Every few weeks a headline tells us that a single ChatGPT prompt “drinks a bottle of water” or burns ten times the energy of a Google search. Policy teams hear this and wonder whether they should feel guilty about drafting a brief with Claude. So I did what an engineer does with a vague worry: I put numbers on it.
I took the three tasks that consume most of a policy analyst’s week and costed each one three ways. The tasks are drafting a presentation with an accompanying report, running a literature or case-study review, and generating ideas. The modes are Google web search, offline databases (JSTOR, Scopus, World Bank and IMF portals, plus your own downloaded archive), and an LLM chat assistant such as Claude or ChatGPT.
The comparison
Figures are for one analyst producing one deliverable: a 10-slide deck with a 5-page brief, a scoping review of about 30 sources, and a shortlist of policy options with rationale. Electricity covers the analyst’s workstation (laptop plus monitor, about 100 W) and the server-side cost of every search or prompt. Environmental resources are operational CO₂e and data-center cooling water. The storage and infrastructure column captures what the headline “per query” numbers leave out: the energy to hold the working files for a month, the amortized share of model training, and the manufacturing (embodied) carbon of the laptop and the servers, spread over their useful lives. Manpower is analyst hours, with the verification time that AI output requires shown separately.
| Task | Mode | Avg. time | Electricity (total, of which servers) | CO₂e / DC water (operational) | Storage & infrastructure (amortized) | Manpower |
|---|---|---|---|---|---|---|
| Presentation + report draft | Google search | 480 min | 824 Wh (24 Wh) | 321 g / 26 mL | ~5 Wh storage · 330 g embodied CO₂e | 8 person-hours, 1 analyst |
| Offline databases | 600 min | 1,006 Wh (6 Wh) | 392 g / 7 mL | ~23 Wh storage (5 GB local archive) · 410 g embodied | 10 person-hours, 1 analyst | |
| LLM chat | 240 min | 480 Wh (80 Wh) | 187 g / 88 mL | ~5 Wh storage · +80 Wh training share · 215 g embodied | 4 person-hours incl. ~1 h fact-checking | |
| Literature / case-study review | Google search | 1,200 min | 2,060 Wh (60 Wh) | 803 g / 66 mL | ~14 Wh storage · 820 g embodied | 20 person-hours, 1 analyst |
| Offline databases | 1,440 min | 2,415 Wh (15 Wh) | 942 g / 16 mL | ~46 Wh storage (10 GB archive) · 980 g embodied | 24 person-hours, 1 analyst (+ librarian support) | |
| LLM chat | 600 min | 1,240 Wh (240 Wh) | 484 g / 264 mL | ~14 Wh storage · +240 Wh training share · 560 g embodied | 10 person-hours incl. ~3 h citation verification | |
| Idea generation | Google search | 180 min | 312 Wh (12 Wh) | 122 g / 13 mL | ~1 Wh storage · 125 g embodied | 3 person-hours, 1 analyst |
| Offline databases | 240 min | 403 Wh (3 Wh) | 157 g / 3 mL | ~5 Wh storage · 165 g embodied | 4 person-hours, 1 analyst | |
| LLM chat | 60 min | 115 Wh (15 Wh) | 45 g / 16 mL | ~1 Wh storage · +15 Wh training share · 50 g embodied | 1 person-hour incl. ~15 min sanity check |
Five things the numbers say
The server side barely registers. A Google search costs about 0.3 Wh, a median Gemini prompt 0.24 Wh, a ChatGPT query around 0.34 Wh, and third-party benchmarks put a Claude Sonnet response near 0.8 Wh. Fire 120 long-context prompts at a literature review and you have spent a quarter of a kilowatt-hour. Your laptop and monitor burn that in two and a half hours of reading PDFs.
Storage is a rounding error, and so is the “data centre” you imagine behind each mode. Holding a 10 GB research archive for a month costs about 46 Wh, less than a single hour at the screen. Google’s index, a publisher’s database servers and an AI serving cluster are all shared across hundreds of millions of users, so the slice attributable to one deck is small. Where infrastructure does show up is in manufacturing: the embodied carbon of the analyst’s own hardware, amortized at roughly 40 g CO₂e per hour of use, is about equal to the electricity it draws. Add it in and every column’s carbon roughly doubles, but the ranking between modes does not change, because the embodied number is driven by the same variable as everything else: hours.
