{"id":"a8fc24a7-5a69-4384-b652-50bd11a308ec","arxiv_id":"2412.17376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Training large AI models has become exponentially more energy-intensive and environmentally impactful since 2012, and efficiency gains and greener electricity alone cannot offset this growth.","lead":"This paper measures how the environmental impacts of training machine learning models changed from 2012 to 2024, using data on hardware and models. It finds that impacts grew exponentially despite efficiency improvements, and that shifting to greener energy would not stop the growth.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exponential-trend claim rests on a non-random 205/825 subset of Epoch AI systems; selection bias is untested, and Section 7 admits excluded models could change the trend.","rationale":"The reader's weakest assumption and my concern converge: the sector-level exponential claim depends on a 25% subset of the Epoch AI database, and the paper does not test whether the excluded systems follow the same growth trajectory. This is load-bearing because the headline conclusion is explicitly about trends, not point estimates: if the missing systems are disproportionately small or slow-growing, the fitted exponential slope is an artifact of sample selection rather than a sector-wide fact. Prior work on compute growth (Sevilla et al., Thompson et al.) gives independent support for exponential compute, but the paper's specific environmental-impact trend, and its claim that reduction strategies cannot curb it, is computed from this subset, so the representativeness question cannot be waved away. The concrete test is feasible with the Epoch AI database and the paper's own formulas. I am not moving the verdict because the reader already appropriately set CONDITIONAL; my check is the kind of robustness analysis that condition should require. I also credit the paper for sharing data and code and for explicitly acknowledging the database incompleteness, but an acknowledgment alone does not establish that the trend is robust.","tokens_in":11543,"tokens_out":5668,"duration_ms":57498,"concrete_test":"Re-fit the main trend using all Epoch AI systems for which both FLOP and documented hardware are available, regardless of whether training duration is reported. Compute GPU-hours as FLOP divided by peak FLOP/s of the documented hardware, apply the paper's ~27% utilization correction (from the Section 3.2 regression), and re-estimate the WLS trend with confidence intervals. Compare this slope and intercept to the 205-system subset used in the paper, and also compare the year, FLOP, and hardware-type distributions of included versus excluded systems. If the all-available slope is materially lower or the subset slope falls outside the confidence interval, selection bias is confirmed; if the slopes agree, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract, Section 5) is that training energy and environmental impacts grew exponentially and that location shifting cannot curb this. The analysis uses only 205 of 825 Epoch AI systems (Table 1): 112 with directly reported GPU-hours and 93 where GPU-hours are derived from FLOP divided by peak hardware throughput. This subset is not a random sample: inclusion requires either direct training-duration reporting or documented hardware plus FLOP, both of which are more common for large, recent, well-known systems. If excluded systems grow more slowly or are systematically smaller, the fitted WLS slope overstates the sector-wide trend. The paper's own Section 7 states that 'the Epoch AI database is incomplete... excluded models could have an impact,' but no test of selection bias is provided. The GPU-h2 correction issue in Section 3.2 is real but less damaging: applying or omitting a constant ~3.7x factor changes levels, not the log-linear slope, so it does not directly threaten the exponential-growth conclusion as much as sample composition does.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the environmental impacts of AI hardware and model training over 2013–2023. It curates a dataset of NVIDIA workstation and other graphics cards, estimates production impacts using the MLCA life-cycle assessment tool, and combines this with Epoch AI's notable-systems database to estimate the energy, carbon footprint (GWP), and abiotic depletion potential (ADP) of training individual ML models. The central claim is that training energy and environmental impacts have grown exponentially over time, that hardware production impacts also increase, and that reduction strategies such as shifting compute to less carbon-intensive electricity or improving PUE cannot curb this growth—interpreted as evidence of a rebound effect. The authors share code and data for reproducibility.","tokens_in":11792,"tokens_out":2991,"duration_ms":31157,"significance":"If the exponential-trend result is robust, the paper is a valuable contribution to the Green AI literature: