Artificial intelligence is increasingly being positioned as a tool for reducing emissions across industries, but the technology now faces a harder question than whether it can improve efficiency. The question is whether those improvements remain meaningful after accounting for the electricity, water and computing infrastructure required to develop and operate AI systems. GlobalData’s latest analysis argues that AI can contribute to targeted emissions reductions in energy, agriculture and buildings, yet those gains cannot automatically be treated as a net environmental benefit. The concern becomes sharper as organizations increasingly deploy larger generative models that can require substantially more computing capacity than many traditional predictive AI applications. That shift places data centers, cooling systems and inference workloads inside the climate equation rather than treating them as an invisible layer underneath software.

The issue matters because evaluating what algorithms can optimize also requires consideration of the infrastructure needed to support them. An AI system that improves a power-grid forecast, reduces unnecessary agricultural inputs or lowers building energy consumption can create measurable resource savings. Those savings, however, exist alongside the energy required for model development, data processing, inference and the cooling systems that keep computing equipment within operating limits. Water adds another layer because cooling requirements can create environmental pressure in locations where data center expansion coincides with constrained water resources. The resulting calculation is not simply whether an AI application produces an efficiency gain, but whether that gain exceeds the environmental burden created across its operating chain. 

Predictive AI Carries the Stronger Climate Record

GlobalData draws a meaningful line between predictive AI and the newer generation of generative systems. Predictive models can analyze patterns and anticipate changes in systems such as electricity networks, agricultural operations and buildings, while their computing requirements can differ substantially from those of large frontier generative models. Renewable-grid optimization can help operators better match supply and demand, while agricultural applications can support more efficient resource use. Building-management applications can identify opportunities to reduce energy consumption by responding to operating conditions and demand patterns. These applications connect AI directly to physical efficiency outcomes that organizations can potentially measure against defined environmental performance metrics. The distinction matters because it shifts the discussion away from the broad claim that AI is inherently climate-positive and toward the narrower question of which workloads deliver demonstrable environmental improvements.

Aoife McGurk, Senior Analyst at GlobalData Strategic Intelligence, makes the company’s position explicit: “AI can support climate mitigation, but it will never be the whole solution. For AI technology to have a net positive impact on the planet, the efficiencies it generates must outweigh the environmental harm caused by data centers and inference.” The statement places the computing layer alongside the intended environmental outcome rather than treating infrastructure as an externality. It recognizes that the climate value of an application depends partly on the resources consumed by the machinery running it. That becomes particularly important when organizations compare optimization workloads with larger models whose computational requirements can vary significantly by model and application. The resulting environmental balance can vary significantly even when both systems sit under the broad label of artificial intelligence.

Generative AI Changes the Environmental Equation

Generative AI introduces a different operating profile because large language models and other frontier systems can require significant computing resources during both development and inference. The scale of that infrastructure can increase as companies move from limited experimentation toward broader enterprise deployment. Each additional inference workload can create incremental computational demand, making inference an important part of the environmental accounting rather than a one-time model-development expense. GlobalData says the demonstrated climate advantages associated with AI currently come overwhelmingly from traditional predictive applications rather than energy-intensive generative AI and large frontier models. The company further says there is limited evidence that generative AI’s environmental benefits can compensate for its higher emissions. However, the question is not whether generative AI has any potential climate applications, but whether individual deployments can demonstrate benefits large enough to justify their resource requirements.

McGurk describes that difference in direct terms: “Predictive AI can help improve efficiency and reduce resource use across key sectors, through renewable grid optimization, sustainable agriculture, and building energy use management. Generative AI, on the other hand, produces significantly higher emissions, and there is little evidence that it mitigates climate change.” The statement does not eliminate the possibility that generative systems could support useful climate applications. Instead, it raises the evidentiary threshold for claiming an environmental benefit from those systems. A sustainability program that uses a large language model, for example, would need to establish what emissions or resource consumption the system actually reduces and compare that outcome with the computing required to produce it. Without that accounting, a technology’s climate narrative can become detached from its physical footprint.

