Trend analysis: Responsible AI & environmental impact — where the value pools are (and who captures them)

The environmental footprint of artificial intelligence has become impossible to ignore. Global data center electricity consumption reached an estimated 460 TWh in 2025, roughly 1.8% of global electricity demand, with AI workloads accounting for approximately 30% of that figure and growing at 35 to 40% annually. Training a single large language model now consumes as much electricity as 1,200 average EU households use in a year, and inference workloads at scale can dwarf training costs by 10x or more over a model's production lifetime. This tension between AI's potential to accelerate sustainability and its own escalating resource demands has created a distinct economic category: responsible AI and environmental impact management. Understanding where value concentrates within this category, and who is positioned to capture it, is now essential for sustainability professionals navigating the intersection of digital transformation and environmental strategy.

Why It Matters

Three forces are converging to transform responsible AI from a voluntary aspiration into a compliance and commercial imperative in the EU.

First, the EU AI Act, which entered into force in August 2024 with phased implementation through 2027, establishes the world's first comprehensive AI regulatory framework. While the Act's primary focus is on safety and fundamental rights, Article 40 requires providers of high-risk AI systems to report energy consumption and resource utilization metrics. The European Commission is developing delegated acts that will specify detailed environmental reporting requirements for AI systems, expected to take effect in 2027.

Second, the Corporate Sustainability Reporting Directive (CSRD) now requires approximately 50,000 EU companies to disclose Scope 1, 2, and 3 emissions, including those attributable to cloud computing and AI services. For organizations where AI workloads represent a material portion of their computing estate, this creates a direct reporting obligation that demands accurate measurement of AI-related energy consumption and carbon emissions. The European Sustainability Reporting Standards (ESRS) E1 on climate change explicitly includes purchased cloud services within Scope 3 reporting.

Third, the market itself is demanding efficiency. The cost of AI compute has become a primary constraint on scaling. Goldman Sachs estimated that global spending on AI infrastructure will reach USD 200 billion annually by 2027, with electricity costs representing 15 to 25% of total cost of ownership for large-scale AI deployments. Organizations that can deliver equivalent AI capabilities with lower energy consumption hold a direct competitive advantage.

Mapping the Value Pools

AI Energy Measurement and Carbon Accounting

The foundational value pool lies in measurement: enabling organizations to understand and report the environmental impact of their AI workloads. This subsegment addresses the basic question that most enterprises currently cannot answer with precision: how much energy does our AI consume, and what are the associated emissions?

The measurement challenge is technically complex. AI workloads share physical infrastructure with other computing tasks, making attribution difficult. GPU utilization fluctuates dynamically, and the carbon intensity of electricity varies by grid region and time of day. Traditional IT energy monitoring tools lack the granularity to isolate AI-specific consumption.

Companies capturing value in this space include CodeCarbon, an open-source project backed by Mila and BCG Gamma that tracks emissions from machine learning code execution, and Climatiq, a Berlin-based company providing emission factor APIs that enable programmatic carbon accounting for cloud and AI workloads. Watershed and Persefoni have added AI-specific modules to their enterprise carbon accounting platforms, recognizing that technology companies and AI-intensive enterprises need granular visibility into compute-related emissions.

The addressable market for AI-specific environmental measurement tools is estimated at EUR 800 million to 1.2 billion by 2028, driven primarily by CSRD compliance obligations and the growing number of enterprises running substantial AI workloads. Margins in this subsegment are moderate (40 to 55% gross margins for SaaS platforms) but retention rates are high because switching measurement platforms disrupts historical baselines.

Energy-Efficient AI Hardware and Architecture

The highest absolute value pool resides in hardware and architecture innovations that reduce the energy required per unit of AI computation. This is where the largest capital investments are flowing and where the most significant competitive advantages are being built.

NVIDIA dominates GPU-based AI compute, and its energy efficiency trajectory defines the market. The H100 GPU delivers approximately 3x the energy efficiency (measured in tokens per watt) of the prior-generation A100, while the B200 Blackwell architecture promises another 4x improvement for inference workloads. However, NVIDIA's pricing power means that much of the efficiency gain accrues to NVIDIA itself rather than to AI operators, an important dynamic for sustainability professionals to understand when forecasting organizational energy costs.

