The Wrong Question About AI and the Planet

May, 2026, Peter Stefanyi Ph.D., MCC

Artificial intelligence uses energy. So does almost everything else modern organizations do. The environmental test is not how many tokens we burn, but what real-world waste we avoid.

Artificial intelligence has acquired a new public charge sheet. It hallucinates. It threatens jobs. It produces mediocre content at industrial scale. And now, increasingly, it is accused of draining the grid and guzzling water.

There is truth in the criticism. Data centers are no longer a rounding error in the energy system. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global electricity demand. By 2030, that could rise to about 945 TWh, just under 3% of global electricity demand. (IEA)

The United States is already further along. A Lawrence Berkeley National Laboratory report found that US data centers used 176 TWh in 2023, or 4.4% of US electricity consumption. By 2028, the report projects 325–580 TWh, equal to 6.7% to 12.0% of US electricity use. (The Department of Energy's Energy.gov)

That is serious. It is not, however, the whole story.

The current debate often asks the wrong question: How much energy does AI use?

The better question is: What does AI replace?

An hour of AI-supported work is not competing with nothing. In business, it may be competing with a car trip, a flight, a defective production run, a truck driving half-empty, or a consultant crossing the Atlantic to deliver a workshop that could have been redesigned for remote delivery.

When that is the comparison, the environmental math changes.

AI is a grid problem before it is a planetary catastrophe

Globally, AI and data centers are still much smaller than the heavyweights of energy use: power generation, industry, transport, buildings and agriculture. Data centers at 1.5% of global electricity are not yet in the same league as steel, cement, road transport, aviation, heating, cooling or food production. (IEA)

But the growth rate matters. The IEA says data-center electricity demand is projected to grow by around 15% per year from 2024 to 2030, more than four times faster than electricity consumption from other sectors. (IEA)

So AI can be globally modest and locally disruptive at the same time.

In Northern Virginia, Ireland, Singapore, parts of the US Midwest and other data-center hubs, the issue is not abstract percentage points. It is grid connection queues, transmission constraints, power prices, backup generation, water permits and local political resistance.

That is why the US is the canary in the coal mine. If data centers reach 6.7% to 12.0% of US electricity by 2028, they are no longer a niche tech issue. They are infrastructure policy. (The Department of Energy's Energy.gov)

Water follows the same pattern. Globally, agriculture dominates freshwater withdrawals. FAO puts the global split at about 69% agriculture, 19% industry and 12% municipal use. Another FAO source estimates roughly 3,600 km³ of freshwater withdrawals per year. (FAOHome)

Against that global total, AI water use is small. A 2025 peer-reviewed estimate put the water footprint of AI systems alone at 312.5–764.6 billion liters per year. That equals roughly 0.009% to 0.021% of global freshwater withdrawals. (ScienceDirect)

But local water politics do not care about global percentages. LBNL estimates US data centers used about 66 billion liters of onsite water in 2023, plus nearly 800 billion liters indirectly through electricity generation. Those figures are small compared with global agriculture. They are not small if a hyperscale facility is placed in a hot, drought-prone region with stressed aquifers. (eta-publications.lbl.gov)

The first correction, then, is this:

AI is not yet one of the world’s largest environmental loads. But it is one of the fastest-growing, densest and most politically visible new electricity loads.

The trap: measuring AI by how much we consume

A new workplace habit captures the problem: token-maxxing.

The term describes the push to maximize AI token consumption, sometimes treating usage volume as a sign of productivity. Recent reporting describes workers and teams competing to spend more tokens, trigger more AI requests, or climb internal usage dashboards. (Axios)

As an adoption metric, this is almost exactly wrong.

Tokens measure consumption. They do not measure value.

A person can burn tokens by asking better questions, or by asking worse ones repeatedly. An autonomous agent can consume millions of tokens solving a real engineering problem, or generate a mountain of useless activity. A company can report higher AI usage while creating no better products, no faster decisions, no lower waste and no happier customers.

That is the same mistake companies made with email, meetings and dashboards: count the activity because it is easy to measure, then confuse activity with work.

