The Thermodynamic
Efficiency Inversion

Introduction and motivation

In 2023, a widely circulated comparison claimed that generating a single response from a large language model consumed ten times more energy than a Google search query. The claim was technically narrow — it compared server-side GPU computation for an unoptimised, low-utilisation research deployment against a decade-mature search stack — but it lodged in public consciousness, shaped ESG discourse, and influenced early regulatory thinking on both sides of the Atlantic.

This paper makes a different comparison. Rather than asking "how much energy does a server consume to answer one query?", we ask: "how much energy does a user consume to satisfy one complex information need?" This reframing — from server-side computation to full-stack session — changes the answer substantially.

The energy cost of navigating that map — downloading pages, rendering JavaScript, processing advertisement auctions, and spending time reading — is borne by the user's device, the telecommunications network, and a largely invisible ad-tech infrastructure. None of these costs appear on the data centre's meter.

The past eighteen months have also transformed the empirical landscape. Google published a peer-reviewed technical paper Elsworth et al. (2025) documenting that the median Gemini text prompt consumed 0.24 Wh in May 2025 — a 33-fold reduction from the same measure twelve months prior. OpenAI's CEO disclosed a comparable 0.34 Wh for ChatGPT. Meanwhile, the HTTP Archive 2025 Web Almanac reports a median mobile page weight of 2.16 MB across page types, with mobile home pages at 2.56 MB and mobile inner pages at 1.77 MB. The 1.769–2.559 MB range captures variation across published page-type medians rather than the full page-weight distribution, whose upper tail is materially heavier. Using the lighter inner-page median as the low bound is conservative with respect to the inversion, since heavier pages raise the search-side allocation. At Nokia's reported 0.17 kWh/GB case-study value, transferring the 2.164 MB overall median page is allocated approximately 0.368 Wh of network electricity. Under the model's central 0.14 kWh/GB coefficient, the corresponding allocation is approximately 0.303 Wh. Both values exceed the central standard-inference value used in Scenario B before the client device draws a single watt.

Related work and analytical gap

2.1 The server-centric measurement tradition

The benchmark for search-engine energy was established in 2009, when Google disclosed that one query consumed approximately 0.3 Wh, including indexing and retrieval. This figure remained the canonical public reference for approximately fifteen years, despite substantial changes in hardware, software and data-centre efficiency. The stability was achieved, however, by optimising what sits inside the data centre, while the energy cost of what happens outside — traversing the network and rendering on the client — grew at an entirely different rate.

The early AI energy literature Strubell et al. (2019)Patterson et al. (2021) correctly identified training costs as a major concern. As deployment scaled, Luccioni et al. (2023) conducted the first systematic inference energy measurement. Epoch AI (2025) synthesised available evidence to estimate ChatGPT at approximately 0.3 Wh per query, noting this was 'relatively pessimistic'.

2.2 The emerging system-level perspective

The Green Software Foundation and related bodies have advocated for software carbon intensity metrics extending beyond the data centre. Morrison et al. (2025) proposed holistic lifecycle evaluation of language model creation, while a recent systematic review Oliveira et al. (2026) confirms that narrow, server-only operational boundaries systematically underestimate the true environmental impact of deployed AI. The critical contribution of the present paper is to extend this system-level thinking across modalities, comparing LLM sessions against search sessions on a common functional-unit basis.

2.3 The unexplored gap

No published peer-reviewed study has, to our knowledge, quantitatively compared session-level energy for LLM versus search modalities while incorporating the programmatic advertising energy overhead. Scope3 (2023) documented advertising's campaign-level carbon footprint, and Khan et al. (2024a, 2024b) measured ad-blocker impact on device power. These contributions have not been synthesised into a cross-modality CELCA using a common functional unit. A recent PRISMA-compliant systematic review Oliveira et al. (2026) corroborates this gap, finding that cross-modal lifecycle comparisons remain absent from the peer-reviewed literature and that heterogeneous functional units preclude cross-study synthesis; a limitation this paper directly addresses.

