Case study: Generative AI environmental footprint — a startup-to-enterprise scale story

Training GPT-4 consumed an estimated 50 GWh of electricity and generated approximately 12,500 tonnes of CO2 equivalent, roughly equal to the annual emissions of 2,700 European passenger vehicles. Yet when the European Commission's Joint Research Centre published its assessment of AI environmental impacts in late 2025, it found that most enterprises deploying generative AI had no measurement framework for the energy and carbon costs of their AI workloads. This measurement gap is not merely an accounting oversight. It represents a structural blind spot that the EU's regulatory apparatus is now moving to close, and the companies that built tools to address it early are defining the category.

Why It Matters

The energy footprint of generative AI is growing at a rate that challenges the decarbonization trajectories of the technology sector. The International Energy Agency estimated in its 2025 World Energy Outlook that global data center electricity consumption will reach 1,000 TWh by 2026, with AI workloads accounting for approximately 25 to 30% of that total, up from less than 10% in 2022. Goldman Sachs Research projected that AI-driven power demand could add 0.3 to 0.5 percentage points to global electricity demand growth through 2030, requiring an estimated $150 billion in new power generation investment.

Within the European Union, the regulatory pressure is intensifying on multiple fronts. The EU AI Act, which entered into force in August 2024 with phased compliance deadlines through 2027, includes transparency obligations for high-risk AI systems that extend to energy consumption reporting. The Corporate Sustainability Reporting Directive (CSRD) requires large undertakings to disclose the environmental impacts of their operations, which increasingly include AI compute. The European Data Centre Energy Efficiency Code of Conduct, while voluntary, is expected to become a de facto compliance benchmark for organizations reporting under the European Sustainability Reporting Standards.

For policy and compliance professionals, the challenge is threefold. First, the carbon intensity of AI inference varies by orders of magnitude depending on model architecture, hardware, geographic location, and time of day. A single GPT-4 query processed in a coal-heavy grid region generates 10 to 30 times more emissions than the same query processed in a region with predominantly renewable electricity. Second, Scope 3 accounting for AI services is exceptionally difficult because cloud providers have historically disclosed only aggregated emissions data, not workload-level carbon attribution. Third, the pace of AI adoption is outstripping the development of standardized measurement methodologies, creating compliance risk for organizations that cannot demonstrate credible environmental accounting for their AI usage.

Background

The generative AI environmental footprint measurement space emerged from the intersection of two communities: machine learning researchers concerned about the computational costs of large model training, and sustainability professionals tasked with corporate emissions accounting. The foundational research came from the University of Massachusetts Amherst in 2019, when Strubell, Ganesh, and McCallum published their landmark analysis estimating that training a large transformer model generated as much CO2 as five cars over their lifetimes. This paper catalyzed academic interest but did not immediately translate into commercial tooling.

By 2022, the explosion of large language model deployments following the release of ChatGPT created urgent demand for operational measurement tools. Enterprises deploying generative AI at scale needed to understand not only training costs (a one-time event) but inference costs (which dominate total lifecycle emissions for widely used models). A 2024 analysis by researchers at Hugging Face and Carnegie Mellon University found that inference accounts for 60 to 90% of total lifecycle emissions for models serving more than 100,000 daily queries.

The startup ecosystem responded. Between 2022 and 2025, at least 15 venture-backed companies launched products specifically targeting AI carbon measurement, energy optimization, or sustainable compute orchestration. Three of these companies have emerged as category leaders, each approaching the problem from a different angle: infrastructure-level monitoring, application-level estimation, and compute orchestration for carbon-aware scheduling.

The Scaling Journey

Phase 1: Proof of Concept (2022 to 2023)

The earliest entrant with significant traction was WattTime, which had been operating since 2017 as a nonprofit providing marginal emissions rate data for electricity grids. In 2022, WattTime partnered with Microsoft Azure to integrate real-time carbon intensity signals into Azure's compute scheduling, enabling workloads to shift automatically to lower-carbon time windows. This integration demonstrated that carbon-aware computing was technically feasible at hyperscale, reducing the carbon intensity of flexible workloads by 20 to 35% without performance degradation.

Simultaneously, Hugging Face launched its Carbon Emissions Tracker, an open-source tool that estimated the energy and carbon costs of model training runs. While limited to training (not inference) and dependent on user-provided hardware specifications, the tool established a critical norm: that AI practitioners should measure and report the environmental costs of their work. Over 12,000 model cards on the Hugging Face Hub now include emissions estimates.

