Is AI Search Really "Free"? What a Query Costs in Electricity, and What That Means for Indonesia
The question behind "free"; Typing a question into Google costs you nothing. Typing it into a chatbot costs you nothing either, if you use the free tier. Yet somebody pays for the electricity, the chips, the cooling and the idle capacity. The interesting que..
The question behind "free"; Typing a question into Google costs you nothing.
Typing it into a chatbot costs you nothing either, if you use the free tier.
Yet somebody pays for the electricity, the chips, the cooling and the idle capacity. The interesting question is not whether AI search is free. It is who pays, how much, and how sure are we of the number?
We tried to answer it for one country which is Indonesia, ages 15 to 49. Every input below is tagged with its source and how reliable it is.
Where the world has not published a number, we say so, and we show what happens to the conclusion when that number moves.
The short version that the headline "AI uses 10x more energy than Google" is no longer safe. Neither is its opposite. The answer depends on numbers that the largest companies have either not published or last published seventeen years ago.
What "one query costs" - the published measurements
| Source | Figure | What it actually is |
|---|---|---|
| Google, technical paper (Aug 2025) | 0.24 Wh per median Gemini Apps text prompt | First-party measurement in production, May 2025. Includes idle machines and data-centre overhead. A narrower accelerator-only view gives 0.10 Wh. Not peer-reviewed at release. |
| Sam Altman (2025), via CBS News | 0.34 Wh per average ChatGPT query | CEO statement. Measurement boundary not disclosed. |
| Epoch AI (2025) | ~0.3 Wh typical; ~2.5 Wh for a 10,000-token input | Independent estimate for GPT-4o. |
| Nature Comment supplement (2025) | 0.4 Wh typical; 3.5 Wh long query | Compiles several studies. |
| Research paper updated June 2026 | 0.31 Wh median; 3.91 Wh with 15× more tokens | Reasoning-heavy use multiplies cost about 13×. |
| Google, classic search (2009) | 0.3 Wh per search | The last official figure. It has not been updated in over 15 years. |
| Unofficial recent estimates | 0.03 to 0.04 Wh per search | Secondary sources, no disclosed method. |
Table 1
Published energy consumption measurements and estimates per query: AI inference vs. conventional search benchmarks (2009–2026).
Two things stand out.
- First, a typical AI text prompt and a 2009-era search sit at almost the same figure, 0.24 to 0.34 Wh versus 0.3 Wh.
- Second, Google reports that its median prompt energy fell 33× in twelve months.
A number that can shrink that fast should be treated as a dated reading, not a constant.
The old "10×" claim came from a 2023 estimate of about 2.9 Wh per query for older models. It is now roughly decade of efficiency behind.
Global volume "who is asking what"
- ChatGPT: 900 million weekly active users (OpenAI, 27 Feb 2026) and about 2.5 billion prompts per day.
- Google Search: more than 5 trillion searches per year, about 13.7 billion per day (Google, 2025 disclosure).
- Internet users worldwide: 6.0 billion in 2025 (ITU).
That gives an anchor of ≈2.3 searches per internet user per day (13.7 B ÷ 6.0 B), and ≈2.8 ChatGPT prompts per weekly-active user per day (2.5 B ÷ 0.9 B).
Google does not publish how many Gemini prompts it serves, so its 0.24 Wh cannot be multiplied into a national or global total by anyone outside the company.
The Indonesian case: building the population
A query’s energy cost is meaningless until it is grounded in real-world demography.
Rather than assuming every connected citizen is an active prompt engineer, we need to map the digital baseline layer by layer. What follows is the empirical filter - turning 284 million people into a verifiable estimate of daily Indonesian search and AI consumption.
Step 1
Who is 15 to 49? BPS mid-2025 projections (via GoodStats)
- 15–19: 22.09 M
- 20–24: 22.15 M
- 25–29: 22.51 M
- 30–34: 22.20 M
- 35–39: 21.73 M
- 40–44: 20.70 M
- 45–49: 19.65 M
Total 15-49 = 151.0 million people, or 53.1% of Indonesia's 284.44 million. Adding age 50 itself would add roughly 3.5 million, but we stop at 49 to keep age bands clean.
Step 2
Who is online? APJII reports internet penetration of 80.66% in 2025 (229 M people) and 81.72% in 2026.
Penetration by generation in 2025 was about 89.1% (Millennials), 87.8% (Gen Z) and 79.5% (Gen X). Mapping those bands to our age groups gives ≈131.9 million online people aged 15–49, or 87.3% of the group.