The energy physics of LLM inference in 2025

3.1 The production benchmark

The most rigorous publicly available production measurement was published by Google in August 2025 Elsworth et al. (2025). The paper measures a comprehensive stack including active TPU/GPU power (0.14 Wh, 58%), host CPU and DRAM (0.06 Wh, 25%), idle machine provisioning (0.02 Wh, 10%), and data-centre PUE overhead (0.02 Wh, 8%), yielding a median of 0.24 Wh per Gemini Apps text prompt.

3.2 Corroborating independent evidence

Available evidence converges on a common order of magnitude rather than on a provider-independent constant. Google reports 0.24 Wh for its median Gemini Apps text prompt in Elsworth et al. (2025). OpenAI has disclosed approximately 0.34 Wh for an average ChatGPT query without publishing a directly comparable system boundary. Oviedo et al. (2026, Joule) estimate a median of 0.31 Wh per standard frontier-scale query under a comprehensive bottom-up simulation, with an interquartile range of 0.16–0.60 Wh, concluding that prior estimates based on non-production assumptions systematically overstate energy use by 4–20×.

Scenario B adopts the peer-reviewed 0.31 Wh median as its central inference value and uses the published interquartile range for adverse-inference sensitivity. These estimates remain provider-, model-, hardware-, and workload-dependent. Vendor production measurements and independent bottom-up modelling provide triangulation, not an independently measured universal production constant.†

3.3 The reasoning model tier (out of scope)

SOTA reasoning models† generate extended chain-of-thought sequences, even in mid-tier variants. Drawing on recent benchmarks Hugging Face AI Energy Score v2, Dec 2025ML.Energy Leaderboard v3.0, 2026, we derive estimates for leading reasoning queries at 1.0–5.0 Wh per query — approximately 3–16× the 0.31 Wh standard-inference reference used in Scenario B. More extreme benchmark configurations report substantially larger multipliers under sustained test-time compute and long output sequences. This tier is explicitly out of scope; for the class of query most comparable to web search, standard models are both adequate and preferred.

3.4 Training and indexing: excluded from the central model

For a frontier model at 50 GWh training energy, deployed over two years serving 500 million queries/day:

While this represents a non-trivial overhead to the operational inference energy, both LLM training and traditional search engine crawling/indexing operations are massive, continuous background processes. They are therefore omitted from the session-level budget.

The resulting 0.14 Wh value is an illustrative amortisation scenario based on assumed training electricity, deployment lifetime, and query volume. It is not an observed per-prompt training value. Training and search crawling/indexing are excluded from the central operational session model because comparable attribution data are unavailable for both modalities, not because the processes are assumed equivalent. Their exclusion is one reason why CELCA is described here as a screening and partial-lifecycle framework rather than a complete lifecycle assessment.

Anatomy of the modern search session

4.1 Web page weight in 2025

The HTTP Archive 2025 Web Almanac reports a median mobile page weight of 2.16 MB across page types. Disaggregated by page type, the median mobile home page is 2.56 MB and the median mobile inner page is 1.77 MB.

Because search-result clicks commonly lead to inner content pages rather than home pages, Scenario B uses 2.16 MB as its central overall-mobile value and reports inner-page and home-page variants separately in §6.3. A typical LLM synthesis response is a structured text payload of 2–10 KB. The network-transmission ratio between a 2.16 MB median webpage and a 5 KB LLM response is approximately 433:1, before accounting for supplementary scripts, advertising payloads, and tracking pixels.

4.2 Mobile network energy intensity

Nokia's engineering white paper reports an average allocation of 0.17 kWh/GB for its Finnish aggregate mobile-network case study. The 0.14 kWh/GB central value used in this paper is a model-derived intermediate coefficient between the 0.10 kWh/GB low-allocation case and Nokia's reported 0.17 kWh/GB value. It is not itself a direct measurement or a universal network constant. At this coefficient, transferring the 2.164 MB overall median page is allocated approximately 0.303 Wh of network electricity; at Nokia's reported 0.17 kWh/GB the corresponding allocation is 0.368 Wh. Across the 0.10–0.17 kWh/GB attributional range, a three-page mobile search session is allocated approximately 0.65–1.10 Wh of network electricity, compared to effectively zero for a text-only LLM response. As of Q1 2026, over 95% of nominally 5G traffic in France and across Europe operates in Non-Standalone (NSA) mode — routing through a 4G core network. The continued prevalence of Non-Standalone 5G supports retaining 4G-era evidence as a reference point, but core-network architecture alone does not establish the complete electricity intensity of the radio-access network Ookla & Omdia, 2026MedUX, 2026.

