The narrative “AI consumes more energy than Google” is based 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. It’s not what you’re looking for. You go back. You click on the second one. Too superficial. Third attempt, reformulated. Fourth. Fifteenth minute.
Now, ask yourself this question: during all this time, how much energy has your phone consumed?
No one is asking themselves that question. And that’s exactly the problem.
Since 2023, the dominant narrative says this: an AI query consumes ten times more energy than a Google search. This assertion has taken over CSR reports, regulatory debates, board meetings. I hear it at every conference. I read it in every AI maturity benchmark that my clients submit to me.
It compares the wrong thing.
It puts the server cost of a Google query (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 happens 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 that finally asks the right question. Not “how much does a server consume to respond to a query?” But: “how much does the user consume to satisfy a complete information need – from the first search to the final response?”
This change in unit changes everything.
The comparison made since 2023 is based on the wrong premise
In 2023, a claim dominates the debates: an LLM query consumes ten times more energy than a Google search. It permeates CSR committees, regulatory articles, expert speeches.
The problem: this comparison pits server vs server. It measures what the Google data center consumes to process your query (about 0.30 Wh), and what a GPU cluster consumes to generate an LLM response (about 0.24 to 0.34 Wh according to Google, Epoch AI, and OpenAI – three convergent independent sources).
That’s where everything stops. And that’s where everything 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 – to download the pages, render the JavaScript, endure the advertising auctions – is borne by your device, your mobile network, and a largely invisible ad-tech infrastructure.
None of this appears on the data center’s meter.
Real numbers, session vs session
Let’s take a concrete task on mobile: comparing two technical solutions on three sources, 5G, non-standalone network (which the majority of French “5G” still is in 2026 – NSA, meaning core 4G).
Web search session, figures by component: – Server query 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 – Ad load (30% of rendering, Khan et al. 2024): 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: 5 KB text payload, negligible – Screen time (2.5 minutes): 0.10 Wh – Total: 0.40 to 0.50 Wh
Central ratio: 5.4 times in favor of LLM.
Duprat validates this figure with a Monte Carlo analysis of 10,000 simulations and 9 free parameters. Result: no combination of values brings the search below the LLM. The observed floor – the most unfavorable case for LLM – is 1.6 times.
Three mechanisms that explain the inversion
First mechanism: the mobile network is the real culprit.
The average mobile page weighed 2.56 MB in 2025. On 4G, it costs 0.44 Wh in transmission energy. For a single page. Before rendering a pixel.
An LLM response is a text payload of 2 to 10 KB. The transmission ratio is around 500:1. The network isn’t a marginal cost – it’s the dominant component of a mobile search session’s footprint.
Second mechanism: programmatic advertising is an invisible energy tax.
When loading an ad-supported page, an auction opens in parallel. Dozens of DSPs receive the demand. The vast majority lose – and consume CPU cycles for nothing.
Khan et al. (2024) directly measured it: integrated ad-blockers reduce device consumption by 15 to 44% compared to normal browsing. This means that between one-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 – shows that LLM users complete synthesis tasks faster, with fewer queries. Less screen time, less watts.
The behavior known as “pogo-sticking” – clicking, finding the disappointing page, returning to the SERP, and restarting – 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 delivering a complete synthesis on 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 falls to 1.5 to 2.5 times on complex tasks, reaches parity on simple queries. The inversion is a mobile phenomenon. It does not automatically generalize to desktop.
On reasoning models, the logic reverses, sometimes sharply. Models like Claude Opus in thinking mode, GPT-o3, Gemini Deep Think generate extensive chains of reasoning. Jin et al. (2025) document an average 4.4 times expansion of the output tokens in production, with extreme cases up to 113 times. The crossing point with mobile search – where LLM becomes more energy-intensive – is at a factor of 4 to 8 times. We are already in the risk zone for current reasoning models. These are not slightly more expensive models: it is a categorically different consumption regime.
Jevons paradox is not swept away by unit efficiency. ChatGPT surpasses 2 billion daily queries by the end of 2025. If this demand is new rather than substituted from web search, the total consumption increases regardless of the unit ratio. Efficiency per session says nothing about aggregate efficiency. These are two distinct questions and both deserve a serious answer.
What this means concretely for your organization
If you are driving an AI deployment strategy or a digital CSR policy, three direct implications.
Firstly: the choice of model is an energy decision, not just a performance decision. Using a reasoning model for standard synthesis tasks – what most companies default to because “it’s the best model” – multiplies the footprint by a factor that no one quantifies in carbon assessments. Smart routing based on task complexity is not a technical luxury. It is an RSE coherence imperative.
Secondly: your mobile teams are making web searches where a standard LLM would be 5 times less energy-intensive. For synthesis, monitoring, multi-source comparison tasks – the substitution is measurable, immediate, and requires no additional investment.
Thirdly: the web’s advertising infrastructure is an energy externality that your employees bear without seeing it. Auditing your organization’s digital footprint without counting the 15 to 44% of device energy sucked by programmatic ads is a significant blind spot.
What I’ve observed on the ground for three years
On the 200 AI projects I’ve deployed in B2B companies between 2022 and 2025, a pattern consistently emerges during digital maturity audits: teams that have shifted their search and synthesis tasks to a standard LLM did not do it for ecological reasons. They did it because it’s faster.
Collateral result: they have unwittingly reduced their mobile digital footprint. Without realizing it. Without measuring it.
Duprat now provides the framework to quantify it. And the figure is stark.
What this changes, concretely
Let’s go back to the scene at the beginning. You’re looking for complex information on mobile.
Scenario A – Google: four queries, seven loaded pages, fifteen minutes of screen time, background ads. Estimated consumption: 2.41 Wh.
Scenario B – Standard LLM: one query, synthesized response in thirty seconds. Estimated consumption: 0.40 Wh.
You haven’t made an ecological gesture. You’ve just asked your question in the right place.
The narrative “AI consumes ten times more than Google” is not only inaccurate. It protects an infrastructure – mobile advertising web – whose real energy cost has never been correctly accounted for because no one had the intention to do so.
A modern web page is not a document. It’s a software package that executes 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 inefficient. Not because it is virtuous.
It’s a difference that matters – for your purchasing decisions, for your CSR policies, and for the next time someone explains 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 on March 25, 2026. Non-peer-reviewed working paper – underlying empirical data (Google arXiv:2508.15734, Khan et al. 2024, Spatharioti et al. CHI 2025) are published and independently verifiable.
