FutureIQ 209: “How Bad is ChatGPT for the Environment”
FutureIQ episode 209 argued that various claims about AI’s environmental cost are mostly inflated. This page checks that argument against published sources — IEA, FAO, EPA, USGS, UNESCO, the World Bank, peer-reviewed papers and the AI companies’ own disclosures — one claim at a time.
Where a figure from the episode doesn’t hold up, the note explains where the number most likely came from and what the evidence actually supports. Most of it holds up; a few things don’t, and those are marked plainly.
Fact check generated on 18 September 2026 · Prompt-level measurements are system-, date-, location- and accounting-boundary-specific, so figures from different sources are not always directly comparable.
People are very worried about ChatGPT's electricity usage, and water usage, and environment impact.
Checks out
The concern itself is documented: the IEA says data centres are a visible flashpoint for public concern over energy prices and the environment. IEA 2026 executive summary
You may have heard that using one ChatGPT prompt takes 5 litres of water. NOT TRUE.
Exaggeration
No published estimate puts a single prompt anywhere near 5 litres, so the verdict is right even if the figure being knocked down is larger than anything in the literature. The likely origin is a chain of misquotation: Li et al. (2023) estimated roughly 500 mL per 10–50 responses, which circulated as “500 mL per query”; the Washington Post and UC Riverside later estimated ≈519 mL for a 100-word GPT-4 email, rising to ≈1,408 mL in Washington state once power-generation water is counted. Rounding a bottle up to a jug gets you to litres. Against company disclosures — ≈0.32 mL (OpenAI) and 0.26 mL (Google) per average text prompt — 5 L is about four orders of magnitude high, and still ≈3.5× above the largest published figure. Washington Post / UC Riverside · Li et al. 2023 · OpenAI figure
You may have heard people saying “Thank you to ChatGPT” costs millions in electricity.
True, with caveats
Sam Altman said courtesy words cost 'tens of millions' in an informal reply; no workload, period, energy figure or calculation was disclosed. The source establishes the remark, not an auditable electricity total. report on the remark
Ok, Sam Altman said that in a casual remark; wasn’t intended to be taken very seriously. Keep in mind, this only applies if you say “thank you” as a independent query, which results in a full response. If it part of another conversation it is just a few tokens and the cost is minimal
True, with caveats
The characterisation is fair: the figure came from a casual reply with no methodology attached. The mechanism is directionally right — a standalone “thank you” triggers a full generated response, whereas courtesy words folded into a message being sent anyway add only a handful of tokens. One wrinkle: a “thank you” typed as its own turn inside an existing conversation is still a fresh request that reprocesses the context and generates a reply, so it is not the cheap case either. The genuinely cheap case is politeness appended to a prompt already being sent. report on the remark
Let's start with electricity
QQ: Yeah, isn’t AI is using more electricity than entire countries!!
AC uses more electricity than 193 of the world's 195 countries, but nobody is comparing AC to countries. So that's not a great way to compare. Instead here are better numbers:
True, with caveats
IEA estimates cooling at about 10% of global electricity, so the scale behind the comparison is sound. The specific “193 of 195” count would need a named country-level dataset and year, and no cited source establishes it. The deeper point being made — that measuring a global end use against individual countries is a rhetorical trick rather than an analysis — applies to this comparison too. IEA cooling report
IEA gives 485 TWh of data-centre consumption and 28,200 TWh of global consumption in 2025. 485 ÷ 28,200 = 1.72%, which rounds to 1.7%. Note that another sentence in the IEA report rounds the share to 1.5%, so retain the underlying TWh values when precision matters. IEA 2026 electricity demand · IEA 2026 AI report
AI = 32% of data centers, so 0.54% of all electricity,
Checked arithmetic
If AI is assumed to be 32% of data-centre electricity, 1.7% × 32% = 0.544%. The arithmetic is correct. The 32% split is not stated in the accessible IEA narrative and needs the report's underlying chart/data or another source before publication. IEA data-centre outlook
Electricity is 21% of all energy (because remember petrol?)
Checks out
IEA gives electricity as 21% of global total final energy consumption in 2025. “All energy” is loose shorthand for total final consumption, which excludes conversion losses, but the figure itself is right. World Energy Outlook 2025
1.7% × 32% × 21% = 0.114%, so 0.11% of final energy follows arithmetically. It inherits the 32% AI-share assumption, the weakest input in the chain; the electricity share itself is sourced. IEA electricity share
So who’s the biggest consumer of electricity right now?