it extends prior compute-trend studies (e.g., Sevilla et al., Thompson et al.) with an environmental multi-criteria assessment that includes embodied hardware impacts, and it explicitly interrogates impact-shifting and rebound effects—issues often ignored in single-phase carbon accounting. The use of MLCA, a previously published tool, applied to new data is appropriate, and the public availability of the curated data and code is a concrete strength. The main significance hinges on whether the fitted exponential trend is an artifact of sample selection and time-correlated modeling assumptions, which the current manuscript does not yet establish convincingly.","major_comments":[{"comment":"The exponential-growth conclusion for training energy and impacts (Figures 6 and 7) is based on only 205 of the 825 systems in the Epoch AI database (112 with GPU-h1, 93 with GPU-h2). This is not a random sample: inclusion requires documented hardware plus either training duration or FLOP, which is systematically more common for large, recent, well-known systems. Section 7 acknowledges that the database is incomplete and that 'excluded models could have an impact,' but the paper provides no test of selection bias. To make the sector-level claim load-bearing, the authors should compare observable characteristics of included versus excluded systems (e.g., release date, reported FLOP) and provide a sensitivity analysis in which missing GPU-hours are imputed under alternative assumptions (e.g., using the fitted GPU-h1~GPU-h2 relation, or assuming excluded systems follow a slower growth path). Without such an analysis, the fitted WLS slope may overstate the sector-wide trend.","section":"Section 5.1 and Table 1"},{"comment":"The modeling assumption of a constant PUE of 1.1 for all years is time-correlated with the trend. The paper itself cites Masanet et al. (2020) and states in Section 7 that 'we supposed that all models were trained in hyper-scale data-centers, leading to using a quasi optimal PUE for all models, masking the increase in PUE over the last decade.' Because older models trained in less efficient datacenters are assigned too low an infrastructure multiplier, their energy is underestimated more than recent models, which can artificially steepen the fitted exponential slope. The authors should report a sensitivity analysis with time-varying PUE (for example, linearly decreasing from 1.8 in 2012 to 1.1 in 2020) and show that the exponential trend and the conclusion that location shifting cannot curb growth remain unchanged.","section":"Section 3.2 and Section 7"},{"comment":"The calibration model log(GPU-h1) = 1.31 + 1.00 * log(GPU-h2) is built after excluding 19 anomalies from 106 observations and is then applied to 93 models without direct duration data. The text reports R²=0.98, but the exclusion criteria are not described in enough detail to assess whether the retained 87 observations are representative, and the paper does not report how sensitive the fitted training-energy trend is to the exclusion list or to the constant performance ratio (exp(1.31) ≈ 3.7). The authors should provide the scatterplot (currently only described in words), the list of excluded anomalies, and a sensitivity check in which GPU-h2 is used without correction or with the upper/lower bounds of the regression confidence interval. This would show whether the exponential slope in Figures 6–7 is robust to the calibration choice.","section":"Section 3.2, GPU-h2 calibration"}],"minor_comments":[{"comment":"There are typos: 'reroducibility' should be 'reproducibility', and 'Cummulative' in Section 4.1 should be 'Cumulative'.","section":"Abstract and Introduction"},{"comment":"The sentence 'we choose a close to optimal PUE of 1.1 and a hardware utilization of 50' appears to be missing a unit or percent sign; it should read '50%.'","section":"Section 3.2"},{"comment":"The description of the GPU-h1/GPU-h2 comparison would benefit from separating the model diagnostics from the interpretation: the text moves from the regression equation directly to the statement 'This model correspond to using a constant performance ratio of ≃ 27%,' which is not immediately obvious from the reported coefficients (exp(1.31) ≈ 3.7, i.e., GPU-h1 is about 3.7 times GPU-h2).","section":"Section 3.2"},{"comment":"The scenario analysis multiplies the carbon intensity by (1 - ratio)^n starting from 2019 for all models, including those released before 2019. This means models released before 2019 have their carbon intensity increased (since n is negative), which may visually affect the left side of Figure 8 and the comparison of real versus simulated impacts. The authors should clarify this choice or restrict the counterfactual to models released after 2019.