{"id":"30ede03d-fd7b-42fd-ac5d-efb8f6ad4c40","arxiv_id":"2411.11540","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Cloud energy consumption is proportional to a cloud platform's cumulative revenue, according to a thermodynamic growth model fitted to Meta and Google data.","lead":"A thermodynamics model imported from climate economics is applied to cloud computing, concluding that cloud energy use grows with the total revenue a cloud company has ever earned. It is worth reading because it offers a simple quantitative story for why more efficient data centers keep consuming more energy, and it exposes weaknesses in current cloud sustainability metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation uses company-wide revenue, not cloud revenue, so the central A = αC claim is not actually tested for cloud computing.","rationale":"The reader identified the weakest assumption as 'cloud revenue is a suitable work metric' and specifically mentioned advertising. My stress-test sharpens this into the concrete empirical mismatch: the validation data are not cloud-segment revenue. This is the most load-bearing concern because the entire derivation and validation rest on the work metric being revenue of the cloud platform. If Meta and Alphabet total revenue are used instead, the claim becomes about internet companies generally, not cloud computing. The mathematical skeleton (Equations 1–6) is internally coherent, so the issue is not a derivation error but an identification failure in the empirical test. The paper also contains a secondary technical issue: deriving A = αC from Equation 3 requires assuming zero initial interface L(0); otherwise an intercept αΔΦL(0) appears. This is potentially fixable by starting cumulative revenue at company founding, but it is not stated. The back-casting circularity in Section 4 affects only the CapEx proxy (Figure 5), not the main A = αC claim, so it is secondary. The proposed concrete test — using actual cloud segment revenue from multiple providers — would settle whether the central claim has empirical support. If it fails, the paper would reduce to a speculative analogy; if it passes, the thermodynamic model gains credibility. Given that the data could be reanalyzed with publicly available segment revenue, the verdict remains CONDITIONAL: conditional on cloud-segment validation.","tokens_in":10306,"tokens_out":6138,"duration_ms":65437,"concrete_test":"Recompute the A = αC analysis using segment-level cloud revenue: Google Cloud segment revenue from Alphabet 10-K filings versus Google data center energy; Microsoft Azure segment revenue versus Microsoft data center energy; and AWS revenue versus Amazon data center energy. Report the fitted α, R², residuals, and confidence intervals for each, and compare against null models where A is fit to contemporaneous revenue or to independent exponential trends. If the segment-level proportionality does not hold with acceptable fit (or the null model fits equally well), the central claim is unvalidated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central relation, A = αC (energy proportional to cumulative revenue), is derived by identifying work with 'cloud revenue' (Section 3: 'we claim that cloud revenue is a suitable work metric'). But the validation in Section 4 uses Meta's total revenue, which is overwhelmingly advertising revenue, and Alphabet's total revenue, not the Google Cloud segment. Meta is not a cloud infrastructure provider; its 'work' is ad-driven user engagement. This contradicts the stated system boundary in Section 4: 'We only consider the cloud provider as part of the system, and not the users' — yet total company revenue is generated by users and advertisers. If A = αC holds for total company revenue, it may simply reflect that both revenue and energy grew roughly exponentially for these companies over 2016–2022, with no evidence that the thermodynamic mechanism is specific to cloud platforms. The fit uses one free parameter per company with no error bars, no R², and no comparison against a null model (e.g., A proportional to contemporaneous revenue, or independent exponential trends), so the 'validation' cannot distinguish the claimed mechanism from a spurious correlation between two growing series. This is the load-bearing weakness because the paper's central claim — that cloud revenue determines cloud energy consumption — is not actually supported by cloud-segment data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a coarse-grained thermodynamic model of cloud computing, adapted from Garrett's models of the global economy. The system consumes energy A = αLΔΦ; work (identified with cloud revenue) is W = εA; and the work is assumed to be reinvested in expanding the system's interface, W = (dL/dt)ΔΦ. Combining these equations yields the system-inertia relation A = αC, where C is cumulative revenue, and exponential energy growth dA/dt = ηA. The author claims this explains the Jevons paradox for hyperscale clouds and validates the model using public energy and financial data from Meta and Google (2016–2022): energy versus cumulative revenue (Figure 3), embodied emissions versus cumulative energy (Figure 4), and capital expenditure versus model-back-cast