For environmental management, token-maxxing is even worse. It rewards exactly what should be disciplined: compute without a clear value test.

The better metric is not tokens used.

It is value per kilowatt-hour.

Or more concretely: waste avoided per model run.

Moving people is expensive carbon

A recent article on the carbon footprint of coaching delivery is useful because it does something rare: it compares delivery modes. The authors estimate one hour of audio coaching at 18.6 g CO₂e, audio-video coaching at 66 g CO₂e, AI coaching at 288 g CO₂e, a local 10-mile car trip at about 4 kg CO₂e, and an international 1,000-mile trip at about 384 kg CO₂e.

The structure is right: compare the mode of delivery, not the technology in isolation.

The AI estimate in that paper is probably too high for ordinary current text AI. Google researchers recently estimated that the median Gemini Apps text prompt uses 0.24 Wh of energy and about 0.26 ml of water, while Epoch AI has estimated a typical GPT-4o text query around the same order of magnitude. (arXiv)

But the deeper lesson survives. The largest variable is usually not whether the session uses a digital platform or an AI assistant. The largest variable is whether people travel.

Consider a modest meeting. Three people attend. One is already onsite. Two drive 100 km each way, a 200 km round trip per traveler. Using UK Government 2025 greenhouse-gas conversion factors as the reporting basis, an average petrol car is roughly 0.209 kg CO₂e per km when direct fuel combustion and fuel-supply emissions are included. Two travelers therefore produce about 84 kg CO₂e just getting to and from the meeting. (GOV.UK)

The online equivalent, even with video and AI support, is plausibly below 1 kg CO₂e for the group. To avoid false precision, I use a broad working range: 0.10 kg CO₂e per person-hour for ordinary AI-supported online work and 0.35 kg CO₂e per person-hour for heavier AI-supported work. Those assumptions are intentionally conservative relative to measured text-prompt estimates and include the device, video meeting and AI support.

Now scale up.

A six-person team day, with each person driving 100 km each way, produces about 251 kg CO₂e from car travel alone. The AI-supported online version is roughly 3.6–12.6 kg CO₂e under the assumptions above.

A three-day training for 15 people flying within Europe can easily reach 4–5 tonnes CO₂e from flights. The AI-supported online version is measured in tens of kilograms.

If 15 people fly economy between Europe and the United States, the flights alone are around 29 tonnes CO₂e. In business class, the same trip can rise to about 84 tonnes CO₂e, because premium seats occupy more aircraft capacity per passenger. (GOV.UK)

This does not mean every meeting should become virtual. Physical presence has real value: trust-building, conflict resolution, embodied learning, sensitive strategy work and deep team formation.

But flying or driving people for routine information transfer, weakly designed training, status updates or slide-reading is environmental waste with a calendar invite attached.

AI can make substitution more credible. It can summarize preparation material, personalize learning, support translation, generate exercises, structure follow-up, assist coaching reflection and keep distributed teams aligned. The environmental win is not that AI is magical. The win is that it can help remote work become good enough to replace some travel.

The strongest case is not email. It is physical waste.

The most serious environmental case for AI is not writing faster memos. It is reducing waste in physical systems.

Take automotive parts.

Suppose a supplier makes 500,000 door handles, locks or latch assemblies per year at an average value of €100 per unit. That is €50 million of production. If the plant has a 3% scrap rate, it scraps 15,000 units a year.

Now assume AI-assisted inspection, process monitoring or root-cause analysis reduces scrap by only 10% relative. Not 10 percentage points. Just a 10% reduction in the existing scrap. The plant avoids scrapping 1,500 units.

At €100 each, that is €150,000 of avoided production waste. If each scrapped unit carries a conservative embodied footprint of 3–10 kg CO₂e, the avoided emissions are 4.5–15 tonnes CO₂e per year. A small AI inspection system using 4,380 kWh/year would produce about 2 tonnes CO₂/year on the 2024 global-average electricity intensity of 445 g CO₂/kWh. (IEA)

The exact result will vary by material, part weight, process energy, defect rate and electricity source. But the principle is stable: avoiding physical scrap can outweigh the compute footprint.

