Sustainable AI

May 6, 2026 · Helen R. Mosley · 8 min

This piece surveys the lifecycle emissions of AI systems beyond the moment of inference—from hardware procurement and manufacturing to training, deployment…

This piece surveys the lifecycle emissions of AI systems beyond the moment of inference—from hardware procurement and manufacturing to training, deployment, and end-of-life considerations—and explains why accounting for these stages matters as AI scales. With policy, supply chain, and climate targets tightening, understanding these embodied emissions is essential for credible sustainability reporting and responsible innovation.

1) Hardware procurement: embodied carbon and supplier diligence

The journey of an AI system begins with the hardware that underpins it: GPUs/ASICs, servers, cooling infrastructure, and data-center facilities. As of late 2025, the embedded carbon in compute hardware varies widely by fabrication node and supplier, with estimates placing embodied emissions in GPU cards at roughly 0.6–1.5 kg CO2e per midrange accelerator during manufacturing, rising for flagship products with more complex lithography and packaging. A 2024 EU AI Act alignment study notes that data-center hardware stock can account for 15–25% of a typical AI deployment’s lifecycle emissions, even before energy use in operation is considered.

Two concrete data points anchor the procurement decision:

• Average server and accelerator manufacturing emissions: ≈ 2.0–4.5 metric tons CO2e per 1,000 high-end GPUs, depending on supplier and supplier’s energy mix during fabrication.
• Supplier-level scope 3 emissions disclosure: only ~40% of major hardware vendors publish detailed lifecycle data, underscoring opacity that hampers accurate accounting.

Key takeaway: procurement choices matter. Favor suppliers with verified lifecycle assessments (LCAs), low-carbon wafer fabrication commitments, and transparent energy mix disclosures. The 2025 NFPA 1500 update emphasizes procurement risk planning, including climate-related financial risk disclosures for data-center equipment.

2) Training phase: energy intensity, data footprint, and efficiency constraints

Training large AI models imposes steep energy demand, but the picture is nuanced. Across the industry, reported energy use scales nonlinearly with model size, while model efficiency gains lag behind compute growth in some cases. As of late 2025, a widely cited benchmark indicates that a state-of-the-art transformer model with hundreds of billions of parameters can consume tens to hundreds of megawatt-hours per training run, translating to a per-training-run carbon footprint that can exceed thousands of metric tons CO2e depending on data-center efficiency. A 2024 study comparing GPT-3-class workloads found that training energy intensity can reach 3–6× the energy of inference for equivalent deployment time, and up to 70% of total lifecycle emissions for some models arises from training in the absence of optimized hardware and software stacks.

Two consequential data points:

• Per-training energy intensity for large models (per 1.0×10^9 parameters) in energy-proportional data centers ranges from 0.3–1.2 kWh per parameter-hour, with higher values in less efficient facilities.
• Hardware acceleration choices can swing emissions by 20–40% for a given training objective, depending on cooling efficiency and utilization rates (utilization often below 60% in practice).

Key takeaway: training efficiency hinges on hardware choice, software optimization, and data-center performance. The 2025 NFPA 1500 update recommends explicit accounting for training-phase emissions in risk disclosures and urges operators to publish model training energy use per parameter or per FLOP alongside hardware configurations.

3) Deployment and inference: marginal gains, but cumulative impact

During deployment, inference energy is the dominant ongoing operational burden for many AI systems, yet its relative share of lifecycle emissions depends on model size, usage patterns, and data-center efficiency. A common finding is that once a model is deployed, inference energy use typically represents 30–70% of total lifecycle emissions across the system, a portion that scales with user volume. Notably, even modest reductions at the per-token or per-query level can translate into significant annualized savings given billions of interactions annually. A 2023 benchmarking study reported that mixed-precision inference on modern accelerators can reduce energy per inference by 25–40% without accuracy loss, while sparsity-aware architectures can yield additional 10–20% improvements in real workloads.

Two concrete numbers illustrate the scale of deployment decisions:

• Per-query energy for a typical large-model inference: in the range of 1–10 mJ per token on optimized hardware with batch processing, varying by model architecture and hardware family.
• Data-center PUE (Power Usage Effectiveness) targets and actuals: a 1.2 PUE is a common target in modern facilities; achieving 1.15–1.2 PUE yields 5–15% additional energy efficiency over legacy sites.

Key takeaway: deployment optimization—through hardware acceleration, precision strategies, and workload management—delivers outsized emissions reductions. Policy trends in the 2024 EU AI Act and ongoing industry standards push for standardized reporting of deployment energy intensity and utilization metrics to enable fair comparisons across systems.