cumulative energy (Figure 5). The paper then draws implications for efficiency metrics, client-side energy, and edge-cloud architectures. The thermodynamic derivation is internally consistent, but the empirical validation uses company-wide rather than cloud revenue, reports no quantitative fit statistics, and includes a circular back-casting step.","tokens_in":10479,"tokens_out":14932,"duration_ms":144942,"significance":"If the central proportionality A = αC could be established with cloud-segment data, the paper would offer a useful quantitative regularity linking an easily measured financial variable to cloud energy use, and a mechanism-based account of the Jevons paradox that goes beyond the standard elasticity story. The strengths of the manuscript are real: the derivation from Eqs. (1)–(6) is transparent and internally consistent; the model makes explicitly falsifiable predictions (A ∝ C with stable α; embodied emissions ∝ cumulative energy; CapEx ∝ cumulative energy); the empirical work uses publicly available data; and the paper candidly discloses the Google emissions failure and the back-casting step. These strengths, however, do not compensate for the validation gap: the headline empirical claim is tested with total-company revenue for two firms, with no fit statistics or baseline comparisons, and with a system boundary that shifts between model and data. The significance for a journal readership is therefore conditional on a substantially strengthened evaluation, in particular a test using actual cloud-platform revenue and quantitative goodness-of-fit measures.","major_comments":[{"comment":"The validation in Fig. 3 does not test the model's work metric. Section 3 defines work as 'cloud revenue' ('we claim that cloud revenue is a suitable work metric'), but the evaluation uses total company revenues from SEC filings: Meta's consolidated revenue is almost entirely advertising (Meta sells no public cloud service), and Alphabet's figure is dominated by Google Services and advertising rather than the Google Cloud segment. This contradicts the system boundary stated at the start of Section 4 ('We only consider the cloud provider as part of the system, and not the users'). Total revenue is generated by users and advertisers outside that boundary, so the fitted proportionality A = αC is a statement about company-wide revenue, not about cloud-platform revenue, and the central claim is not actually tested on cloud revenue. The derivation also assumes that all work (revenue) is reinvested in interface growth, Eq. (3), whereas both companies paid substantial dividends and conducted share repurchases during 2016–2022; the paper should at least quantify how much of the revenue is excluded by that assumption.","section":"§4 (Model Evaluation), Fig. 3; §3 (work metric)"},{"comment":"The CapEx analysis in Fig. 5 is not independent confirmation of the model. As the text discloses, 'To estimate the cumulative historical (pre-2016) energy usage, we use our model's linear relation from Figure 3.' The cumulative-energy series used to compute the ratio CapEx/cumulative energy is therefore generated from the very A = αC relation that the paper is validating, so the agreement in Fig. 5 is partly built in. The disclosure is commendable, but the section should be reframed: Fig. 5 illustrates a back-casting application of the model; it does not provide independent evidence for the size-proxy claim or for the proportionality.","section":"§4 (Model-based back-casting), Fig. 5"},{"comment":"No goodness-of-fit or baseline comparison is reported for the central relation. Figure 3 shows the declared linearity visually, but the paper gives no R², no residuals, no error bars, and no uncertainty on the fitted slope α; nor is α derived independently, so the dashed lines are a functional-form consistency check rather than a parameter-free prediction. Because both revenue and energy grew roughly monotonically over 2016–2022 for both companies, visual proportionality between two growing series is a weak test: an alternative such as A proportional to contemporaneous revenue, or two unrelated exponential trends, would be hard to distinguish by eye. With two firms and seven annual points, this is the main empirical support for A = αC, and quantitative measures of fit plus at least one null model are needed before the statement 'the model prediction holds true' is warranted.","section":"§4 (Model Evaluation), Fig. 3"},{"comment":"The system boundary is inconsistent between the model and the data. Section 3 defines A as 'the electricity consumption of the cloud data centers and the client-side consumption' and states that the system includes users, whereas Section 4 restricts the system to the provider-only boundary and uses data-center electricity from sustainability reports only. The relation A = αC is thus fitted to a different quantity than the A in Eqs. (1)–(6). The author should either re-derive the model for the provider-only boundary or validate the full-boundary model with the client-side and internet-energy estimates that Section 5.2 shows to be significant; as it stands, the data and the model do not refer to the same system.","section":"§3 vs. §4 (system boundary for