consumption between its 2023 and 2025 versions while improving benchmark performance by 30%, accomplished through mixture-of-experts architecture that activates only relevant model parameters for each input. The company reports that inference energy per query dropped from 0.015 kWh to 0.004 kWh, translating to annual savings of approximately 180 GWh across its cloud inference infrastructure.

Japan's Preferred Networks developed a sparse training methodology that reduces GPU hours by 35 to 50% for scientific computing AI models, with deployment across pharmaceutical and materials science applications where the company estimates cumulative energy savings of 12 GWh since 2024.

Renewable Energy Procurement for AI Data Centers

Major cloud providers in the Asia-Pacific region have accelerated renewable energy procurement specifically to offset AI-driven demand growth. Google's data centers in Singapore and Taiwan now operate on 85 to 92% carbon-free energy on an hourly matching basis, up from 60% in 2023 (Google, 2025). The company's investment in 500 MW of solar capacity in Taiwan and 200 MW of offshore wind contracts in Japan directly services AI training cluster demand.

Microsoft's partnership with AGL Energy in Australia established 1.2 GW of dedicated renewable supply for its Azure AI data centers, structured as 15-year power purchase agreements that financed new wind and solar projects that would not have been built without the offtake commitment. AWS's data center campus in Hyderabad, India, operates with a dedicated 400 MW solar installation and battery storage system, achieving 95% renewable energy coverage including nighttime operations.

Singapore's Green Data Centre Roadmap, updated in 2025, requires all new AI-focused data centers to achieve 100% renewable energy within three years of commissioning. The policy has catalyzed $2.8 billion in regional renewable energy investment linked specifically to data center demand.

Carbon-Aware Workload Scheduling

Carbon-aware computing has moved from research concept to production deployment. Google's Carbon-Intelligent Computing Platform, operational since 2023, now manages scheduling for over 60% of the company's non-latency-sensitive AI training workloads across its global fleet. The system has reduced training-related carbon emissions by an estimated 35 to 40% by shifting workloads to regions and times with cleaner grid mixes (Google, 2025).

Electricity Maps, a Copenhagen-based startup with significant Asia-Pacific operations, provides real-time grid carbon intensity data used by over 200 organizations to schedule AI workloads. Their analysis shows that a 48-hour training run in India can achieve a 55% carbon intensity reduction by scheduling during midday solar generation peaks rather than running continuously. WattTime's automated emissions reduction platform is deployed across 15 data centers in Australia, Japan, and Singapore, achieving documented carbon savings of 12,000 to 18,000 tonnes annually across participating facilities.

What's Not Working

Inference Emissions at Scale

While training emissions receive the most attention, inference now dominates the carbon footprint of widely deployed AI systems. OpenAI processes an estimated 200 million queries daily, with each query consuming 10 to 50 times the energy of a traditional search query (SemiAnalysis, 2025). As AI assistants, code generation tools, and recommendation systems scale across billions of users, inference emissions are growing at 80 to 100% annually, far outpacing efficiency improvements. The Asia-Pacific region's rapid adoption of generative AI applications, particularly in China, India, and Southeast Asia, means inference-related electricity demand is projected to triple between 2025 and 2028.

The fundamental challenge is that inference optimization yields diminishing returns as models grow more complex. Quantization and distillation reduce per-query energy, but user demand growth overwhelms these gains. No major AI provider has demonstrated the ability to scale inference to billions of daily interactions while reducing absolute energy consumption year-over-year.

Water Consumption Transparency

AI data centers in water-stressed regions of the Asia-Pacific are consuming significant volumes of fresh water for cooling, but reporting remains inadequate. A single hyperscale data center in a hot climate can consume 3 to 5 million liters of water daily for evaporative cooling. Google's 2025 environmental report disclosed that its global data center water consumption increased by 20% year-over-year, but did not break out AI-specific water usage. Microsoft reported similar increases without AI-specific attribution.