IEA's global analysis found motor-driven systems accounted for more than 40% of electricity consumption. It is an older (2011) estimate and motor electricity overlaps other end-use labels such as cooling and appliances, so this list is not additive. IEA motor systems report
If anything this undersells households. IEA puts appliances alone — refrigerators, washing machines, computers, phones — at close to a quarter of global electricity, with air conditioning a further 8% on top of that. On those figures residential use sits comfortably above 15%, not below it. The 15% may come from a narrower definition of “appliances”, or from a particular national breakdown rather than the global one. Worth noting too that the list as a whole is not additive: motor systems overlap both cooling and appliances. IEA buildings analysis
IEA's global cooling analysis estimated air conditioners and fans at about 10% of global electricity. A later IEA page gives AC alone at 8%; the scope and year explain much of the difference. IEA Future of Cooling · IEA appliances analysis
IEA's 2026 update estimates about 8% for lighting in buildings and outdoors, excluding industry and agriculture; estimates including industrial lighting imply roughly 9–12% using 2,500–3,500 TWh against 2025 global demand. IEA lighting update
QQ: But AI (and therefore electricity consumption) is growing at a ridiculous rate, isn't it?
Data centers = 3% of all electricity (up from 1.7%); and electricity itself will be 25% of all energy (up from 21%)
True, with caveats
The 3% data-centre share is supported. The companion figure runs slightly high: IEA projects electricity rising from about 21% to about 24% of total final consumption by 2030, not 25%. IEA 2026
AI = 55-60% of data centers (up from 32%), 1.7-1.8% of all electricity (up from 0.54%), 0.45% of all energy (up from 0.11%)
True, with caveats
The arithmetic works only as a scenario assumption: 3% × 55–60% = 1.65–1.8% of electricity; multiplying by a 24% electricity share gives 0.40–0.43% of final energy (or 0.41–0.45% if 25% is assumed). IEA says AI-focused data-centre consumption triples by 2030, but the accessible text does not establish a 55–60% share. IEA 2026
This is and will continue to be a big fraction of their cost
True, with caveats
Electricity is a significant and rising operating cost, and IEA expects it to grow with load. Which cost it is a big fraction of matters: for AI providers, chip capex and depreciation currently dominate total cost, so this is a large share of facility operating cost rather than of the business as a whole. IEA 2026
They're already being forced to charge customers for it
True, with caveats
AI providers charge customers, but the claim that electricity costs are specifically forcing those charges needs provider pricing/cost evidence. As written it confuses ordinary monetisation with cost pass-through.
So, economics will do one of two things (actually both):
This is the standard demand response mechanism: as the user-facing price rises, quantity demanded usually falls, all else equal. It is an economic mechanism, not a forecast of how elastic AI demand will be.
B: because there's so much demand, and so many people are willing to pay high prices, there is a lot of money to be made, so labs will invest:
True, with caveats
IEA documents very large investment and links AI demand to investment in energy equipment, nuclear and startups. 'Labs will invest' and the causal claim about willingness to pay are broader than the source; utilities, hyperscalers, financiers and governments may invest. IEA 2026
IEA projects new renewables, gas and nuclear generation to meet data-centre load and reports technology companies backing generation projects. It does not imply every lab will build its own plant. IEA energy supply for AI
In making everything more efficient (already happening: 10x less energy usage every year)
Checks out
IEA reports that energy per AI task fell by at least an order of magnitude annually in recent years, which is what this says. It describes recent history rather than a guaranteed trend. IEA 2026
Will significantly increase investment in renewables
Checks out
IEA projects renewables to supply about half the growth in data-centre electricity demand. This supports increased renewable deployment, though not necessarily that AI causes a net economy-wide increase after displacement effects. IEA energy supply
IEA expects the first small modular reactors in its data-centre supply mix around 2030 and notes hyperscalers among key corporate backers. This is a projection, not a completed supply source today. IEA energy supply
QQ: Okay, moving on to water… Even if we invent new ways of generating electricity, we can't do that with water! If all water is being used up for cooling data centers, what happens to regular people? We have seen this with almond farms in California…