A)"},{"comment":"The embodied-emissions size proxy is confirmed for only one of the two companies: the text reports that the Google analysis 'fails to show the proportional relation between Scope 3 emissions and cumulative energy.' The paper explains this by the non-standard nature of emissions accounting, but the consequence is that the interface-size branch of the validation rests on a single company. The finding should be stated accordingly (for example, 'consistent for Meta only'), rather than as a general confirmation that emissions are a usable proxy for system size.","section":"§4 (Model Evaluation), Fig. 4"}],"minor_comments":[{"comment":"Equation (7), A(t) = e^{ηt}, omits the prefactor A(0); as written it is dimensionally inconsistent and would imply A(0) = 1 in whatever units are used. It should read A(t) = A(0)e^{ηt}.","section":"§3, Eq. (7)"},{"comment":"Please correct the typos and garbled sentences: 'phenomonological' → 'phenomenological' (§3), 'minitua' → 'minutiae' (§1), 'inspite' → 'in spite' (§§1, 4), 'reason-detrie' → 'raison d'être' (§3), and 'conventional PUE-improving have data center optimizations' in §5.3, which is not a grammatical sentence.","section":"Throughout"},{"comment":"The caption 'Cumulative Revenue/Energy' is ambiguous: it does not state that the two series are normalized for comparison, nor how the slope of the dashed model line was determined (fitted to the same data? derived from an independent constant?). Both points should be stated explicitly.","section":"Fig. 3 caption"},{"comment":"The assumptions are stated inconsistently: 'We assume 2 hours a day of usage, with 1 Watt energy usage' gives 2 Wh/day only if the 1 W is the average draw over the 2 hours, and the later sentence states 'we assume 2 Watt-Hours per day.' The client-usage (2 h/day) and internet-attribution (10%) parameters are free choices; since the figure's purpose is to show that client energy can be significant, a brief sensitivity discussion would strengthen it.","section":"§5.2, Fig. 6"},{"comment":"The statement that the model is 'adapted from prior work in [10]' under-specifies the provenance: Eqs. (1)–(3) and the L = N_S^{1/3}·N_R remark appear to come from Garrett's series [8, 9, 11, 12] and the bio-economics literature [14, 15]. Please cite each equation to its source.","section":"§3, references"},{"comment":"The ACM template placeholders ('Conference'17, July 2017', '7 pages', 'ACM ISBN 978-x-xxxx-xxxx-x/YY/MM', 'https://doi.org/10.1145/nnnnnnn.nnnnnnn') remain in the manuscript and should be removed.","section":"Manuscript formatting"},{"comment":"No scripts or data repository are provided for reproducing Figures 3–5. Given that the empirical claims are the load-bearing part of the paper, a small release of the extracted time series and fitting code would materially help reviewers and readers.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closer in structure to a workshop or vision paper, and its contribution over Garrett's existing thermodynamic models of the global economy is the translation to the cloud setting, which is legitimate but should not be overclaimed. The load-bearing weakness is the validation: the paper's central relation is tested against total-company revenue rather than cloud-segment revenue, with no quantitative goodness-of-fit, and Fig. 5 is partly circular. These problems are fixable within the manuscript's scope (cloud-segment revenue for Google Cloud or Microsoft Azure; R², residuals, and a null model; reframing the back-cast as an application; aligning the system boundary), so I recommend major revision rather than rejection. I would also encourage the author to include at least one true public cloud provider and to reconsider using Meta, whose ad-driven business model conflicts with the provider-only boundary adopted in Section 4."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper is the first to apply Garrett's thermodynamic growth model to cloud platforms, and the central relation A = αC is a genuinely interesting empirical target. But the validation doesn't actually test it. In Section 4, 'cloud revenue' turns out to be Meta's total revenue (overwhelmingly advertising) and Alphabet's total revenue, not Google Cloud segment revenue. Meta is an advertising company, not a cloud infrastructure provider. So the clean-looking linear fit between energy and cumulative revenue may simply reflect two growing series over 2016–2022, with no evidence that the thermodynamic mechanism is specific to cloud.\n\nWhat's new and good: the domain application itself, revenue as a work metric, the derived exponential growth and system-inertia framing, and the side analyses of embodied emissions and CapEx as interface-size proxies. The derivation from A = αLΔΦ, W = εA, and W = (dL/dt)ΔΦ is internally consistent. The paper is clearly written, uses public data, and is honest when the Scope 3 emissions proxy fails for Google. That intellectual honesty is real.