The “AI pollutes more than Google” narrative is built on the wrong unit of measurement. Here’s what the data really says, session by session.
Familiar scene. You search for something on Google from your smartphone. You type. You get ten links. You click on the first one. The page takes three seconds to load. This is not what you are looking for. You come back. You click on the second. Too superficial. Third request, reformulated. Fourth. Fifteenth minute.
Now ask yourself: in all this time, how much power has your phone used?
Nobody asks this question. And that’s exactly the problem.
Since 2023, the dominant narrative has been this: an AI query consumes ten times more energy than a Google search. This assertion has colonized CSR reports, regulatory debates, and management committees. I hear it at every conference. I read it in every AI maturity benchmark that my clients submit to me.
She is comparing the wrong thing.
It compares the server cost of a Google request (what the Mountain View data center pays) against the server cost of an LLM response. Server versus server. And it stops there.
It forgets everything that’s happening on your phone, on the network, and behind the scenes of programmatic advertising while you search.
In March 2026, Charles Duprat (ICOM’Provence) published a working paper which finally asks the right question. Not “how much does a server consume to respond to a request?” But: “how much does the user consume to satisfy a complete information need — from the first search to the final answer?”
This change of unit changes everything.
The comparison we have been making since 2023 is built on the wrong basis
In 2023, one statement colonizes the debates: an LLM query consumes ten times more energy than a Google search. It is printed in CSR committees, regulatory articles, expert speeches.
Problem: this comparison pits server vs. server. It measures what the Google data center consumes to process your request (around 0.30 Wh), and what a GPU cluster consumes to generate an LLM response (around 0.24 to 0.34 Wh according to Google, Epoch AI and OpenAI — three converging independent sources).
This is where it all ends. And that’s where it all goes wrong.
Because a Google search doesn’t give you information. It gives you a map to information hosted elsewhere. The energy to navigate this map — downloading pages, rendering JavaScript, undergoing ad auctions — is carried by your device, your mobile network, and a largely invisible ad-tech infrastructure.
None of this shows up on the data center meter.
The actual numbers, session vs. session
Let’s take a concrete task on mobile: compare two technical solutions on three sources, 5G, non-standalone network (which almost all of French “5G” still is in 2026 – NSA, that is to say 4G core).
Web search session, figures by component:
- Server request processing: 0.30 Wh
- Network: 3 pages × 2.56 MB (median HTTP Archive 2025) × 0.14 kWh/GB = 1.08 Wh
- Page rendering (CPU/GPU device): 0.60 Wh
- Advertising load (30% of rendering, Khan et al. 2024 measurement): 0.18 Wh
- Screen time (6 minutes × 2.5 W): 0.25 Wh
- Total: 2.41 Wh
Equivalent LLM session:
- Inference (standard model, not reasoning): 0.30 to 0.40 Wh
- Network: text payload of 5 KB, negligible
- Screen time (2.5 minutes): 0.10 Wh
- Total: 0.40 to 0.50 Wh
Central ratio: 5.4 times in favor of the LLM.
Duprat validates this figure with a Monte Carlo analysis on 10,000 draws and 9 free parameters. Result: no combination of values lowers the search below the LLM. The observed floor — the worst case for LLM — is 1.6 times.
Three mechanisms that explain the inversion
First mechanism: the mobile network is the real culprit.
The median mobile page weighs 2.56 MB in 2025. On 4G, it costs 0.44 Wh in transmission energy. For a single page. Before having rendered a pixel.
An LLM response is a text payload of 2 to 10 KB. The transmission ratio is of the order of 500:1. The network is not a marginal cost — it is the dominant component of the footprint of a mobile search session.
Second mechanism: programmatic advertising is an invisible energy tax.
When you load an ad-supported page, an auction opens in parallel. Dozens of DSPs receive the request. Almost all lose — and consume CPU cycles for nothing.
Khan et al. (2024) measured it directly: integrated ad-blockers reduce device consumption by 15 to 44% compared to normal browsing. That is to say, between a sixth and almost half of the energy your smartphone burns while browsing fuels the advertising ecosystem, not the content you read.
An LLM completely bypasses this infrastructure.
Third mechanism: completion speed reduces screen time.