Training does not rescue the anti-AI case. Epoch AI’s analysis suggests that over a model’s deployed lifetime, training energy is roughly comparable to total inference energy, so I doubled the server-side figure for the LLM column to account for it. That takes the literature review’s AI-side cost from 240 Wh to 480 Wh. It is still a fifth of what the workstation consumes over the same task.
Water is the one column where AI loses. LLM prompts consume three to ten times more data-center cooling water than the equivalent searches, and offline databases win outright because the heavy lifting happens on your own machine. The absolute quantities are small, a quarter of a litre for the most AI-intensive task, but if your organisation or your cloud region is water-stressed, this is the number to watch.
Manpower does not simply shrink; it changes shape. In the AI column roughly a quarter of the remaining hours are verification: confirming that a cited paper exists, that a case study happened in the country the model claims, that a statistic was not invented. Skip that and the saving is an illusion that resurfaces later as a correction or a retraction.
Caveats worth stating
These are point estimates built on published averages and reasonable task assumptions, not measurements of your team. The Google search figure dates from 2009 and has probably fallen. The Gemini and ChatGPT figures come from the companies themselves and are unaudited. Anthropic has published no per-query numbers at all, so Claude rests on academic benchmarks. Reasoning models change the picture sharply: a long o3-class response can cost 20 to 40 Wh, a hundred times a plain chat prompt, and a workflow built on them would push the AI column’s server share from a few percent to a real fraction. Storage figures use a mid-range 55 kWh per terabyte-year and ignore replication. And the biggest unknown is the rebound effect: if faster research means three times as many decks, the aggregate footprint grows even as the per-deck footprint falls.
The bottom line
For the three tasks that define policy research, the choice between Google, a database and an AI assistant is not primarily an environmental decision. Per deliverable, the AI workflow is the lowest-energy and lowest-carbon option, even after charging it for training and embodied hardware, because it shortens the hours. It is modestly more water-hungry, and it is only cheaper in manpower if verification is budgeted honestly. The environmental variable that actually matters is not which tool you open but how long the screen stays on, and whether the hours you get back go into better analysis or simply more of it.
Method note: workstation 100 W; US grid average 0.39 kg CO₂e/kWh; data-center water 1.1 L/kWh (Google’s disclosed ratio); 0.3 Wh per web search, 0.1 Wh per database query, 0.5 Wh per short prompt and 2 Wh per long-context prompt. Storage at 55 kWh/TB-year for one month of the working corpus (1–10 GB). Training amortized at 100% of inference energy (Epoch AI). Embodied carbon: 40 g CO₂e per workstation-hour (a ~300 kg laptop-plus-monitor footprint over four years of office use) plus 67% of server-side operational CO₂e (embodied ≈ 40% of data-center lifecycle emissions). Task durations draw on presentation-time benchmarks (4–6 h per 10-slide deck), librarian time-tracking for systematic reviews (median 22 h on the search side), and the productivity studies cited below (MIT: 40% faster; Harvard–BCG: 25% faster).
Sources: Google Cloud — Measuring the environmental impact of AI inference · MIT Technology Review on Google’s Gemini disclosure · DCD — Altman: 0.34 Wh per ChatGPT query · Epoch AI — How much energy does ChatGPT use? · Hannah Ritchie — AI footprint, August 2025 update · Devera — ChatGPT vs Claude vs Gemini energy and water · Windows Forum — Anthropic emissions reporting gap · Search Engine Land — carbon footprint of a Google search (2009) · IEEE Spectrum — Data center sustainability metrics and embodied carbon · EcoFlow — energy cost of 1 TB cloud storage · MIT News — ChatGPT boosts productivity for writing tasks · Noy & Zhang, Science 2023 · Dell’Acqua et al., Navigating the Jagged Technological Frontier · Bullers et al., librarian time on systematic review tasks · TextDeck — average time to create a presentation · Jackery — computer power consumption
sadhanamanthapuri.com 
Leave a comment