4.3 Client device energy

Modern laptops draw 6–18 W during active browsing; flagship smartphones 2–4 W. The CHI 2025 experimental study by Spatharioti et al. found that LLM participants completed tasks more quickly with fewer queries than traditional search users, directly reducing total device energy through shorter screen-on time.

4.4 Zero-click asymmetry and the hidden cost of AI-augmented search

Similarweb data from July 2025 reported that 69% of Google searches end without a click to any website. For these queries, search energy approximates the query cost alone (≈0.3 Wh), matching the LLM baseline. The efficiency advantage emerges for the ≈31% of queries requiring website visits.†

The programmatic advertising energy overhead

5.1 The real-time bidding mechanism

When a user lands on an ad-supported webpage, a programmatic auction initiates in parallel with content loading. The publisher's SSP broadcasts a bid request to dozens or hundreds of DSPs. Each DSP processes the request within a 100–300 ms deadline. Research has documented extreme cases of a single ad slot auctioned across thousands of intermediaries, with the vast majority of bid computations producing no output of value to the user.

5.2 Client-side energy overhead from ad rendering

Khan et al. (2024a) report reductions of up to 44% in measured component power for integrated ad-blocking browsers relative to conventional browsing under their tested conditions. A companion comparison Khan et al. (2024b) reports a reduction of approximately 15%. These values are not a general 15–44% share of whole-device smartphone electricity: they are device-, browser-, website-, and protocol-specific results.

The 30% value used in Scenario B is a rounded model midpoint of the reported 15–44% endpoints. It is not a published median. In the central ledger, it is applied only to the modelled incremental rendering term.

Advertising is not assumed absent from the web session. Advertising assets are already represented in observed webpage payloads, while the separate 0.18 Wh line models an incremental client-computation burden. Because that decomposition remains uncertain, §6.3 reports a single-inclusive-client-term stress test removing both rendering and the standalone advertising increment.

5.3 Server-side ad-tech carbon footprint

Scope3's Q1 2023 State of Sustainable Advertising report estimated 215,000 metric tonnes of CO₂-equivalent per month generated by programmatic advertising in five major economies. We do not apportion this campaign-level server-side figure to individual page views. Client-side advertising overhead is instead treated through the explicitly modelled and stress-tested specification described in §5.2 and §6.3.

5.4 Server-side ad-tech energy: a quantified estimate

The Ad Net Zero Global Media Sustainability Framework V1.2 (June 2025) now provides explicit formulas permitting quantitative allocation of server-side programmatic overhead. Using the framework's published defaults — server use-phase intensity of 3.41× 10⁻⁷ kWh per ad opportunity, server factor 1.412, call factor 1.464, and average RTB payload of 3 KB — a standard ad-supported page with 3–5 ad slots (each triggering dozens of DSP bid requests) generates approximately 0.05–0.12 Wh of server-side energy from RTB bidding alone. Adding creative delivery and network overhead brings the estimated total server-side ad-tech burden to 0.10–0.25 Wh per page load.

These figures are distinct from, and additive to, the client-side rendering overhead quantified in §5.2. For a three-page mobile search session (Scenario B), they represent a structural server-side overhead of approximately 0.30–0.75 Wh — energy entirely absent from an LLM session. This server-side ad-tech estimate is deliberately excluded from the quantitative session totals in §6 and Appendix A, which reflect only the primary search inference and modelled network and client-side energy. Because real-time bidding energy scales highly variably per page, the server-side estimate is excluded from the central ledger and reported separately as the RTB-inclusive specification in §6.3, where it raises the Scenario B ratio from 5.4× to 6.1×.