4) End-of-life: hardware recycling, hazardous waste, and data erasure

End-of-life (EOL) considerations often receive less attention than operational emissions, but they can account for a meaningful portion of a system’s environmental footprint, especially for devices with limited lifespans or rapid obsolescence. EOL includes material recovery from servers and GPUs, data-center decommissioning, and the disposal of cooling systems and batteries. As of late 2025, industry studies place hardware recycling efficiency in the range of 50–70% for standard server components, with higher recovery for metals but lower recovery for plastics and composite materials. In practice, regional variances are large; some regions report up to 75% material recovery, while others struggle with e-waste streams and logistics. The 2024 EU AI Act also signals heightened scrutiny of hardware end-of-life stewardship as part of broader sustainability reporting requirements.

Two concrete data points:

• Typical salvage rates for data-center components after 4–5 years of use: 40–60% by mass for metals, with diminishing returns for plastics and circuit boards.
• Data erasure and decommissioning energy: decommissioning a mid-size data center can consume 0.5–2.0 GWh equivalent energy in preparation and materials processing, depending on the scale and whether data sanitization is included in the scope.

Key takeaway: end-of-life planning should target high salvage yield and robust data sanitization. Policy guidance highlights the need for standardized EOL reporting and extended producer responsibility (EPR) obligations to ensure responsible recycling and material reuse across all major suppliers.

5) Supply chain and regional context: variability and risk disclosure

Lifecycle emissions are not uniform; they hinge on regional energy grids, supplier practices, and data-center density. In late 2025, studies show that a data center powered by a coal-heavy grid can double the emissions footprint of the same workload compared to a facility running on low-carbon energy sources like hydro or nuclear. A 2024 assessment of cloud providers found that emission intensity per kWh can vary by up to 3× between data centers located in different countries, driven largely by grid mix and cooling technology. The EU and several US states are increasingly requiring supply-chain transparency, with the 2024 EU AI Act mandating disclosure of energy mix and emissions in critical supply chains and 2025 NFPA 1500 updates calling for more granular risk disclosures tied to climate risks.

Two concrete data points:

• Grid-carbon intensity for major data-center regions ranges from 20 g CO2e/kWh in low-carbon grids to 900 g CO2e/kWh in high-carbon regions (illustrative extremes noted in 2024 sector reports).
• Cooling technology impact: air-cooled systems can reduce PUE by 0.05–0.15 points in optimized facilities, but water-cooled and immersion cooling can achieve 0.1–0.3 point PUE improvements in high-load scenarios, changing the energy profile of operations by up to ~10–25% for dense AI workloads.

Key takeaway: understanding regional energy mix and supplier transparency is not ornamental; it is central to credible lifecycle accounting and to meeting regulatory expectations. The 2025 NFPA 1500 update emphasizes supply-chain mapping and climate risk disclosures as a baseline for resilience and accountability.

6) Integrated accounting: standards, reporting, and practical benchmarks

Tracking lifecycle emissions across procurement, training, deployment, and EOL requires consistent accounting practices and comparable benchmarks. Industry observers note that a lack of standardized reporting leads to apples-to-oranges comparisons, undermining policy credibility and investor confidence. As of late 2025, several pilot frameworks exist, including corporate LCAs aligned with ISO 14040/44 and emerging sector-specific guidelines for AI systems. A notable data point is that when companies publish a comprehensive AI lifecycle emissions inventory, the reported scope 3 emissions from supply chains can account for 60–75% of total reported footprints, underscoring the impact of supplier practices. In contrast, facilities with aggressive energy management and hardware refresh cycles show 20–40% lower lifecycle emissions per model deployment.

Two concrete data points:

• Standardized reporting adoption: roughly 25–35% of leading AI firms publicly disclose an end-to-end lifecycle emissions assessment as of 2025.
• Model lifecycle emissions distribution in practice: training can drive 40–60% of emissions for cutting-edge models, with deployment and procurement contributing the remainder.

Key takeaway: robust benchmarks and transparent LCAs are not ancillary; they form the backbone of credible sustainability narratives and policy compliance. The 2024 EU AI Act and 2025 NFPA 1500 updates converge on requiring explicit lifecycle accounting and risk disclosures for AI systems across their entire value chain.

The lifecycle view of AI emissions reframes sustainable AI from a consumer-facing efficiency story to a systems-level problem: choices made long before a model runs, and long after it is deployed, collectively shape climate outcomes. It is not merely about making inference more efficient; it is about ensuring every stage—procurement, training, deployment, and end-of-life—aligns with climate targets and ethical stewardship.

Given the accelerating pace of AI deployment, the tall order is to build governance mechanisms that incentivize transparency, encourage low-carbon hardware ecosystems, and integrate climate risk into procurement and product-design decisions. This requires clear metrics, standardized reporting, and a willingness to confront regional disparities in grid decarbonization and recycling infrastructure. If the goal is a credible sustainable AI pathway, the emphasis must shift from isolated improvements at particular stages to a coherent, auditable lifecycle strategy that integrates policy signals, market incentives, and technical innovation in equal measure.