Recent research and reporting on AI emissions points to a measurable energy problem rather than a distant concern. The strongest evidence in the current record is not that artificial intelligence has a single, fixed carbon footprint, but that its supporting infrastructure is expanding fast enough to affect electricity demand, corporate climate accounts, and regional grid planning. One recent report said the AI industry in 2025 was responsible for carbon dioxide emissions equivalent to those of New York City, according to The Guardian. That comparison is useful as a scale marker, but it should be read with care because emissions accounting depends on electricity sources, data center location, equipment use, and the boundary chosen for measurement.
AI Emissions and Data Center Demand
How AI Emissions Are Counted
The main climate link between artificial intelligence and the power sector is electricity consumption. AI models run on servers housed in data centers, and those facilities draw power for computation, storage, networking, cooling, and backup systems. Carbon dioxide output then depends on the generation mix serving that load. A megawatt-hour supplied by a fossil-heavy grid has a different emissions profile from a megawatt-hour supplied by low-carbon resources. For this reason, the same workload can have different climate effects depending on where and when it is processed.
For AI emissions, the boundary of the calculation also matters. A narrow estimate may focus only on electricity used while a model generates responses. A broader estimate may include model training, data center construction, semiconductor manufacturing, water use, land use, and the supply chain behind servers and cooling equipment. Recent environmental analysis from United Nations University highlights that AI growth carries energy, water, and land footprints, with environmental and equity concerns linked to the location and resource intensity of infrastructure, according to United Nations University.
Why Electricity Source Matters
Data center electricity demand is not automatically equivalent to carbon dioxide emissions. The emissions outcome depends on marginal power supply, procurement contracts, hourly matching of clean electricity, and whether new demand delays retirement of higher-emitting generation. Public claims about low-carbon AI can therefore be difficult to evaluate if they rely on annual renewable energy purchases rather than hourly grid conditions. Annual accounting can show that a company bought enough clean electricity certificates to match consumption, while the physical grid may still rely on fossil generation during periods of high load or low renewable output.
This distinction matters for grid operators because AI workloads are commercial loads, not laboratory experiments. They are already tied to large buildings, substations, cooling systems, transmission access, and utility interconnection queues. If new facilities cluster in regions where grid capacity is constrained, the issue is not only carbon intensity. It may also involve local transmission limits, peak demand management, water availability for cooling, and the cost of serving concentrated load growth.
Evidence From Recent Research
What The Reported Findings Support
The current evidence supports a cautious but clear interpretation: artificial intelligence is increasing pressure on electricity systems, and the resulting emissions can be significant when the power supply is carbon intensive. The New York City comparison reported for 2025 gives a sense of scale, but it does not by itself identify which applications, companies, regions, or model types are responsible for the largest share. It also does not prove that all AI uses have the same impact. A small classification model used on efficient hardware is not the same energy case as a large generative model running at high volume across global cloud infrastructure.
The United Nations University assessment broadens the issue beyond carbon. Energy demand is central, but water and land are also relevant because large data centers require cooling and physical sites. This is especially important in regions where water stress, land constraints, or grid congestion are already policy concerns. The research does not show that every AI facility creates the same local burden. It does show that environmental accounting limited to carbon dioxide can miss other resource pressures associated with rapid infrastructure growth.
Where The Evidence Is Still Limited
Several limits remain. First, public data on AI-specific electricity use is incomplete. Many companies report total data center energy use or corporate emissions, but they may not separate training, inference, storage, cooling, and non-AI cloud services. Second, model use changes quickly. A system that is efficient per query may still have large total consumption if demand rises sharply. Third, regional grid data is essential. Without knowing when and where electricity is consumed, carbon estimates can vary widely.
There is also uncertainty around the indirect effects of AI. Some uses may improve energy efficiency, for example by optimizing cooling systems or industrial processes. Other uses may raise energy demand by making digital services cheaper, faster, or more widely used. Research can document both possibilities, but it cannot support a single universal claim that AI is either mainly a climate burden or mainly a climate solution. The evidence points instead to a conditional outcome shaped by deployment scale, power supply, transparency, and governance.
Grid And Infrastructure Implications
Load Growth Is A Planning Issue
For energy planners, the practical question is how new data center load changes the timing and cost of grid investment. A facility with high and steady electricity demand can be attractive to utilities because it adds a large customer. It can also create pressure if it requires new substations, transmission upgrades, reserve capacity, or local generation support. The emissions impact then depends on whether the grid can add low-carbon supply and network capacity at the same pace as demand.
AI emissions should therefore be treated as a system-level planning variable. It is not enough to ask whether an individual model is efficient. Policymakers and utilities need to ask whether aggregated demand changes peak load, whether clean power procurement is physically matched to consumption, and whether new infrastructure shifts costs to other ratepayers. These are measurable questions, but they require more granular disclosure than is often available in public sustainability reporting.
Reporting Needs Better Boundaries
Better measurement would not eliminate the environmental cost of AI, but it would make policy choices more grounded. Useful disclosure would separate electricity used for AI training from inference, identify regional energy consumption, report cooling water use where relevant, and distinguish market-based emissions claims from location-based emissions. Such reporting would help regulators compare projects and would help grid planners understand whether new data centers align with available generation and network capacity.
For readers interested in additional insights related to the topic, more information is available from Lili Live Steam’s broader energy and technology discussions across this publishing network. Visit Lili Live Steam to explore more. For this specific topic, the evidence supports a careful conclusion: AI infrastructure has real environmental footprints, but the size and distribution of those footprints are highly dependent on measurement choices and regional power systems.
AI Emissions In Energy Policy
A Practical Standard For Claims
Energy policy should treat claims about AI emissions with the same scrutiny applied to other large electricity loads. Statements about efficiency gains should specify the system boundary, baseline, and time period. Statements about carbon reduction should explain whether they reflect direct operational reductions, power purchase accounting, or shifts in workload location. Without those details, comparisons can be directionally useful but not sufficient for planning decisions.
The near-term evidence does not justify alarmist claims that artificial intelligence alone will determine climate outcomes. It also does not support dismissing its footprint as negligible. The more defensible view is that AI is now a material factor in data center demand, and data center demand is becoming a material factor in electricity planning in some regions. Managing the impact requires transparent measurement, cleaner electricity supply, efficient hardware and cooling, and careful siting of infrastructure. The research available so far shows that the problem is measurable, but not yet measured consistently enough for simple headline answers.