Custom silicon represents a growing competitive alternative. Google's TPU v5p delivers inference efficiency comparable to NVIDIA GPUs at lower cost for workloads optimized for Google's ecosystem. Intel's Gaudi 3 accelerator targets price-performance rather than absolute performance, appealing to cost-sensitive enterprise deployments. Graphcore and Cerebras pursue fundamentally different architectures (intelligence processing units and wafer-scale engines, respectively) that promise step-change efficiency improvements for specific workload types.

Beyond chip-level efficiency, algorithmic and architectural innovations are reducing the computational requirements of AI itself. Model distillation, quantization, and pruning techniques can reduce inference costs by 50 to 90% with minimal accuracy loss. Hugging Face has made efficient model deployment accessible through its Optimum library and model hub, while OctoAI and Together AI offer inference platforms that automatically select the most efficient model architecture for each query.

This value pool is enormous, with the global AI chip market projected to exceed USD 150 billion by 2028. However, value capture is highly concentrated among a small number of hardware manufacturers and hyperscale cloud providers. Sustainability professionals should focus less on influencing the hardware market and more on optimizing their organizations' use of available efficient options.

Renewable Energy Procurement for AI Infrastructure

The third major value pool connects AI infrastructure directly to clean energy supply. Hyperscale data center operators and enterprise AI users are among the largest corporate purchasers of renewable energy globally, and their procurement strategies are becoming more sophisticated.

Microsoft committed to matching 100% of its electricity consumption with zero-carbon energy purchases by 2025, and has signed over 13 GW of renewable energy agreements since 2020. Google goes further, targeting 24/7 carbon-free energy matching at every data center by 2030, an approach that requires location-specific procurement rather than portfolio-level balancing. Amazon Web Services is the world's largest corporate buyer of renewable energy, with 500+ renewable energy projects globally.

For sustainability professionals outside the hyperscale sector, the key dynamic is that cloud providers' clean energy investments directly affect the carbon intensity of AI workloads run on their platforms. Choosing between AWS, Google Cloud, and Azure for AI workloads involves not just price and performance but also the carbon intensity of each provider's regional infrastructure. Google Cloud's Carbon Footprint dashboard and Microsoft's Emissions Impact Dashboard make these comparisons increasingly transparent.

An emerging subsegment within this value pool is temporal and spatial energy matching, where AI workloads are scheduled to run during periods of high renewable generation or in regions with cleaner grids. Electricity Maps provides real-time carbon intensity data for 200+ grid zones globally, enabling carbon-aware workload scheduling. Google's internal carbon-intelligent computing platform reportedly shifts 25 to 30% of non-time-sensitive workloads to periods with cleaner energy, reducing associated emissions without affecting service quality.

AI for Environmental Optimization

The fourth value pool is AI applied to reduce emissions across the broader economy, creating a positive feedback loop where AI's environmental costs are offset by its environmental benefits. This category spans building energy optimization, grid management, supply chain emissions reduction, and industrial process efficiency.

Siemens Building X platform uses AI to optimize energy consumption across 500,000+ connected buildings. Google DeepMind demonstrated 40% cooling energy reductions in data centers through reinforcement learning. Uptake applies AI to industrial asset optimization, reducing energy waste in manufacturing and heavy industry.

The strategic insight for sustainability professionals is that AI's net environmental impact depends heavily on deployment choices. An AI system consuming 100 MWh annually but enabling 1,000 MWh of energy savings elsewhere delivers a 10:1 return on environmental investment. The organizations best positioned to capture value in this space are those that can credibly measure and communicate this net impact, a capability that connects back to the measurement value pool.

Responsible AI Environmental Impact KPIs: Benchmark Ranges

Metric	Below Average	Average	Above Average	Top Quartile
AI Compute Carbon Intensity (gCO2e/kWh compute)	>400	200-400	100-200	<100
Model Efficiency (tokens per watt)	<500	500-2,000	2,000-5,000	>5,000
Clean Energy Match (% of AI workload)	<50%	50-75%	75-95%	>95%
Inference Cost Reduction (vs. baseline)	<20%	20-40%	40-60%	>60%
AI Environmental ROI (savings/footprint)	<2:1	2:1-5:1	5:1-10:1	>10:1
CSRD AI Disclosure Readiness	Ad hoc	Partial	Systematic	Auditable

Who Captures the Value

Value capture in responsible AI and environmental impact follows a clear hierarchy. Hardware manufacturers and hyperscale cloud providers capture the largest share because they control the physical infrastructure layer. NVIDIA alone generated USD 60 billion in data center revenue in fiscal year 2025, with margins exceeding 75%. Cloud providers capture the next layer through managed AI services that bundle compute, energy, and optimization.