Step 3
Who uses Google? StatCounter puts Google at 93.72% of Indonesian search-engine usage (May 2026).
Applying that share gives ≈123.6 million Google-type searchers. This is an approximation: StatCounter measures page-view share, not people.
Step 4
Who uses AI? APJII's 2026 survey reports that 27.3% of respondents have accessed AI. That figure means "has ever accessed", not "uses it daily".
The model
Energy per day = people × queries per person × Wh per query. Cost = kWh × tariff.
Tariff: PLN large-industrial I-4 tariff, Rp 996.74/kWh, unchanged for Q3 2026. The I-3 tariff (Rp 1,114.74) would add about 12%. Where the servers physically sit, and what they pay, is not disclosed. We use PLN rates as a common yardstick, not as a claim about anyone's bill.
Conventional search (fixed)
123.6 M searchers × 2.3 searches/day = 284.3 million searches/day, at 0.3 Wh (upper bound) and 0.03 Wh (low sensitivity).
AI search (three scenarios)
Only the 27.3% figure and the Wh values come from published sources. The rest are our stated assumptions:
| Low | Base | High | |
|---|---|---|---|
| Share of online 15–49 who are AI-active on a given day | 5% | 13.65% (= 27.3% × 50%) | 27.3% (every "ever" user is daily) |
| AI-active people | 6.6 M | 18.0 M | 36.0 M |
| Prompts per person per day | 3 | 6 | 12 |
| Wh per prompt | 0.24 | 0.34 | 0.656 |
Table 2
Modeled daily active user shares, prompt frequencies, and blended per-query energy intensities across three adoption scenarios.
The High-case 0.656 Wh is a blend: 90% typical prompts at 0.34 Wh plus 10% long or reasoning prompts at 3.5 Wh.
Results
| Scenario | Prompts or searches/day | kWh/day | Rp/day | GWh/month | Rp/month | GWh/year | Rp/year |
|---|---|---|---|---|---|---|---|
| Search @ 0.3 Wh | 284.3 M | 85,284 | Rp 85.0 M | 2.56 | Rp 2.55 B | 31.1 | Rp 31.0 B |
| Search @ 0.03 Wh | 284.3 M | 8,528 | Rp 8.5 M | 0.26 | Rp 0.26 B | 3.1 | Rp 3.1 B |
| AI Low | 19.8 M | 4,748 | Rp 4.7 M | 0.14 | Rp 0.14 B | 1.7 | Rp 1.7 B |
| AI Base | 108.0 M | 36,724 | Rp 36.6 M | 1.10 | Rp 1.10 B | 13.4 | Rp 13.4 B |
| AI High | 432.1 M | 283,422 | Rp 282.5 M | 8.50 | Rp 8.48 B | 103.5 | Rp 103.1 B |
Table 3
Projected daily, monthly, and annualized electricity consumption and cost across conventional search baselines and AI adoption scenarios in Indonesia.
Months are 30 days, years 365. All figures verified by recalculation.
At roughly Rp 16,400 per US dollar, the Base AI case costs about US$0.81 million per year in electricity. Per AI-active person, that is about Rp 742 per year, or about Rp 2 per day.
Three uncomfortable findings
- The ranking flips depending on a number nobody has published. If classic search still costs 0.3 Wh, the Base AI scenario is only 43% of search energy. If it costs 0.03 Wh, Base AI is 4.3× larger, and the High case is 33×. One unpublished figure decides the headline. Anyone who states a confident "AI is X times more than search" is choosing their assumption, not reporting a fact.
- The median hides the tail. A median prompt is cheap. A long input or a reasoning run costs 10 to 13 times more. If Indonesia's AI use drifts from short questions toward research modes, energy per user can jump without any increase in user numbers. OpenAI itself says Indonesia ranks in the top 5% for reasoning-token usage.
- "Free" is a boundary decision. Rp 742 per user per year is the electricity for a typical AI-active person. A paid ChatGPT Plus plan costs about Rp 328,000 per month. Electricity is one line of cost, and our model deliberately excludes chips, model training, staff and margins. So "free" is not a claim about electricity. It is a decision about who absorbs everything else: the company, the advertisers, the paying minority, or investors.
A fourth, quieter point: one AI prompt often replaces several searches. Comparing per-event energy is not the same as comparing per-task energy. Nobody has published data that settles this.
