Is AI a threat to the environment?

The Real Scale of AI's Energy and Environmental Footprint

AI's environmental footprint is real, measurable, and growing fast — but the numbers behind the headlines deserve closer scrutiny. Data centres consumed roughly 415 TWh of electricity in 2024, about 1.5% of global electricity, and the IEA projects this will roughly double to ~945 TWh by 2030. That growth is more than four times faster than every other sector of the economy.

Beyond electricity, the broader footprint is substantial. A peer-reviewed review in Patterns estimates AI's 2025 carbon footprint at 32.6–79.7 million tons of CO2, with water use between 312.5 and 764.6 billion liters. Notably, most of that water isn't on-site cooling — it's consumed indirectly by the thermal and hydropower plants supplying the data centres.

Yet some widely shared figures don't hold up. A study in Joule found that high-profile per-query estimates overstate AI inference energy by 4–20×, largely because they assume non-production settings. In real deployments, frontier-scale inference uses a median of just 0.31 Wh per query.

The genuine emerging concern isn't simple chatbot prompts — it's agentic AI. Long reasoning queries consume roughly 13× more energy than standard ones, and if even 10% of daily requests fall into that category, total demand more than doubles. For anyone building with AI agents, workload design now has direct energy consequences.

Several factors shape the true scale of AI's footprint:

• Geographic concentration: the US and China account for nearly 80% of data centre growth through 2030, straining local grids.
• Workload type: AI-generated video uses at least 1,000× more energy than a chatbot prompt.
• Efficiency gains: per-task power consumption is falling "at a rate unprecedented in energy history," per the IEA, with 8–20× per-query reductions identified as achievable.
• Opacity: operators don't disclose the data needed to measure impact precisely, so even the best estimates remain approximations.

That last point matters most. "You can't solve a problem if you can't measure it," as Stanford's Peter Henderson has argued. This is the same principle behind why SkillzBench is building independent verification for AI agent skills — well-scoped instruction packages that avoid inflating token generation or triggering unnecessary reasoning loops are measurably more efficient, and efficiency is one of the few levers proven to bend this curve. The footprint is significant, but whether it becomes a crisis depends on transparency and design choices being made right now.

Why Agentic AI Is the Emerging Environmental Concern

The most consequential environmental question in AI isn't how many chatbots the world runs — it's how many of them stop to think. A peer-reviewed Joule study found that long reasoning queries, roughly 5,000 output tokens, consume about 13× more energy than a standard query, with median per-query energy rising from 0.31 Wh to 3.91 Wh at frontier scale.

For teams building and deploying AI agents, that multiplier matters more than any headline about data centre growth. The same research modeled a service handling one billion queries per day: standard queries require 0.7 GWh/day, but if just 10% of requests are long reasoning queries, total demand more than doubles to 1.7 GWh/day. A small shift in workload mix produces a disproportionate spike in energy use.

The IEA flags the same dynamic from the systems level: energy-intensive uses like AI agents are offsetting per-task efficiency gains, even as per-task power consumption falls at a rate the agency calls unprecedented in energy history. Agentic workloads are the reason efficiency gains aren't translating into flat demand.

What inflates an agent's energy footprint? In practice, several patterns recur:

• Verbose instruction packages that force the model to generate thousands of tokens before acting
• Unnecessary reasoning loops, where an agent re-derives conclusions it already reached
• Overly broad tool access that triggers exploratory chains instead of direct action
• Skills that pad context with redundant references, multiplying token generation

The encouraging news is that the problem is tractable. The Joule study identifies 8–20× "line-of-sight" per-query energy reductions across models, serving systems, and hardware — and shows that efficiency interventions can bring that 1.7 GWh/day scenario back down to 0.8 GWh/day. Concise, well-scoped agent behavior isn't just better engineering; it's a measurable environmental lever.

This is why SkillzBench treats efficiency as a quality signal, not an afterthought. As it builds its benchmarking pipeline — testing every skill in a sandbox with and without the skill to measure Lift — it will be positioned to surface which skills inflate token usage or trigger wasteful reasoning loops, alongside the security and license checks that keep agent stacks safe.

For organizations deploying agents at scale, the takeaway is direct: audit what your agents are prompted to do, how long they think before acting, and whether that reasoning earns its energy cost. In agentic AI, the biggest environmental wins often come from the smallest instruction edits.

The Counterweights: Efficiency Gains and the Green AI Movement

For all the alarming headlines about AI's energy appetite, the same research that documents the growth also documents a powerful countercurrent: efficiency gains at a pace the energy sector has rarely seen. The question isn't whether AI consumes more energy — it's whether efficiency, clean investment, and better measurement can outpace the demand curve.