Software and measurement companies capture a smaller but faster-growing share. The market for AI governance and environmental measurement tools is growing at 50 to 60% annually from a smaller base, with the best-positioned companies building data moats through integration with enterprise systems and historical baselines.

Consulting and advisory firms capture value through implementation services, particularly for CSRD compliance and AI governance frameworks. The Big Four accounting firms have all launched dedicated AI sustainability practices, recognizing that their audit relationships provide a natural channel to advisory revenue.

Organizations that deploy AI efficiently, rather than building AI tools, capture value indirectly through lower operating costs and stronger sustainability positioning. For most sustainability professionals, this is the relevant value pool: optimizing their organization's AI deployment to minimize environmental impact while maximizing business value.

Action Checklist

• Inventory all AI workloads across cloud and on-premise infrastructure, mapping compute consumption and associated emissions
• Evaluate cloud providers' regional carbon intensity and factor emissions data into procurement decisions for AI infrastructure
• Implement AI-specific carbon accounting using tools like CodeCarbon, Climatiq, or cloud provider dashboards
• Assess CSRD reporting obligations for AI and cloud computing under ESRS E1 Scope 3 requirements
• Explore model optimization techniques (distillation, quantization, pruning) to reduce inference energy by 50% or more without sacrificing performance
• Establish an internal AI environmental impact review process for new model deployments and training runs
• Evaluate carbon-aware workload scheduling to shift non-time-sensitive AI tasks to periods of cleaner grid electricity
• Calculate net environmental ROI for AI deployments by comparing operational footprint against documented sustainability benefits

FAQ

Q: How should sustainability professionals prioritize responsible AI initiatives given limited budgets? A: Start with measurement. You cannot manage what you cannot measure. Implementing AI-specific carbon accounting typically costs EUR 20,000 to 80,000 for enterprise deployments and provides the baseline for all subsequent optimization. Next, focus on inference optimization, which often delivers 50 to 70% cost reductions alongside equivalent emissions reductions. Carbon-aware scheduling is a third priority that adds incremental benefit at relatively low implementation cost.

Q: How material is AI's environmental impact relative to other corporate emissions sources? A: For most enterprises, AI currently represents 1 to 5% of total Scope 2 and 3 emissions. However, growth rates of 35 to 40% annually mean that AI could represent 10 to 15% of emissions within three to four years if left unmanaged. For technology companies, the figures are significantly higher, with some reporting AI compute as 15 to 30% of total emissions already.

Q: What does CSRD require specifically for AI environmental reporting? A: CSRD does not single out AI explicitly, but ESRS E1 requires disclosure of energy consumption and GHG emissions across all material sources, including purchased cloud computing services. Organizations must report Scope 3 Category 1 (purchased goods and services) emissions, which encompasses cloud AI services. Companies should work with their auditors to establish materiality thresholds and reporting methodologies for AI workloads specifically.

Q: How reliable are cloud providers' carbon reporting tools? A: Cloud provider dashboards have improved significantly but remain imperfect. Google Cloud's Carbon Footprint tool and Microsoft's Emissions Impact Dashboard provide regional and temporal granularity that is generally sufficient for CSRD reporting. However, they rely on provider-disclosed energy data and emission factors that cannot be independently verified at the facility level. Organizations should use these tools as a starting point while developing capability to cross-reference with third-party grid emission data from sources like Electricity Maps or the European Environment Agency.

Sources

• International Energy Agency. (2025). Data Centres and Data Transmission Networks: Electricity Consumption Tracking Report. Paris: IEA.
• Goldman Sachs Research. (2025). AI Infrastructure: The USD 200 Billion Question. New York: Goldman Sachs.
• European Commission. (2024). EU AI Act: Regulation (EU) 2024/1689 on Artificial Intelligence. Brussels: Official Journal of the European Union.
• Google. (2025). Environmental Report 2025: Progress Toward 24/7 Carbon-Free Energy. Mountain View, CA: Google LLC.
• Electricity Maps. (2025). Real-Time Carbon Intensity Data: Methodology and Coverage Report. Copenhagen: Electricity Maps.
• European Financial Reporting Advisory Group. (2023). ESRS E1: Climate Change Reporting Standard. Brussels: EFRAG.
• NVIDIA Corporation. (2025). Fiscal Year 2025 Annual Report: Data Center Segment. Santa Clara, CA: NVIDIA.