One problem: 2023 paper, was based on an "estimate" (that 20-50 questions take maybe 500ml). Actual measurements showed it was off by a factor of 40 to 100
Checked arithmetic
The 2023 paper estimated about 500 mL for 20–50 GPT-3-era responses: roughly 10–25 mL each. Dividing by Google's measured 0.26 mL gives 38–96×. That produces the stated 40–100× range, but it is not a clean error estimate: the studies cover different systems, dates, locations and boundaries, and Google's figure excludes offsite electricity-generation water. 2023 estimate · Google methodology
Another problem: Karen Hao's book, misread one key statistic and reported a number 1000x of what the actual number was; She corrected it later, but by then the meme had spread
Checks out
Karen Hao's own correction says a government document labelled cubic-metre values as litres, causing her comparison to be off by exactly 1,000. She also corrected 'consumption' to 'use/withdrawal' elsewhere. Karen Hao correction
that one chatgpt takes a bottle of water per query
Clarification
The episode is describing the meme accurately — this claim genuinely circulated, so what needs correcting is the meme, not the characterisation of it. The “bottle of water per query” framing trended on Threads and TikTok and was catalogued as an AI-incident entry in the form “a bottle of water per email”; harsher variants spread too, including comparisons to ten people showering. The underlying research says something far narrower: Li et al. (2023) modelled roughly 500 mL per 10–50 responses for a GPT-3-era deployment, “depending on when and where” — one modelled scenario, in a hot region, on older cooling technology, not a universal per-query measurement. Company figures for an average text prompt are ≈0.32 mL (OpenAI) and 0.26 mL (Google). Li et al. 2023 · origin of the figure · AIAAIC incident entry · Google · OpenAI
(According to OpenAI and Google reports): One query: 0.24-0.34 Wh electricity, 0.03 g CO2, 0.26-0.32ml water = 5 drops
Checks out
The attribution matters and is right. The detailed set is Google's, for a median Gemini Apps text prompt: 0.24 Wh, 0.03 gCO₂e, 0.26 mL — “five drops” is Google's own analogy. OpenAI separately disclosed ≈0.34 Wh and ≈0.32 mL. Two limits worth knowing: the CO₂ figure is Google's only, with no OpenAI equivalent, and both are company self-disclosures, point-in-time, not independently audited, and excluding electricity-generation and hardware-manufacturing water. Google measurement · OpenAI disclosure
Are we sure THESE numbers are not misinterpreted and/or misquoted?
250 mL ÷ 0.26 mL/prompt = 962 prompts, rounded to 1,000. Valid only for the median Gemini prompt and Google's onsite-water accounting boundary. Google measurement
A peer-reviewed estimate gives 12 litres per California almond. 12,000 mL ÷ 0.26 mL/prompt = 46,154 prompts, rounded to 46,000. This compares a full agricultural water footprint with Google's onsite operational water only, so the accounting boundaries are not like-for-like. California almond study · Google measurement
Here '3L' appears to mean 3 lakh (300,000), not 3 litres. EPA gives an eight-minute shower at more than 16 US gallons (≈61 L): 61,000 ÷ 0.26 ≈ 235,000 prompts; an older 18-gallon assumption gives ≈262,000. 'About 2.5 lakh prompts' is better supported than 3 lakh. EPA shower facts · Google measurement
Here '5L' appears to mean 5 lakh (500,000). A standard water-footprint source gives 140 L for a 125 mL cup of coffee: 140,000 ÷ 0.26 ≈ 538,000 prompts. Again, coffee's supply-chain footprint is being compared with Gemini's onsite operational water. Water Footprint Network · Google measurement
Average American person per day = 1.2 million prompts
Checked arithmetic
EPA reports 82 gallons (≈310 L) per American per day: 310,000 mL ÷ 0.26 mL/prompt ≈ 1.19 million prompts. Holds on Google's onsite-water boundary. EPA statistics · Google measurement
These reproduce once the denominator is identified as global municipal water withdrawal rather than all freshwater use. ICEF puts global direct data-centre water at about 1.5 million m³/day now and about 3.3 million m³/day in 2030 — 0.55 and 1.2 km³/year. FAO AQUASTAT puts municipal withdrawal at about 12% of roughly 4,000 km³/year of global withdrawals, or ≈480 km³/year. That gives ≈0.11% today and ≈0.25% in 2030, close enough to the 0.13% and 0.24% quoted to be the intended calculation. Measured against total withdrawals including agriculture, the shares would be roughly ten times smaller. ICEF water roadmap · FAO AQUASTAT
Water usage: 70% agriculture, 19% industry, 12% city use, 0.1% data centers
True, with caveats