The CHI 2025 study by Spatharioti et al. — randomized design between two groups — measure that LLM users complete summary tasks faster, with fewer queries. Less screen time, less watts.
So-called “pogo-sticking” behavior — clicking, finding the page disappointing, returning to the SERP, starting again — creates a penalty that static models never capture. Each return to the SERP on mobile costs an additional 0.30 to 0.60 Wh. The LLM structurally eliminates this pattern by providing a complete synthesis in the first exchange.
Blind spots that need to be named
Three real limits, not nuances of comfort.
On fixed Wi-Fi, the advantage collapses. On a fixed network (0.006 kWh/GB), the transmission cost drops by 95%. The LLM advantage drops to 1.5–2.5 times on complex tasks, reaching parity on simple queries. Inversion is a moving phenomenon. It does not generalize mechanically to the desktop.
On reasoning models, the logic is reversed, sometimes violently. Claude Opus in thinking mode, GPT-o3, Gemini Deep Think: these models generate extended chains of reasoning. Jin et al. (2025) document an average expansion of 4.4x output tokens in production, with extreme cases at 113x. The crossover threshold with mobile search — the point where LLM becomes more energy intensive — is at a factor of 4 to 8 times. We are already in the risk zone for common reasoning models. These are not slightly more expensive models: they are a categorically different consumption regime.
The Jevons paradox is not swept away by unitary efficiency. ChatGPT exceeds 2 billion daily requests by the end of 2025. If this demand is new rather than replacing web search, total consumption increases independently of the unit ratio. Efficiency per session says nothing about aggregate effectiveness. These are two separate questions and both deserve a serious answer.
What this actually changes for your organization
If you are leading an AI deployment strategy or a digital CSR policy, there are three direct implications.
First: the choice of model is an energy decision, not just a performance decision. Using a reasoning model for standard synthesis tasks — which the majority of companies do by default because “it’s the best model” — multiplies the footprint by a factor that no one quantifies in carbon footprints. Intelligent routing by task complexity is not a technical luxury. This is an imperative for CSR consistency.
Second: your mobile teams do web research where a standard LLM would be 5 times less energy intensive. For synthesis, monitoring, and multi-source comparison tasks — substitution is measurable, immediate, and requires no additional investment.
Third: the web advertising infrastructure is an external cost that your employees bear without seeing it. Auditing the digital footprint of your organization without counting the 15 to 44% of device energy sucked up by programmatic advertising is a significant accounting blind spot.
What I have observed in the field for three years
Of the 200 AI projects that I deployed in B2B companies between 2022 and 2025, a pattern systematically comes up during digital maturity audits: the teams which switched their research and synthesis tasks to a standard LLM did not do so for ecological reasons. They did it because it’s faster.
Collateral result: they have mechanically reduced their mobile digital footprint. Without realizing it. Without measuring it.
Duprat now provides the framework to encrypt it. And the figure is brutal.
What it changes, concretely
Let’s go back to the beginning scene. You are looking for complex information on mobile.
Scenario A — Google: four queries, seven pages loaded, fifteen minutes of screen time, ad auctions in the background. Estimated consumption: 2.41 Wh.
Scenario B — Standard LLM: one query, response summarized in thirty seconds. Estimated consumption: 0.40 Wh.
You have not made an ecological gesture. You just asked your question in the right place.
The “AI consumes ten times more than Google” narrative is not only inaccurate. It protects an infrastructure — the mobile advertising web — whose real energy cost has never been properly accounted for because no one had any interest in doing so.
A modern web page is not a document. It’s a software package that performs hundreds of operations for advertisers you’ll never see. You pay the cost in time. Your battery pays the cost in watts.
The LLM wins this comparison because its opponents are extraordinarily ineffective. Not because he is virtuous.
It’s a difference that matters — for your purchasing decisions, for your CSR policies, and for the next time someone explains to you in a meeting that “AI is bad for the planet.”
Main source: Duprat, C. (2026). The Thermodynamic Efficiency Inversion: A Comparative Energy Lifecycle Assessment of Generative AI Inference versus Ad-Supported Web Search Sessions. Working paper, ICOM’Provence. Revised March 25, 2026. Non-peer-reviewed working paper — the underlying empirical data (Google arXiv:2508.15734, Khan et al. 2024, Spatharioti et al. CHI 2025) are published and independently verifiable.