The same logic applies to logistics.

Consider a medium fleet: 20 diesel trucks, each driving 80,000 km per year. That is 1.6 million km annually. At 30 liters of diesel per 100 km, the fleet burns about 480,000 liters of diesel. The US Energy Information Administration puts diesel CO₂ at 10.21 kg per US gallon, or about 2.70 kg per liter. That fleet therefore emits roughly 1,300 tonnes CO₂ per year from fuel combustion. (U.S. Energy Information Administration)

If AI-supported scheduling, routing and load planning save just 5%, the avoided emissions are about 65 tonnes CO₂ per year. A 10% saving is about 130 tonnes CO₂.

This is not fantasy. UPS’s ORION routing system was expected to reduce annual driving by 100 million miles, save 10 million gallons of fuel, and avoid 100,000 metric tons of greenhouse-gas emissions once implemented across the US. (bsr.org)

The serious environmental case for AI is therefore not “AI is green.”

It is more precise:

Small improvements in physical systems can dwarf the footprint of the software that enables them.

The counterargument is real

This argument has a shadow side.

If AI is used to replace travel, scrap, fuel, rework and downtime, it can be environmentally positive.

If it is used to create synthetic spam, redundant content, unnecessary video, low-value

automation and uncontrolled agents, it becomes a new waste machine.

The unit of responsibility is not the prompt. It is the workflow.

A company that gives every employee a chatbot and celebrates prompt volume has not adopted AI responsibly. It has adopted an energy-consuming novelty meter.

A company that uses AI to eliminate failed site visits, cut scrap, reduce empty truck miles, shorten troubleshooting and replace routine travel has a different environmental story.

This is why tokenmaxxing is the wrong symbol for the AI era. It confuses consumption with capability. It encourages people to burn compute to prove they are “using AI.” In sustainability terms, it is like measuring manufacturing excellence by how much raw material a factory consumes.

A better AI metric would ask:

• How much travel was avoided?

• How much scrap was prevented?

• How much fuel was saved?

• How much rework disappeared?

• How much expert time moved from administration to problem solving?

• How much value was created per kilowatt-hour?

The right metric is not tokens per employee.

It is value per kilowatt-hour.

What responsible AI looks like

The practical rules are simple.

Use AI first where it replaces something heavier: travel, rework, scrap, downtime, fuel, emergency logistics, failed meetings. Use the smallest model that reliably does the job. Prefer text when text is enough. Put budgets and stop rules on agents. Cache and reuse outputs instead of regenerating the same work. Treat image, video and autonomous multi-step workflows as higher-impact tools that need stronger justification.

And do not stop at user behavior. The infrastructure matters.

Enterprise buyers should demand AI-specific energy reporting, location-based carbon data, water-usage metrics, cooling disclosure, hourly clean-energy matching and evidence that data centers are not drawing potable water in water-stressed regions.

The EU has already moved in this direction. Under the revised Energy Efficiency Directive, data centers above 500 kW installed IT power demand face reporting obligations. (IEA)

Data centers should be pushed toward low-water cooling, closed-loop systems where feasible, recycled or non-potable water where appropriate, cleaner power procurement, better utilization and waste-heat reuse. In cold climates and dense urban areas, data-center heat can become an input to district heating rather than an unwanted byproduct.

A better environmental standard

The AI energy debate is real. It is also often lazy.

It is lazy to pretend AI is immaterial because a single prompt is small. At data-center scale, the load is growing fast enough to affect grids, water systems and climate targets.

It is equally lazy to treat every AI use as environmental sin. A one-hour AI-supported meeting may replace hundreds of kilometers of car travel. A virtual training may replace 15 flights. A diagnostic assistant may prevent an engineer from flying to inspect a machine. A routing system may save more diesel in a week than its compute uses in a year.

The environmental question is not whether AI consumes resources. It does.

The question is whether AI helps us consume less of something larger.

Used badly, AI becomes a factory for noise. Used well, it can reduce the movement of bodies, the scrapping of materials, the burning of fuel and the repetition of low-value work.