Verification ledger
| Input | Value | Source | Reliability | Action before publishing |
|---|---|---|---|---|
| Population 15–49 | 151.0 M | BPS mid-2025 projection, via GoodStats | Secondary. BPS table did not load in our tool. Age bands sum to 284.67 M against 284.44 M total (rounding). | Confirm on the BPS table directly. |
| Internet penetration | 80.66% (2025); 81.72% (2026) | APJII | 2025 confirmed by two outlets. 2026 via a secondary blog. | Confirm 2026 in the APJII report. |
| Cohort penetration | 89.1 / 87.8 / 79.5% | APJII 2025 via Teknologi.id | Secondary. Generation-to-age mapping is approximate. | Confirm in the APJII report. |
| "Accessed AI" | 27.3% | APJII 2026 via a secondary blog | Unverified. The blog's generation figures (29.4%, 16.7%, 7.5%) are internally inconsistent with a 27.3% total. | Do not publish until checked against the primary report. |
| Google share, Indonesia | 93.72% | StatCounter, May 2026, via secondary summary | Secondary. | Read the StatCounter page directly. |
| Searches per person per day | 2.3 | Our derivation: Google 5 T/yr ÷ ITU 6.0 B users | Global average, not Indonesia-specific. | Flag as assumption. |
| Google search Wh | 0.3 | Google, 2009 | Official but 17 years old. | Flag as upper bound. |
| Gemini Wh | 0.24 | Google paper, May 2025 | Primary, first-party, not peer-reviewed at release. | Cite the paper. |
| ChatGPT Wh | 0.34 | Altman, via CBS | CEO statement, boundary undisclosed. | Cite as claim, not measurement. |
| PLN tariff | Rp 996.74 (I-4) | Secondary industry blog, consistent across two sources | Q3 2026 unchanged per RRI. The same blog lists two different I-3 values. Q4 adjustment is due in October. | Confirm on pln.co.id. |
| Exchange rate | ≈Rp 16,400/USD | Industry.co.id | Secondary, implied by price conversion. | Use Bank Indonesia rate. |
| OpenAI "1 in 5 Indonesians" | not used | Media report | Contradicts its own "over 50%" claim. | Excluded. |
Table 4
Data provenance, empirical reliability assessment, and pre-publication verification status for demographic, technical, and economic model inputs.
Not modelled: training energy, hardware manufacturing, water, network and device energy, image and video generation, Baidu/Amazon/Yahoo/Bing (no official per-query energy or usable 2025–26 volumes found), and Google AI Overviews layered onto classic search.
What would change our conclusion
- Google publishes a current per-search figure. That resolves the largest uncertainty in the table.
- A daily-active AI share for Indonesia replaces our 50% assumption.
- A shift to reasoning and agent use, which multiplies energy per person.
- Hardware and software efficiency. Google's own 33× drop suggests today's Base case may be an overestimate a year from now.
The question to leave with
If an AI answer costs about as much electricity as a 2009 search, but replaces three searches and a page of reading, is AI search the wasteful one, or the efficient one?
And if the true cost of a query is unpublished by design, who is positioned to find out?
Reference
- Google, Measuring the environmental impact of delivering AI at Google Scale
https://arxiv.org/abs/2508.15734 - CBS News on Gemini and ChatGPT per-query figures
https://www.cbsnews.com/news/ai-environment-impact-study-energy-usage-google-gemini-prompt/ - Epoch AI, How much energy does ChatGPT use?
https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use - Nature Comment supplement (Luers, 2025)
https://media.nature.com/original/magazine-assets/d41586-025-02641-4/51368154 - Google 2009 figure and its age
https://www.zmescience.com/science/news-science/how-much-energy-chatgpt-query-uses/ - ANTARA on 900 M weekly users
https://www.antaranews.com/berita/5444662/openai-catat-jumlah-pengguna-aktif-mingguan-chatgpt-tembus-900-juta - ITU Facts and Figures 2025
https://www.itu.int/itu-d/reports/statistics/2025/10/15/ff25-internet-use/ - BPS age table (via GoodStats)
https://goodstats.id/article/indonesia-masih-didominasi-penduduk-muda-hampir-separuh-berusia-di-bawah-30-tahun-jUs5H - APJII 2025 (Bloomberg Technoz)
https://www.bloombergtechnoz.com/detail-news/79778/survei-apjii-pengguna-internet-indonesia-2025-ada-229-juta-orang - APJII 2026 summary (secondary)
https://www.jununx.com/2026/08/penetrasi-internet-indonesia.html - StatCounter Indonesia
https://gs.statcounter.com/search-engine-market-share/all/indonesia - PLN tariff adjustment
https://www.pln.co.id/pelanggan-id/tarif-tenaga-listrik-id/penyesuaian-tarif-id - Gemini prompt volume not published
https://crbonfree.com/blog/how-much-energy-does-gemini-use