UN-Water/FAO reports approximately 69% agriculture, 19% industry and 12% municipal withdrawals (rounded draft values sum to 101%). Data centres are a subset of industry, not a fourth additive category. The separate 0.1% figure is not established and must specify withdrawal/consumption and scope. UN-Water synthesis report
Entire world's annual data center water usage is 1/20 of just US lawns
Checked arithmetic
ICEF's cited current global direct data-centre use is about 1.5 million m³/day. EPA says US residential outdoor use is nearly 8 billion gallons/day ≈30.3 million m³/day, mainly landscaping. 1.5 ÷ 30.3 ≈ 1/20. This compares global data centres with US outdoor residential use, not strictly lawns alone, and excludes indirect data-centre water. ICEF · EPA outdoors
Entire world's annual data center water usage is 1/3.6 of just US golf
Checked arithmetic
1.5 million m³/day of global direct data-centre water against USGA's ≈1.5 billion gallons/day (5.68 million m³/day) for US golf courses gives ≈1/3.8, close to the 1/3.6 quoted. Both sides are direct use; the years are not perfectly matched. ICEF · USGA
Simple explanation of how it is done
QQ: Yeah, but because of increased demand from AI and datacenters, won’t it make electricity and water expensive for us retail consumers?
How worried should consumers be? Globally, not very: for the next 5-10 years it is still a few percent. But that can still mean that some areas, which are next to a huge data center, can still see significant increases (and some are being seen already in the US). In the short term this could be a problem for some areas. And localities will be justified in fighting datacenters in their backyard, or at least demanding enough compensation.
True, with caveats
IEA supports the key distinction: globally the share remains small, while large concentrated loads can raise local system costs and prices if investment and cost allocation are poorly managed. The blanket 'next 5–10 years' reassurance and water-price claim are not quantified by this source. IEA affordability discussion
But in the long term, this will work itself out; because companies are being charged for it; they will invest, innovate and solve this problem.
Speculative
A forecast rather than a finding, and it could go either way. It holds if prices genuinely reflect full costs — including local water scarcity and the grid upgrades new load requires — and if the efficiency gains of recent years continue. It fails where those conditions don't: subsidised industrial tariffs, long permitting and interconnection queues, water priced below its scarcity value, and costs shifted onto other ratepayers all blunt the signal. IEA describes these constraints and recommends tariff design, disclosure and queue reform rather than expecting the market to resolve them unaided. IEA 2026
Data centers = 1.5% of all electricity; and electricity itself is a fraction of all energy related CO2. So 0.55% of emissions are because of data centers.
True, with caveats
The ≈0.55% conclusion is well supported: IEA estimates data centres used about 1.5% of world electricity in 2024 and caused about 180 Mt of indirect CO₂, ≈0.5% of global fuel-combustion CO₂. The intermediate step is loose — electricity generation is roughly 35–40% of energy-related CO₂ — and multiplying shares assumes data centres draw power at the global average carbon intensity, which they do not. The destination is sounder than the route. IEA climate chapter
AI = 32% of data centers, so 0.45% of all electricity, or 0.18 of all energy related CO2
Checked arithmetic
If AI is 32% of data-centre electricity/emissions, 0.5–0.55% × 32% = 0.16–0.18% of energy-related combustion CO₂. The arithmetic is reasonable; the 32% workload share remains the unsupported input, and emissions share need not equal electricity share if AI facilities use a different grid mix. IEA climate chapter
Data centres = ~2% of electricity-sector CO₂ ≈ 1% of all energy-related CO₂
Checks out
IEA's 2026 update projects data-centre emissions at about 2% of global electricity-sector emissions in 2035. Its 2025 base case puts data-centre indirect emissions at about 1% of global fuel-combustion CO₂. IEA 2026 · IEA 2025 climate chapter
AI ≈ 55–60% of data-centre emissions, so 1.1–1.2% of electricity-sector CO₂, or 0.55–0.60% of all energy-related CO₂.
Checked arithmetic
Assuming AI is 55–60% of data-centre emissions: 2% × 55–60% = 1.1–1.2% of electricity-sector emissions; 1% × 55–60% = 0.55–0.60% of energy-related emissions. The multiplication is correct, but the AI-share assumption is not established by the cited IEA text and equal carbon intensity is assumed. IEA 2026
QQ: I have a gotcha for you!! The world is now all about agents and zero-human orgs. Doesn’t that make it 1000s of times worse?