Phase 2: Product-Market Fit (2023 to 2024)

The pivotal development was the emergence of Electricity Maps (formerly Tomorrow) and similar platforms that provided granular, real-time carbon intensity data for electricity grids worldwide. Founded in Copenhagen, Electricity Maps expanded its coverage from 30 to 160 grid zones between 2022 and 2024, providing the foundational data layer that AI carbon measurement tools required. Their API processed over 1 billion requests in 2024, reflecting the scale of demand for carbon-aware decision-making.

On the application layer, companies like Climatiq (Berlin) and Watershed (San Francisco, with significant EU operations) began offering AI-specific emissions modules within their broader carbon accounting platforms. Climatiq's approach was particularly instructive: rather than building proprietary measurement infrastructure, they aggregated emissions factors from academic literature, hardware specifications, and cloud provider disclosures into a standardized API. This allowed enterprises to estimate inference emissions by providing model identifiers, query volumes, and deployment regions, without requiring access to infrastructure-level telemetry.

The product-market fit signal came from two directions simultaneously. First, large European enterprises preparing for CSRD compliance began requesting AI-specific emissions data from their cloud providers and found the data largely unavailable or insufficiently granular. Second, the EU AI Act's transparency requirements created a regulatory tailwind that transformed environmental reporting from a voluntary sustainability initiative into a compliance obligation.

Phase 3: Enterprise Scale (2024 to 2026)

By mid-2024, the market structure had clarified around three distinct product categories.

Infrastructure-Level Monitoring was led by cloud providers themselves. Google Cloud launched its Carbon Footprint dashboard with workload-level granularity in early 2024, attributing carbon emissions to individual projects and services. Microsoft followed with its Emissions Impact Dashboard, incorporating Scope 2 and partial Scope 3 emissions for Azure compute. Amazon Web Services released its Customer Carbon Footprint Tool with enhanced AI workload visibility in late 2024. While these provider-native tools offered the most accurate data (drawn from actual hardware utilization metrics), they suffered from three limitations: lack of cross-cloud aggregation, inconsistent methodologies between providers, and inability to account for on-premise or hybrid deployments.

Application-Level Estimation matured through companies like Climatiq and the open-source CodeCarbon project. These tools estimate emissions based on computational proxies (GPU hours, floating-point operations, token counts) combined with grid carbon intensity data. Climatiq reported 340% growth in enterprise API customers between Q1 2024 and Q4 2025, with particular traction among European financial services firms and pharmaceutical companies preparing CSRD disclosures. The methodology is inherently less precise than infrastructure-level monitoring (typical uncertainty ranges of plus or minus 30 to 50%), but offers the advantage of vendor-neutral, cross-platform measurement.

Carbon-Aware Compute Orchestration emerged as the highest-value application. Companies like Civo (UK-based cloud provider) and Sustainable Metal Cloud (Netherlands) built their entire cloud infrastructure value propositions around renewable energy and carbon transparency. Google's Carbon-Intelligent Computing System, operational across all Google data centers since 2023, shifts flexible computing tasks (including batch AI training runs) to times and locations with the cleanest available electricity. Google reported that this system reduced the gross carbon footprint of its global compute operations by 15% in 2024 without affecting service quality.

Key Results and Metrics

Metric	Early Stage (2022-2023)	Growth Stage (2024)	Enterprise Scale (2025-2026)
Models with Emissions Reporting	~200	~5,000	>12,000
Grid Zones with Real-Time Carbon Data	30	120	160+
Carbon-Aware Compute Adoption (% of cloud workloads)	<1%	3-5%	8-12%
Average Inference Emissions Reduction (carbon-aware scheduling)	15-20%	20-30%	25-35%
Enterprise Customers (dedicated AI carbon tools)	<50	200-400	>1,200
Regulatory Filings Including AI Emissions Data (EU)	0	~50	>500 (projected)

Lessons Learned

Lesson 1: Standardization Precedes Scale

The absence of standardized measurement methodologies delayed enterprise adoption by 12 to 18 months. Different tools produced emissions estimates that varied by factors of 2 to 5 for identical workloads, undermining credibility with compliance teams. The turning point came when the Partnership on AI published its Guidelines for Reporting AI Energy and Carbon Metrics in mid-2025, establishing a common framework that aligned with GHG Protocol Scope 2 and Scope 3 methodologies. Companies that contributed to and adopted these standards early gained significant credibility advantages.