Comparative energy lifecycle assessment

6.1 Methodology and system boundary

The assumption of 2–5 pages visited for complex synthesis tasks draws on two convergent sources: the CHI '25 randomised experiment Spatharioti et al. found participants in the traditional-search condition issued an average of 2.5 queries per task (95% CI [2.1, 3.0]); and cross-industry benchmarks place research-oriented organic-search sessions at 5–7 pages per session LuckyOrange, 2025Databox, 2025. The 2–5 page range is therefore a stylised scenario range informed by convergent behavioural evidence; it is not an empirical population distribution of pages visited.

Included for both modalities: Server-side computation (including data-centre PUE); core and last-mile network transmission; client-device energy during active task engagement; advertising payload rendering for search sessions.

Excluded from the central model for both modalities: Model training/index crawling (see §3.4); embodied carbon; idle device energy.

Accounting frame: The primary model is attributional. Network and server terms allocate shares of current operating electricity according to transferred data and query counts. They do not estimate the short-run marginal electricity that would physically appear or disappear if one session were added or removed. Aggregate substitution figures are therefore differences in attributed operational electricity, not demonstrated grid savings.

Functional unit: The complete user session required to satisfy one complex information need, defined as a task requiring synthesis or comparison of information from multiple sources.

Methodological status: This is an attributional, study-specific Comparative Energy Lifecycle Assessment (CELCA). The framework is not presented as an ISO 14040/44-conformant comparative assertion, a complete multicriteria LCA, or a cradle-to-grave assessment. It is best described as a screening-level assessment of final operational electricity within the stated service-delivery boundary. Model training, search crawling and indexing, embodied infrastructure, primary-energy demand, end-of-life impacts, and other environmental impact categories remain outside the central quantitative model.

6.2 Comparative scenarios

6.3 Primary specification set and stress tests

The 5.4× efficiency ratio reported in Scenario B reflects a central specification. To establish robustness, we report the outcome across seven deterministic specifications testing selected parameters individually around the central ledger, plus three stress tests that deliberately disable core model components or adopt adverse extremes.

Sensitivity analysis

7.1 Parameter ranges

Primary sources for CELCA scenarios.†‡§

7.2 Monte Carlo sensitivity results (Scenario B)

The explorer below recomputes Scenario B from the canonical ledger. The Monte Carlo facility samples uniformly across the Table 1 bounds; reported statistics are generated on execution and are bounded-sensitivity outputs, not empirical confidence intervals.

7.3 The Wi-Fi boundary constraint

The magnitude of the inversion is strongly network-dependent. It is largest on cellular connections, where network allocation is a major component of the search session; on fixed connections the result is carried primarily by the client-side decomposition. On fixed Wi-Fi (0.006 kWh/GB), the central Scenario B search total falls to approximately 1.37 Wh and the search/LLM ratio to approximately 3.3×. Because network allocation becomes small, this result is dominated by the client-side decomposition and should be read alongside the single-inclusive-client-term stress test in §6.3, which removes the separate rendering and advertising increments. Combining both concessions — fixed Wi-Fi allocation and removal of the separate rendering and advertising increments — yields a search total of 0.589 Wh against 0.414 Wh for the LLM session, a ratio of approximately 1.4×. This is the deterministic floor of the specification space explored here: the inversion persists even when network allocation is reduced to fixed-line levels and both disputed client increments are removed.

Behavioural dynamics and the time-on-task multiplier

Energy efficiency and time efficiency are coupled through device power draw. The CHI 2025 study by Spatharioti et al. used a randomised between-subjects design for product research tasks. Key findings: LLM participants completed tasks more quickly with fewer queries; the modal query count for LLM users was one versus two for search users; decision accuracy was comparable when LLM output was accurate.

The 'pogo-sticking' behaviour documented in web usability research — clicking a result, finding it unsatisfactory, returning to the SERP, trying another — creates an energy penalty not captured in static page-count models. Under the Scenario B ledger, each additional page interaction adds approximately 0.30–0.60 Wh through network transfer, rendering, advertising and active-device time, depending on page weight and specification. LLM interfaces structurally eliminate this penalty by delivering a synthesised answer in a single interaction.

Counter-arguments: a rigorous interrogation

9.1 The Jevons paradox

Making information retrieval cheaper will induce more demand. ChatGPT reached 800 million weekly active users by late 2025, with 2 billion daily queries. If this represents new demand rather than substituted demand, aggregate energy grows regardless of per-session efficiency gains.