The efficiency numbers are striking. A peer-reviewed analysis in Joule identifies 8–20× "line-of-sight" per-query energy reductions available across models, serving systems, and hardware — meaning substantial savings are already visible, not speculative. The IEA reports that per-task power consumption is declining "at a rate unprecedented in energy history." And there's historical precedent: the same analysis notes the current pattern echoes how data centers tempered energy demand through efficiency gains during the internet and cloud buildout.

The math illustrates how much these gains matter. At scale, 1 billion queries per day require roughly 0.7 GWh of energy daily. If just 10% of those become long reasoning queries — which consume about 13× more energy than standard prompts — demand more than doubles to 1.7 GWh per day. But efficiency interventions pull that figure back down to 0.8 GWh per day, nearly erasing the agentic-AI penalty. That finding is directly relevant to anyone building agent workflows: concise, well-scoped instructions that avoid unnecessary reasoning loops are an energy decision as much as a design decision, which is part of why SkillzBench evaluates skills in controlled benchmarks rather than assuming verbose instruction packages help.

AI is also reshaping where clean energy gets built. The IEA reports that the tech sector accounted for roughly 40% of all corporate renewable power purchase agreements signed in 2025, and conditional offtake agreements between data center operators and small modular reactor nuclear projects grew from 25 GW to 45 GW. As IEA Executive Director Fatih Birol puts it, "While AI is still an energy taker, it is also becoming an energy maker" — driving next-generation nuclear, flexible data centers, and long-duration storage.

A "Green AI" ecosystem is formalizing around these efforts. Researchers distinguish between "Green-in AI" (energy-efficient ML design) and "Green-by AI" (AI applied to environmental problems elsewhere), and practical tools are emerging:

• The AI Energy Score, developed with Hugging Face and Salesforce, rates model energy efficiency
• The CLEER methodology, with Etsy as its first corporate user, standardizes cloud emissions reporting
• Token-based carbon accounting from Watershed ties emissions to actual usage
• Corporate guidelines at companies like Okta and GitLab steer employee AI use toward efficiency

None of this means the environmental challenge is solved. Even the optimists acknowledge that measurement remains immature — automated carbon footprinting tools came within two multiples of expert answers only 77% of the time in one Watershed study. But the direction of travel is clear: efficiency gains, clean energy procurement, and emerging measurement standards give the industry credible tools to manage a footprint that would otherwise grow unchecked. The threat is real, but so is the response.

What Teams Deploying AI Agents Should Actually Do

If your team runs AI agents, the environmental question is no longer abstract — agentic workloads are the emerging risk. Research shows long reasoning queries consume roughly 13× more energy than standard queries, and even if just 10% of daily requests are long reasoning queries, total energy demand more than doubles (peer-reviewed analysis in Joule). The good news: efficiency interventions can cut that demand nearly in half at scale. The catch is knowing where your agents waste energy in the first place.

Start with skill design, not infrastructure. Verbose, poorly scoped instruction packages inflate token generation and trigger unnecessary reasoning loops — the exact pattern that drives the 13× energy penalty. Favor concise, well-scoped skills that tell the agent exactly what to do and nothing more. A skill that pads output with 5,000 tokens of reasoning when 500 would suffice is an environmental cost you pay on every single run, multiplied across every agent in your fleet.

Demand disclosure from your vendors. Operators do not publicly disclose the inputs needed to quantify AI's environmental impact (a Patterns review found), so footprint estimates remain approximations. When the Data Center Air Pollution Tracker ranked Amazon, Anthropic, Google, Meta, Microsoft, OpenAI, Oracle and xAI, none scored above 70 out of 100 on air pollution (according to a sustainability analysis). Request emissions and water data during contract negotiations — the leverage point is before you sign, not after.

Treat automated footprinting tools as a starting point, not ground truth. Tools like the AI Energy Score and token-based carbon accounting are useful, but in one Watershed study, automated footprinting landed within two multiples of expert answers only 77% of the time (sustainability reporting notes). Use them to identify trends and outliers; commission expert validation before making claims you'll publish or report.

Finally, verify and benchmark what your agents actually run. This is why SkillzBench is building sandboxed benchmarking that tests every skill with and without it installed — the resulting Lift measurement surfaces skills that inflate token usage without delivering proportional value. A skill that adds 20 percentage points of Lift is worth its tokens; one that adds noise while tripling output is not.

For teams deploying agents, the practical playbook looks like this:

• Audit your skill library for verbose, overlapping, or redundant instruction packages that inflate token generation.
• Require emissions and water disclosure from model and infrastructure vendors before contract renewal.
• Use automated footprinting tools for directional insight, then validate with expert analysis.
• Benchmark skills in a controlled sandbox to identify which ones trigger unnecessary reasoning loops.

The research is clear that efficiency is the counterweight — peer-reviewed work identifies 8–20× "line-of-sight" per-query energy reductions available across models, serving systems, and hardware (the Joule study found). Teams that treat agent efficiency as a measurable, benchmarkable property of their stack will catch the waste others miss.