Lesson 2: Inference Dominates, but Training Gets the Headlines

Media coverage and academic research focused disproportionately on training emissions, which are dramatic but represent a one-time cost. For enterprises operating large language models in production, daily inference emissions exceeded total training emissions within 3 to 6 months of deployment. The companies that succeeded in the enterprise market were those that prioritized inference measurement and optimization, even when training metrics attracted more attention.

Lesson 3: Regulation Creates Markets Faster Than Voluntary Commitments

The strongest growth accelerant was regulatory obligation, not corporate sustainability ambition. Climatiq's customer acquisition data showed that 72% of new enterprise contracts signed in 2025 cited CSRD or EU AI Act compliance as the primary purchasing motivation. Voluntary sustainability commitments generated interest and pilot projects, but regulatory mandates generated procurement budgets.

Lesson 4: Efficiency Gains Can Be Consumed by Demand Growth

Hardware efficiency improvements (the transition from NVIDIA A100 to H100 GPUs reduced energy per floating-point operation by approximately 3x) and algorithmic optimizations (mixture-of-experts architectures reduced inference compute by 40 to 60%) were real and significant. However, the total volume of AI inference grew by an estimated 10x between 2023 and 2025. Net energy consumption increased despite per-query efficiency gains, a dynamic known as Jevons' paradox. This finding has significant implications for policy: efficiency standards alone are insufficient without absolute consumption caps or renewable energy requirements.

What's Not Working

Scope 3 Attribution Remains Unsolved

Cloud customers consuming AI-as-a-service from providers like OpenAI, Anthropic, or Google cannot determine the actual energy consumed by their specific API calls. Providers report aggregate carbon data, but per-customer or per-query attribution requires infrastructure-level telemetry that providers consider proprietary. Until API providers disclose per-request energy data or adopt standardized reporting protocols, downstream Scope 3 accounting will rely on estimates with large uncertainty bounds.

Small and Medium Enterprises Are Excluded

Current measurement tools are priced and designed for large enterprises with dedicated sustainability teams. SMEs deploying generative AI through cloud APIs, SaaS integrations, or open-source models locally have neither the budget nor the technical capacity to measure their AI environmental footprint. Given that SMEs represent over 99% of EU businesses, this coverage gap undermines the effectiveness of disclosure-based regulatory approaches.

Action Checklist

• Inventory all generative AI workloads across your organization, including third-party API usage, cloud-hosted models, and on-premise deployments
• Establish baseline emissions estimates using at least two independent measurement tools to identify methodology-driven variance
• Engage cloud providers to request workload-level carbon attribution data for AI compute
• Implement carbon-aware scheduling for flexible AI workloads (batch processing, model retraining, non-real-time inference)
• Align internal AI emissions reporting with the Partnership on AI reporting guidelines and GHG Protocol methodologies
• Include AI environmental impact assessments in procurement criteria for new AI services and platforms
• Monitor EU AI Act implementation timelines and prepare environmental transparency disclosures for high-risk AI systems
• Evaluate model efficiency optimization (distillation, quantization, mixture-of-experts) as both cost and carbon reduction strategies

Sources

• International Energy Agency. (2025). World Energy Outlook 2025: Data Centres and AI Energy Demand. Paris: IEA Publications.
• European Commission Joint Research Centre. (2025). Environmental Impact Assessment of Artificial Intelligence in the European Union. Luxembourg: Publications Office of the EU.
• Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3645-3650.
• Luccioni, A. S., Viguier, S., & Ligozat, A.-L. (2024). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Journal of Machine Learning Research, 24(253), 1-15.
• Partnership on AI. (2025). Guidelines for Reporting AI Energy and Carbon Metrics. San Francisco: PAI.
• Goldman Sachs Research. (2024). AI, Data Centers, and the Coming US Power Demand Surge. New York: Goldman Sachs.
• Google. (2025). Environmental Report 2024: Carbon-Intelligent Computing and Data Center Sustainability. Mountain View, CA: Google LLC.
• Dodge, J., et al. (2022). Measuring the Carbon Intensity of AI in Cloud Instances. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 1877-1894.