\n\nThe soft spots are in the empirics, not the math. The A vs. cumulative-revenue fit uses one free parameter per company, with no error bars, no R², and no comparison against a null model like a contemporaneous revenue relation or independent exponential trends. The back-casting step for Google in Figure 5 uses the model's own A–C linear relation to generate pre-2016 energy data, and then uses that generated data to validate the CapEx proxy. That is partly circular. The assumptions that revenue is fully reinvested in expanding the system interface, and that ΔΦ stays constant, are strong and not independently supported.\n\nThese problems are fixable. Use actual cloud-segment revenue (Google Cloud, AWS, Azure), test out-of-sample, and compare against a simple exponential baseline. If the relation survives that, the paper would be much stronger. As it stands, the word 'validated' is too strong.\n\nWho this is for: researchers in cloud sustainability, Jevons paradox/rebound effects, and macro energy modeling. It deserves a serious referee because the claim is important and testable, even if the current evidence is weak. My recommendation: engage with the model and the empirical strategy, but push for a proper validation before accepting the central claim.","headline":"A clean, first-of-its-kind application of Garrett's thermodynamic growth model to cloud platforms, but the central validation claim is undercut by the use of total company revenue instead of cloud revenue.","tokens_in":11064,"tokens_out":1798,"would_cite":true,"duration_ms":20138,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that cloud energy consumption is proportional to cumulative historical revenue, so efficiency improvements accelerate total energy growth rather than curbing it.","keywords":["Jevons paradox","rebound effect","cloud energy efficiency","thermodynamic model","cumulative revenue","hyperscale data centers","sustainability metrics"],"falsifier":"Compute, year by year over a decade, the ratio of a hyperscale cloud's total electricity use to its cumulative historical revenue. If that ratio trends away from a constant while revenue grows, the identity $A = \\alpha C$ and the implied constant-potential exponential growth are falsified; a second check is whether data-center energy plateaus or declines for several consecutive years while cumulative revenue continues to rise, which would contradict $dA/dt = \\eta A$.","tokens_in":9959,"feed_emoji":"⚡","tokens_out":8689,"duration_ms":76955,"temperature":0.7,"pith_summary":"The paper asks why the total energy consumed by cloud computing keeps rising even as data centers grow more efficient. Its answer is a feedback loop borrowed from thermodynamics: efficiency lowers the cost of doing work, that work (measured by revenue) expands the cloud's boundary with an energy reservoir, and the larger boundary draws more energy from the reservoir. The central prediction is that current energy consumption is proportional to cumulative historical revenue, $A = \\alpha C$, so any efficiency improvement raises the exponential growth rate of energy use instead of reducing it. The author reports that energy and financial data from two hyperscale cloud providers over 2016–2022 fit this linear relation, together with capital-expenditure and emissions proxies for system size. If accepted, the result shifts the goal of cloud sustainability from point efficiency improvements to limiting system size and inertia.","feed_headline":"Why greener clouds burn more energy: efficiency fuels growth","feed_subtitle":"A thermodynamic model ties cloud energy to cumulative revenue, so efficiency gains accelerate total energy use.","key_machinery":"The carrying mechanism is the efficiency-driven growth loop: revenue $W$ is work, work expands the interface $L$, and energy draw $A$ is proportional to $L$. Algebraically, $\\varepsilon = W/A = (1/\\alpha)\\,d\\ln L/dt$, so a constant efficiency is a constant exponential growth rate for the system; with $\\eta = \\alpha\\varepsilon$, the energy growth equation $dA/dt = \\eta A$ follows directly. The testable identity $A = \\alpha C$ (current energy proportional to cumulative revenue) is the model's signature, and the validation section checks it against 2016–2022 data.","core_discovery":"The paper claims that a hyperscale cloud behaves like an open thermodynamic system whose energy consumption is $A = \\alpha L \\Delta\\Phi$, where $L$ is the size of the interface with the energy reservoir and $\\Delta\\Phi$ is the reservoir potential. Taking cloud revenue as the system's work output, work expands the interface via $W = (dL/dt)\\Delta\\Phi$, and combining the equations yields $dA/dt = \\eta A$ with $\\eta = \\alpha\\varepsilon$, so energy grows exponentially as $A(t) = e^{\\eta t}$ and is proportional to cumulative revenue: $A = \\alpha C$. The paper validates this proportionality on public energy and financial data from Meta and Google, and shows that capital expenditure and embodied emissions scale with cumulative energy, supporting their role as proxies for interface size.","pith_inferences":["The same $A = \\alpha C$ proportionality could be tested on other large cloud platforms and on national data-center aggregates; a clear failure there would delimit the model's domain.","A saturation test is directly available: if energy growth visibly slows while cumulative revenue keeps rising, then $\\Delta\\Phi$ is not constant and the constant-potential exponential phase is ending.","Disaggregating