Data Centers Become Part of the Climate Calculation

Data centers are central to this calculation because they translate digital demand into physical requirements for electricity, cooling, equipment and, depending on the cooling system, water. AI workloads can increase computational density, which can alter the power and thermal characteristics of facilities even when the application itself exists entirely in software. Cooling infrastructure must remove the heat generated by computing equipment, and the environmental consequences depend on the technology used, local climate conditions and available resources. Water consumption can therefore become a regional issue rather than a purely global metric, particularly when new computing capacity enters areas already facing resource constraints. Carbon impacts can likewise vary according to how electricity is generated and when computing demand occurs. Therefore, an AI climate assessment that considers only the emissions avoided by an application can miss the infrastructure emissions associated with delivering that application.

The same logic applies to inference, which can become a persistent operating cost as AI moves deeper into enterprise workflows. Training attracts considerable attention because it represents a significant concentration of computing activity, while inference can continue after a model enters production. A model used across customer service, analytics, software development or internal decision-making can therefore create a recurring infrastructure requirement. The environmental balance depends on the workload’s frequency, model size, hardware efficiency, utilization and cooling requirements, among other variables. Organizations that deploy AI at scale will need to connect those operational metrics with the environmental outcomes they claim to achieve. A climate strategy that treats computing consumption as separate from business performance risks overlooking the rapidly growing electricity demand associated with AI-focused data centers.

Sustainability Plans Face a New Compliance Risk

GlobalData is particularly cautious about companies using generative AI as a central component of corporate sustainability programs. Large language models can help process documents, generate analysis and support reporting workflows, but their outputs can contain inaccurate information. In a regulatory environment where companies must substantiate environmental claims, inaccurate AI-generated material can create more than a technical problem. It can complicate compliance processes and expose organizations to reputational consequences if environmental claims rely on information that cannot withstand scrutiny. The environmental footprint of the model adds another layer because the organization must account for both the reliability of the output and the resources required to generate it. GlobalData warns that these combined risks could contribute to greenwashing when companies present AI-enabled sustainability claims without sufficient evidence.

That concern changes how executives can evaluate AI inside environmental programs. Instead of assuming that an AI deployment qualifies as a sustainability initiative because it improves an operational process, companies can establish measurable environmental criteria before scaling it. The first test is whether the deployment produces a tangible reduction in greenhouse-gas emissions or ecosystem degradation. The second is whether the environmental harm associated with operating the AI system remains below the intended climate benefit. These tests create a direct connection between technology procurement and environmental performance. They can further encourage organizations to examine model selection, workload requirements, infrastructure efficiency and cooling considerations before declaring an AI project climate-positive.

Smaller Models Could Change the Economics

The response does not necessarily involve abandoning generative or agentic AI, but can instead focus on limiting unnecessary computational requirements and reducing the environmental harm associated with these systems. McGurk recommends safeguards including smaller language models, automatic model triage, carbon-aware computing, edge infrastructure, emissions budgets, algorithmic efficiency improvements and novel data center cooling. Those measures address different parts of the AI infrastructure chain, from deciding which model should handle a task to reducing the resources consumed by the underlying computing environment. Model triage can prevent large systems from handling workloads that smaller models can complete adequately. Carbon-aware computing can shift selected workloads toward periods or locations with lower associated emissions, while edge infrastructure can reduce some data movement and latency requirements.

McGurk concludes: “If a deployment of generative or agentic AI would deliver significant climate mitigation benefits, companies should do what they can to limit the environmental harm caused by using these systems. Strict safeguards—such as using small language models, automatic model triage, carbon-aware computing, edge infrastructure, emissions budgets, algorithmic efficiency improvements, and novel data center cooling—can prevent AI-enabled ESG strategies from inadvertently accelerating the climate crisis.” The recommendation effectively turns model architecture and infrastructure design into components of corporate climate strategy. It suggests that the environmental performance of AI should be considered alongside decisions about workload requirements, computing efficiency and the infrastructure needed to operate these systems. The approach could encourage companies to treat computational efficiency as an operational sustainability consideration rather than merely an engineering optimization. Ultimately, that could make AI deployments easier to evaluate against measurable environmental outcomes.