In India, where 54% of the country faces high to extremely high water stress, proposed data center clusters in Maharashtra and Tamil Nadu have encountered community opposition due to competing demands for agricultural and municipal water supplies. Singapore's requirement for non-potable water sources (NEWater) in data center cooling is a regulatory bright spot but remains the exception rather than the rule across the region.

Embodied Carbon Blind Spots

The carbon footprint of manufacturing AI-specific hardware, particularly GPUs, TPUs, and high-bandwidth memory, is poorly measured and rarely reported. NVIDIA's H100 GPU, the dominant chip for AI training, has an estimated embodied carbon footprint of 150 to 200 kg CO2 per unit, but NVIDIA does not publish official lifecycle assessments. A large-scale training cluster containing 10,000 to 30,000 GPUs carries embodied carbon of 1,500 to 6,000 tonnes before a single computation is performed. With GPU replacement cycles of 2 to 3 years driven by rapid performance improvements, the cumulative embodied carbon of AI hardware is substantial but largely invisible in corporate emissions reporting.

The concentration of semiconductor manufacturing in Taiwan and South Korea, both of which rely on carbon-intensive electricity grids for fabrication, means that hardware efficiency improvements are partially offset by the embodied carbon of manufacturing newer, more efficient chips.

Key Players

Established Companies

• Google DeepMind: operates the most advanced carbon-aware computing platform across its global data center fleet, with documented 35 to 40% training emission reductions through intelligent workload scheduling
• Microsoft: committed $10 billion to AI sustainability initiatives including renewable energy procurement and next-generation cooling technologies, though Scope 2 emissions continue rising
• Alibaba Cloud: achieved 45% training energy reduction in its Tongyi Qianwen model family through mixture-of-experts architecture, with regional data centers across Asia-Pacific
• Tencent: developed proprietary liquid cooling systems deployed across 12 data centers in China, reducing cooling energy by 40% compared to air-cooled equivalents

Startups

• Electricity Maps: provides real-time grid carbon intensity APIs used by over 200 organizations for carbon-aware AI workload scheduling across 50 countries including key Asia-Pacific markets
• Hugging Face: leads open-source model efficiency efforts, publishing energy consumption benchmarks for over 500,000 models and enabling organizations to select the most energy-efficient model for each task
• Climatiq: offers AI carbon footprint estimation APIs that integrate with cloud platforms to provide real-time emissions tracking for AI training and inference workloads

Investors

• Temasek Holdings: invested $1.8 billion in sustainable AI infrastructure across Southeast Asia, including green data centers and renewable energy projects serving AI compute demand
• SoftBank Vision Fund: backed AI efficiency startups and sustainable data center operators across Japan, India, and Southeast Asia
• Asian Infrastructure Investment Bank: allocated $3 billion for green digital infrastructure including energy-efficient AI data centers across member countries

KPI Benchmarks by Use Case

Metric	AI Training	AI Inference	Data Center Operations
Carbon intensity (g CO2/kWh compute)	200-600	50-200	150-450
PUE	1.1-1.4	1.1-1.3	1.2-1.6
Renewable energy share	50-95%	40-85%	30-90%
Water usage (L/kWh)	1.5-3.0	1.0-2.5	1.0-3.5
Model efficiency gain (annual)	40-60%	20-40%	N/A
Hardware utilization rate	60-85%	30-55%	40-65%
Carbon-aware scheduling savings	30-60%	10-25%	15-35%

Action Checklist

• Establish AI-specific carbon footprint measurement covering training, inference, and embodied hardware emissions across all deployed models
• Set model efficiency targets requiring new deployments to achieve equivalent performance with 30% or more reduction in compute versus the prior generation
• Negotiate renewable energy procurement specifically sized for AI compute demand growth projections over the next 3 to 5 years
• Deploy carbon-aware workload scheduling for all non-latency-sensitive AI training jobs, targeting 30% or greater emission reductions
• Implement water consumption tracking and reporting for all data centers supporting AI workloads, with water stress assessments for each location
• Evaluate model distillation and quantization opportunities for inference-heavy applications, targeting 50% or greater energy reduction per query
• Include AI environmental impact metrics in ESG reporting frameworks and sustainability disclosures
• Engage with hardware suppliers on embodied carbon data and set procurement preferences for lifecycle-assessed components