IEA Figure 2.1 estimates 1.14 Wh for a moderate non-reasoning agentic task against 0.05 Wh for its medium-model text baseline: ≈23×, which rounds to the 25× quoted. The figure models four to six sequential model calls, and IEA warns that agentic systems vary widely. IEA report, pp. 28–29
Extended reasoning: 15 to 150x more energy per prompt
True, with caveats
The upper bound is well sourced: IEA Figure 2.1 gives 7.6 Wh for reasoning against a 0.05 Wh baseline, ≈152×. The lower bound turns on which baseline is chosen rather than on how you count — against OpenAI's 0.34 Wh average query, the same 7.6 Wh is ≈22×. So the range is really two different comparisons rather than one span. IEA report, pp. 28–29
Agentic + extended reasoning: 1000x, per top prompt
Checked arithmetic
IEA Figure 2.1 gives 50 Wh for an indicative agentic-with-reasoning task versus 0.05 Wh baseline: 1,000× energy per task. 'Per top prompt' is the right clarification — it is per top-level task, not per underlying model call. IEA stresses this is an order-of-magnitude illustration, not a universal agent figure. IEA report, pp. 28–29
But, that’s not the full story. Keep in mind, that extended reasoning and agentic also increase time taken per prompt and do a lot more work
Checks out
IEA explains that agents plan, iterate, call tools and execute multi-step tasks, so they accomplish more than a simple completion and can decouple compute from direct human attention. That supports the conceptual caveat, though not any specific productivity ratio. IEA report, pp. 28–31
I think it is reasonable to assume that if you’re using simple/quick chatgpt instant, you will do 1 query per minute; with reasoning you will do one task per 5 minutes, and with an agent, you do one every 15 minutes.
Checked arithmetic
These rates are a reasoned estimate rather than a measurement, and they are what make the per-hour comparisons below reproducible. Worked through: 60 × 0.34 Wh ≈ 20 Wh/hour instant; 12 × 7.6 Wh ≈ 91 Wh/hour with reasoning; 4 × 50 Wh = 200 Wh/hour for agents. One caveat: this mixes OpenAI's per-query figure with IEA's per-task figures, which use different boundaries. On IEA's own 0.05 Wh baseline throughout, the same rates give ≈30× and ≈67× instead of 5× and 10×. IEA report, pp. 28–29 · OpenAI disclosure
With the usage rates above: 12 reasoning tasks/hour × 7.6 Wh = 91.2 Wh against 60 instant queries × 0.34 Wh = 20.4 Wh, giving 4.5× — the 5× quoted. It holds on the OpenAI baseline; on IEA's 0.05 Wh baseline the same rates give ≈30×. IEA task-level estimates · OpenAI disclosure
Same basis: 4 agentic tasks/hour × 50 Wh = 200 Wh against 20.4 Wh/hour instant gives 9.8× — the 10× quoted. Same dependency on the 0.34 Wh baseline; on IEA's 0.05 Wh baseline it would be ≈67×. IEA task-level estimates · OpenAI disclosure
Google equates one median Gemini prompt (0.24 Wh) with under nine seconds of television. Converting an hour of ChatGPT use to 11 TV-minutes requires an explicit query rate and a ChatGPT-vs-Gemini adjustment; neither is stated. Show the assumed prompts/hour. Google measurement
1 hour of paid chatgpt with high reasoning = 1 hour of TV
Checked arithmetic
This reproduces. An hour of reasoning use at the rates given above is ≈91 Wh (12 × 7.6 Wh). A 55-inch LED television draws roughly 77–100 W, commonly quoted at about 0.082 kWh per hour, so an hour of television is ≈82 Wh. The two land within about 10% of each other, which is as precisely as either figure deserves to be read. IEA task-level estimates · TV wattage data
1 hour of agent with high reasoning = 2 hours of TV or 8 minutes of AC
Checked arithmetic
Both halves reproduce. An hour of agentic use at the rates above is 200 Wh (4 × 50 Wh). Against ≈82 Wh/hour for a 55-inch LED TV that is ≈2.4 hours of television. Against a 1.5-ton split air conditioner drawing about 1.5 kW — mid-range for the 1.2–1.8 kW typical in India — 200 Wh buys 200 ÷ 1500 = 0.13 hours, or 8 minutes. IEA's 50 Wh agentic figure is explicitly indicative, so this is an order-of-magnitude comparison. IEA task-level estimates · TV wattage · AC power draw
QQ: This can’t all be a nothing-burger! People are angry and these explanations also feel very dismissive…
It is not a nothing burger; 1% is a big amount; it is probably in the top 5-10 users of energy
True, with caveats
IEA's ≈1.5% of global electricity (2024) does make data centres a large single category, and the hedge is warranted: no cited source ranks them fifth to tenth among energy end uses, and the answer would differ depending on whether you rank by electricity or by total final energy. IEA 2026
Companies have treated people dismissively; not been open with the data, have spent lots of money on funding political campaigns (which is seen as bribes), used NDAs to hide things, and lied about other things, so now people don’t believe the companies claims
True, with caveats
Opacity and NDAs are documented: WIRED reports companies hiding basic project information and unresolved scope questions in public prompt figures. The broader claims about political funding, lying and bribe perceptions need named examples and separate sources; avoid presenting them as one undifferentiated industry-wide fact. WIRED investigation
QIs there enough electricity to go around for everyone?