The scope clarification here is essential: this paper evaluates unit efficiency for a defined task, not aggregate societal energy consumption. The Jevons paradox validates rather than refutes the unit-efficiency argument — demand rises because efficiency improves. Policy responses at the aggregate level are legitimate and complementary, not contradictory.

9.2 The hallucination verification penalty

If users must verify LLM outputs with a follow-up search, the session energy becomes additive. Even in a hybrid workflow with one verification search, total energy typically remains below the unstructured multi-page session. The LLM session (approximately 0.414 Wh) can be supplemented with a single verification query. Adding one search query (0.30 Wh), one page download (1× 2.164 MB × 0.14 kWh/GB = 0.303 Wh), one page render (0.20 Wh), advertising overhead (0.20× 0.30 = 0.06 Wh), and two minutes of reading time (2.5× 2/60 = 0.083 Wh, a scenario assumption) yields a hybrid total of approximately 1.36 Wh — against the central search total of 2.24 Wh, a ratio of approximately 1.6×.

9.3 Scope limitation: agentic and reasoning workflows

The efficiency advantage applies specifically to standard non-reasoning LLM inference serving text synthesis queries on optimised commercial infrastructure. It does not apply to reasoning models (§3.3), agentic workflows combining programmatic web retrieval with LLM inference, test-time compute (TTC) architectures that extend inference through chain-of-thought generation, or multi-turn conversations consuming reasoning tokens implicitly.

This risk mirrors a fundamental thermodynamic constraint of deep learning documented by Yang et al. (2024), where a tenfold (10×) increase in energy yields only a marginal ~3% increase in model accuracy at the frontier. This "accuracy-at-all-costs" regime is rapidly arriving in LLM inference. Harvard's audit of TTC protocols Jin et al. (2025) recorded an average 4.4× token explosion. The thermodynamic cost of this expansion is severe: Oviedo et al. (2026) estimate that long reasoning queries (~5,000 tokens) consume a median of 3.91 Wh, a 13× increase over standard queries.

We can invert the analysis to locate the thermodynamic breaking point: at what reasoning-token expansion factor does the LLM efficiency advantage disappear entirely? Depending on the base inference assumptions applied from Table 1, the model identifies a crossover threshold between roughly 4× and 8× against the mobile search baseline. Jin et al. (2025) independently report a mean production reasoning expansion of 4.4×, with more demanding cases reaching 10× and, in the extreme, 113×. The convergence here is structural, not statistical: the mechanistically derived parity threshold sits within the observed operating envelope of current reasoning models. The efficiency advantage of standard LLM inference over mobile web search cannot be assumed to extend to reasoning workflows.

Below approximately 3× token expansion, the LLM retains a meaningful efficiency advantage across the vast majority of modelled scenarios. Between 3× and 10×, that advantage collapses into parity or net loss. Above 10×, a threshold documented in extreme production cases, the LLM becomes substantially more energy-intensive than the web session it replaces. This is a regime change. The efficiency narrative surrounding standard LLM inference cannot be straightforwardly extended to reasoning or agentic workflows; they constitute a categorically distinct, potentially electricity-intensive operating regime. This reinforces the absolute necessity of difficulty-aware model routing as a first-order sustainability intervention.

9.4 Asymmetric embodied carbon

GPU/TPU manufacturing (TSMC 3nm/4nm nodes) is energy-intensive. We flag this as a limitation and recommend a full Scope 3 lifecycle assessment for future work, noting that the web's continuously refreshed ad-tech server fleet also carries substantial embodied carbon.

9.5 The "conservative baseline" validation

Early drafts of this assessment relied on Google's 0.24 Wh production measurement, which invited legitimate questions regarding provider control of the underlying infrastructure and telemetry. Oviedo et al. (2026) provide peer-reviewed bottom-up modelling in the same broad order of magnitude, estimating a median of 0.31 Wh and an interquartile range of 0.16–0.60 Wh for standard frontier-scale inference. This convergence supports the plausibility of the selected order of magnitude without establishing a provider-independent production constant. Scenario B adopts the 0.31 Wh median centrally and the 0.60 Wh upper-quartile value as an adverse-inference stress test, under which the search/LLM ratio narrows to approximately 3.2×.