revenue by product line could tighten or break the coupling, since advertising or non-compute income included in total revenue would distort the physical work metric.","The model implies that the only escape from exponential growth is constraining interface expansion, for example through capacity limits or edge-cloud boundaries, rather than through efficiency alone."],"forward_implications":["If the model is right, hardware and data-center efficiency improvements raise $\\eta$ and therefore steepen the exponential rise in total cloud energy, so point-level metrics like PUE become insufficient for sustainability.","Energy consumption will keep growing with revenue as long as the reservoir is treated as a constant potential; growth turns sigmoid only if the energy or material reservoir depletes.","Cumulative historical revenue, capital expenditure, and embodied emissions become usable proxies for system size, enabling energy back-casting from public financial data.","The system's inertia means today's energy use is set by past growth, so efficiency levers act on a system whose size has already been determined.","Including clients and the network in the system boundary shows that data-center-only optimizations can push consumption into less efficient parts of the system."],"supporting_citations":[{"why":"Supplies the base thermodynamic model relating global energy consumption to cumulative economic production, which the paper adapts to clouds.","marker":"[8]"},{"why":"Derives the exponential modes of growth and the entropy relation used for the growth rate $\\eta$.","marker":"[9]"},{"why":"States the efficiency-growth double-bind that motivates applying the thermodynamic model to cloud systems.","marker":"[10]"},{"why":"Provides the historical-inertia argument that current energy consumption is set by cumulative past work.","marker":"[12]"},{"why":"Google sustainability report, the public data source for validating the energy and cumulative-revenue relation.","marker":"[2]"},{"why":"Meta sustainability report, the second public data source for validating the model's proportionality.","marker":"[3]"}],"fun_headline_variants":["Cloud efficiency paradox: greening fuels bigger energy bills","Thermodynamics of clouds: why efficiency boosts energy use","More efficient clouds still burn more energy, model shows","Cloud energy growth tied to revenue, not efficiency alone","Jevons paradox in clouds: efficiency fans the flames"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on treating a cloud's revenue as a faithful measure of the useful work it does and assuming that all of that work is reinvested to grow the system against an effectively unlimited energy supply; if revenue includes value unrelated to computation, or if growth is capped by markets or capital, the proportionality between energy and cumulative revenue does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Cloud efficiency paradox: greening fuels bigger energy bills","Thermodynamics of clouds: why efficiency boosts energy use","More efficient clouds still burn more energy, model shows","Cloud energy growth tied to revenue, not efficiency alone","Jevons paradox in clouds: efficiency fans the flames"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000564,"raw_usage":{"total_tokens":2621,"prompt_tokens":839,"completion_tokens":1782,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":1704}},"tokens_in":455,"tokens_out":1782,"duration_ms":11408,"temperature":1.0,"reasoning_tokens":1704,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:23:48.753783+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, year by year over a decade, the ratio of a hyperscale cloud's total electricity use to its cumulative historical revenue. If that ratio trends away from a constant while revenue grows, the identity $A = \\alpha C$ and the implied constant-potential exponential growth are falsified; a second check is whether data-center energy plateaus or declines for several consecutive years while cumulative revenue continues to rise, which would contradict $dA/dt = \\eta A$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base thermodynamic model relating global energy consumption to cumulative economic production, which the paper adapts to clouds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the exponential modes of growth and the entropy relation used for the growth rate $\\eta$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"States the efficiency-growth double-bind that motivates applying the thermodynamic model to cloud systems."},{"cited_title":"J., Grasselli, M","cited_arxiv_id":null,"evidence_quote":"Provides the historical-inertia argument that current energy consumption is set by cumulative past work."},{"cited_title":"https://sustainability.google/reports/google- 2023-environmental-report/, 2023","cited_arxiv_id":null,"evidence_quote":"Google sustainability report, the public data source for validating the energy and cumulative-revenue relation."},{"cited_title":"https://sustainability.fb.com/2023-sustainability- report/, 2023","cited_arxiv_id":null,"evidence_quote":"Meta sustainability report, the second public data source for validating the model's proportionality."}],"review_version":1}