No; and the companies are very worried about that. Their projected electricity usage is far in excess of current and planned capacities. And
True, with caveats
IEA documents severe grid and supply-chain bottlenecks and estimates about 20% of planned data-centre capacity could face grid-connection delays. 'Far in excess of current and planned capacities' is too absolute globally; projects are constrained precisely because not all requested capacity will be built or connected. IEA energy security
No; but the primary problem there isn’t AI. AI uses some water, not a whole lot. And remember, it uses water for cooling; so that water isn’t disappearing. Some water evaporates, some is discarded or lost; but most is reused.
True, with caveats
Broadly right, with one qualification. Water consumption is withdrawal minus what returns to the same watershed, so it counts cooling-tower blowdown and discharge as well as evaporation. How much is actually reused depends on the cooling design: evaporative towers consume most of what they withdraw, while closed-loop and air-cooled sites consume very little. Offsite power generation can dominate the total footprint either way. ICEF definitions, pp. 3–4
QAre the emissions likely to reach a dangerous level? And are we prepared for it if we do?
Again; emissions from AI are and will continue to be a tiny fraction of emissions from your cars and cows.
True, with caveats
IEA puts all data centres at about 0.5% of global fuel-combustion CO₂ today and ≈1% in its 2035 base case; AI is a subset. That is far below road transport and livestock today, but 'will continue' is scenario-dependent. FAO's older lifecycle estimate puts livestock at 14.5% of anthropogenic GHG, a different accounting basis. IEA climate chapter · FAO livestock assessment
If you use ChatGPT right now, you are not doing a horrible thing. Using ChatGPT is cheaper than using a TV; and even using agents+high reasoning is only like using two TVs simultaneously.
True, with caveats
The personal moral conclusion is a value judgement. Simple text inference is low energy: OpenAI cites ≈0.3 Wh for a typical GPT-4o query and IEA says simple text tasks use less electricity than a TV over the same interval. The 'two TVs' agent comparison still needs the missing usage-rate assumptions noted above. OpenAI Academy · IEA 2026
At a personal level, you don’t need to worry; at a macro level, it is a good idea to keep track of whats going on and see how things are trending. Bad trends = increase in number of data centers, and electricity usage. Good trends = increase in power plants and new/more efficient ways of generating energy.
Mechanism
This is advice about scale and monitoring. IEA likewise emphasises uncertainty, frequent updates, disclosure, generation investment and bottlenecks. More data centres are not inherently a 'bad' metric unless paired with energy, carbon, water stress and social benefit. IEA 2026
One good thing is that the use of electricity and water is actually costing the companies; they’re having to pay for it; and they are having to charge customers for it. That significantly reduces the chances of bad misuse (unlike CO2 emissions)
True, with caveats
The mechanism is real: metered inputs that providers pay for and pass on do create pressure to economise, which is genuinely unlike emissions that cost nothing to release. It is weaker than it sounds where local monopoly power, subsidised tariffs, unpriced externalities and cost-shifting onto other ratepayers intervene — which is why IEA recommends tariff design, disclosure and queue reform rather than relying on price alone. IEA policy discussion