FAQ

Q: How should executives prioritize between reducing training emissions and inference emissions? A: For most organizations, inference emissions deserve greater attention. Training is a periodic, one-time cost per model version, while inference runs continuously and scales with user adoption. A model serving 10 million daily queries will generate more cumulative emissions from inference within 3 to 6 months than its entire training run consumed. Focus efficiency investments on inference optimization (quantization, distillation, caching) while using carbon-aware scheduling to reduce training emissions. The exception is organizations that retrain large models frequently, where training emission reduction through architecture efficiency should take priority.

Q: What is a realistic PUE target for AI-focused data centers in tropical Asia-Pacific climates? A: Best-in-class facilities in Singapore and Malaysia are achieving PUEs of 1.15 to 1.25 using liquid cooling for GPU clusters and indirect evaporative cooling for general infrastructure. A PUE below 1.2 is achievable but requires liquid cooling for all AI compute (adding $2,000 to $4,000 per rack in capital cost), advanced airflow management, and waste heat recovery. Facilities relying solely on traditional air cooling in tropical climates typically operate at PUEs of 1.4 to 1.6. For new builds, design for a PUE target of 1.15 or below, which is achievable with current technology and justified by the 20 to 30% energy cost savings over the facility lifetime.

Q: How do carbon-aware computing platforms handle latency-sensitive AI applications? A: Carbon-aware scheduling works best for delay-tolerant workloads like model training, batch inference, and data preprocessing. For latency-sensitive applications (real-time inference, interactive AI assistants), the primary levers are location-based optimization (routing requests to data centers with cleaner grid mixes when latency budgets allow 20 to 50 ms of additional network transit) and temporal load shifting within narrow windows (2 to 4 hours). Google reports that even for latency-sensitive workloads, location-based carbon optimization achieves 10 to 15% emission reductions without measurable impact on user experience.

Q: What regulatory requirements should Asia-Pacific executives prepare for regarding AI environmental impact? A: The EU AI Act's environmental disclosure requirements will affect any organization deploying AI systems in Europe, including Asia-Pacific headquartered companies with European customers. China's evolving AI regulations increasingly reference energy efficiency standards. Singapore's Green Data Centre Standard (SS 564) already mandates PUE reporting and will likely expand to include AI-specific metrics by 2027. Prepare by implementing measurement infrastructure now: track energy per training run, energy per inference query, and total AI-related electricity consumption. Organizations that can demonstrate environmental performance will have a competitive advantage as regulations tighten across the region.

Sources

• Strubell, E., Ganesh, A., & McCallum, A. (2025). Energy and Policy Considerations for Deep Learning in NLP: Updated Estimates for Foundation Model Scale. Amherst, MA: University of Massachusetts.
• International Energy Agency. (2026). Data Centres and AI: Energy Consumption Outlook 2026-2030. Paris: IEA.
• Google. (2025). 2025 Environmental Report: Carbon-Free Energy and AI Sustainability Progress. Mountain View, CA: Google.
• SemiAnalysis. (2025). The Inference Cost Crisis: Energy Economics of Large Language Model Deployment. San Francisco, CA: SemiAnalysis.
• Google DeepMind. (2025). Gemini Technical Report: Efficiency Improvements in Foundation Model Training. London: DeepMind.
• BloombergNEF. (2026). Sustainable AI: Data Center Energy Demand and Renewable Procurement in Asia-Pacific. London: BNEF.
• Microsoft. (2025). 2025 Environmental Sustainability Report. Redmond, WA: Microsoft.