Policy implications and research agenda

10.1 For corporate sustainability officers

Organisations seeking to minimise their digital information-retrieval footprint should: (i) prioritise mobile-first LLM deployments for research and synthesis tasks over traditional search workflows on cellular connections; (ii) audit ad-tech exposure: selected controlled studies report reductions of approximately 15% to 44% in measured browser-component power under their tested desktop and browser conditions, but these values should not be transferred directly to whole-device smartphone sessions; (iii) resist reasoning-model adoption for tasks that standard models handle adequately; (iv) incorporate session-level energy accounting into digital sustainability reporting.

10.2 For regulators and policy-makers

This distinction is now entering institutional analysis. In their 2026 joint report, the French telecom regulator (ARCEP), working with PEReN, highlights that evaluating generative AI requires moving beyond isolated query metrics toward the total energy required to deliver the service rendered, including the advertising-auction overhead of traditional search alternatives. Imposing unit-energy taxes on LLM queries without accounting for this full-stack alternative-use-case risks creating perverse incentives.

10.3 Research agenda

1. Empirical hallucination rate data disaggregated by query type, with energy impact modelling for verification workflows.
2. Independent, multi-provider inference energy benchmarks across production-realistic workloads with comprehensive system boundaries.
3. Full Scope 3 lifecycle assessment for LLM and search infrastructure including embodied hardware carbon.
4. Field measurement of cellular modem energy during LLM vs. search data payloads.
5. Economic analysis of the content-creator/publisher externality: LLMs substituting for web visits reduce advertising revenue for publishers whose content trained the models.

Conclusions

This study provides a bounded, screening-level comparative lifecycle assessment identifying a structural thermodynamic efficiency inversion between ad-supported mobile search and standard LLM inference. For complex synthesis tasks performed on mobile devices, the central model allocates approximately 5.4× as much final operational electricity to the web search session as to the equivalent LLM session. This advantage is structurally driven by three compounding factors:

• The high energy intensity of mobile cellular data transmission applied to the large payloads of modern webpages
• The device energy overhead of the ad-tech supply chain, which consumes browsing power without delivering informational value to the user
• Reduced device screen-on time from faster task completion, validated experimentally at CHI 2025

The directional finding is resilient across the specifications tested. The search session remains 3.5× more energy-intensive under an inclusive-client stress test, and 3.2× under an adverse-inference assumption.

These advantages narrow for simple queries on Wi-Fi, and narrow toward parity or reverse for reasoning-model inference or agentic workflows. The Jevons paradox ensures that unit efficiency gains do not guarantee aggregate efficiency gains, and the rapid growth of AI query volume is a legitimate supply-side concern independent of unit efficiency.

The practical implication is conditional but actionable: for eligible, non-reasoning synthesis tasks on cellular connections, routing the task through a standard LLM interface materially reduces the final operational electricity allocated to the retrieval session under the modelled boundary. At population scale, the efficiency delta is significant. As an illustrative attributional scenario, if 500 million daily sessions already meeting the Scenario B conditions were substituted — 182.5 billion sessions annually — the difference in attributed operational electricity would be approximately 333 GWh per year (182.5× 10⁹ × 1.825 Wh). This is an allocation difference under the accounting frame of §6.1, not an estimate of physical grid electricity saved. A consequential estimate would require network, data-centre, and capacity-response models outside the scope of this assessment.

This finding should inform corporate digital sustainability strategies, regulatory impact assessments, and the emerging discipline of sustainable information retrieval. The ultimate implication of this assessment is not merely that LLM queries can be cheaper than search queries. It is that the dominant paradigm of the ad-supported, multi-page mobile web carries a structural, largely invisible energy tax — and that under the attributional framework modelled here, substitution for eligible tasks would reduce the final operational electricity allocated to information retrieval. For the class of synthesis tasks identified here, generative AI acts as a thermodynamic compression engine, sparing the mobile network and client device from the accumulated burden of the modern web.