That should be the standard for responsible adoption:

Not more AI everywhere. Better AI where it removes real-world waste.

And the metric should be just as clear:

Value per kilowatt-hour. Waste avoided per model run. Not tokens burned.

Appendix:

1. Global and US context

Metric	Audited value	Source / note
Global data-center electricity, 2024	415 TWh	IEA
Share of global electricity, 2024	~1.5%	IEA
Projected global data-center electricity, 2030	945 TWh	IEA base case
Projected 2030 share	Just under 3%	IEA
US data-center electricity, 2023	176 TWh	LBNL / DOE
US share, 2023	4.4%	LBNL / DOE
US projected data-center electricity, 2028	325–580 TWh	LBNL / DOE
US projected share, 2028	6.7–12.0%	LBNL / DOE
Global freshwater withdrawals	~3,600 km³/year = 3.6 quadrillion L/year	FAO
Global water split	69% agriculture, 19% industry, 12% municipal	FAO AQUASTAT
AI-only water estimate, 2025	312.5–764.6 billion L/year	Patterns, 2025
AI water as % of global withdrawals	0.009–0.021%	Calculation: AI water / 3.6 quadrillion L
US data-center onsite water, 2023	66 billion L/year	LBNL
US data-center indirect water, 2023	~800 billion L/year	LBNL
US total water withdrawals, 2015	322 billion gal/day = ~445 trillion L/year	USGS
US data-center onsite water as % of US withdrawals	~0.015%	66B L / 445T L
US data-center indirect water as % of US withdrawals	~0.18%	800B L / 445T L

2. Work-practice comparisons

Assumptions:

• Typical AI-supported online work: 0.10 kg CO₂e/person-hour

• Heavy AI-supported online work: 0.35 kg CO₂e/person-hour

• Average petrol car including fuel supply: ~0.209 kg CO₂e/km

• Medium-haul air travel: ~0.153 kg CO₂e/passenger-km

• Long-haul economy: ~0.142 kg CO₂e/passenger-km

• Long-haul business: ~0.411 kg CO₂e/passenger-km

Scenario	Calculation	Result
3-person online meeting, typical AI	3 × 1h × 0.10	0.3 kg CO₂e
3-person online meeting, heavy AI	3 × 1h × 0.35	1.05 kg CO₂e
3-person onsite meeting, two people drive 200 km each	2 × 200 × 0.209	83.6 kg CO₂e
6-person online team day, typical AI	6 × 6h × 0.10	3.6 kg CO₂e
6-person online team day, heavy AI	6 × 6h × 0.35	12.6 kg CO₂e
6-person onsite team day, all drive 200 km	6 × 200 × 0.209	250.8 kg CO₂e
15-person online 3-day training, typical AI	15 × 18h × 0.10	27 kg CO₂e
15-person online 3-day training, heavy AI	15 × 18h × 0.35	94.5 kg CO₂e
15-person EU flight training	15 × 2,000 km × 0.153	4.59 t CO₂e
15-person transatlantic economy training	15 × 13,600 km × 0.142	29.0 t CO₂e
15-person transatlantic business training	15 × 13,600 km × 0.411	83.8 t CO₂e

3. Industrial examples

Use case	Calculation	Result
Automotive parts production	500,000 units × €100	€50M annual production value
Baseline scrap	3% × 500,000	15,000 scrapped units/year
AI-enabled improvement	10% relative reduction in scrap	1,500 units saved/year
Financial waste avoided	1,500 × €100	€150,000/year
Embodied CO₂ avoided	1,500 × 3–10 kg CO₂e	4.5–15 t CO₂e/year
Example AI inspection system	4,380 kWh × 0.445 kg CO₂/kWh	~2.0 t CO₂/year
Logistics fleet	20 trucks × 80,000 km	1.6M km/year
Diesel use	1.6M km × 30 L/100 km	480,000 L/year
Diesel CO₂	480,000 L × ~2.70 kg CO₂/L	~1,300 t CO₂/year
5% optimization saving	5% × 1,300 t	~65 t CO₂/year
10% optimization saving	10% × 1,300 t	~130 t CO₂/year