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important question and the authors have taken care to make their data and code available. However, the central exponential-trend claim rests on a non-random subset of the Epoch AI database and on time-correlated modeling assumptions (constant PUE, 100% power draw, no memory). The Section 7 limitations paragraph candidly acknowledges the database incompleteness, but the manuscript does not quantify how this affects the fitted slope. I recommend major revision: the authors should add a selection-bias sensitivity analysis and a time-varying PUE sensitivity analysis, and they should make the GPU-h2 calibration exclusion list explicit. If these analyses confirm the slope, the paper would be a solid contribution; if the slope is sensitive to these choices, the conclusions would need to be substantially softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a competent, honest sector-level LCA study. The headline—AI training impacts grow exponentially and carbon-intensity reductions don't bend the curve—is not new; exponential compute growth is established. The novel package is the combination of a decade of GPU production LCA with Epoch AI training data, including embodied impacts and counterfactual decarbonization scenarios. That combination is worth having, and the code/data sharing is a real plus. The split between ADP (dominated by hardware production) and GWP (dominated by energy) is a useful nuance, and the rebound-effect framing is fair rather than alarmist.\n\nThe main soft spot is selection. The analysis uses 205 of 825 Epoch AI systems, and inclusion requires either direct duration reporting or hardware plus FLOP, which is more common for large, recent, well-known models. The paper's own Section 7 says \"the Epoch AI database is incomplete... excluded models could have an impact,\" but no selection-bias test is offered. That could overstate the slope. Also, the GPU-hours calibration excludes 19 anomalies post hoc, and the WLS trends are presented without confidence intervals, so we can't judge the precision of the slopes. Less damaging: the constant ~3.7x GPU-h2 correction changes levels, not the log-linear slope, so it doesn't threaten the exponential conclusion. The modeling assumptions (100% power draw, PUE 1.1, no memory) are acknowledged and mostly push in known directions.\n\nThat said, the qualitative conclusion is likely robust. Exponential compute growth is independently established by prior work, and the embodied-impact trend is clear from hardware specs. The selection issue could change the slope but probably not the direction. The reader's conditional verdict is about right: credible but needing robustness checks before the strong policy claim is taken as definitive.\n\nWho is this for? The AI sustainability and Green AI community, plus policymakers looking for sector-level evidence. It's a solid empirical contribution, not a paradigm shift. It deserves a serious referee—send it to review with a request for a selection-bias analysis, sensitivity checks on the subset, and confidence intervals on the trends.","headline":"A useful sector-level LCA study with a known headline: the exponential growth story is likely right, but the trend estimate is shakier than the abstract suggests because of an untested non-random sample.","tokens_in":12305,"tokens_out":1913,"would_cite":true,"duration_ms":20569,"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":"AI training's environmental footprint grew exponentially from 2012 to 2024, and simulated green-energy strategies did not flatten the curve.","keywords":["AI environmental impact","machine learning training carbon footprint","life cycle assessment","rebound effect","impact shifting","graphics cards production","metallic resource depletion","energy efficiency trends"],"falsifier":"A complete accounting of all notable AI systems released by 2024 that shows training energy or carbon footprint per model plateauing after 2019 would disprove the exponential claim, as would datacenter-level measurements showing AI training electricity grew linearly while model counts grew exponentially.","tokens_in":11367,"feed_emoji":"🌍","tokens_out":8967,"duration_ms":80594,"temperature":0.7,"pith_summary":"Machine learning's energy use and environmental impact have been growing exponentially, and this paper argues that current green strategies cannot stop that growth. The authors quantify how graphics-card production and model training evolved from 2013 to 2024, measuring carbon footprint and metallic-resource depletion. They show that even a simulated annual 25% drop in electricity carbon intensity leaves the exponential trend in training emissions intact. They conclude that the AI sector exhibits a rebound effect and that reducing AI's environmental impact requires limiting AI activity, not just improving efficiency.","feed_headline":"Greener grids can't stop AI training's exponential footprint","feed_subtitle":"Even cutting carbon intensity 25% a year, training impacts kept climbing; efficiency alone won't flatten the curve.","key_machinery":"The analysis is built on two life-cycle metrics: Global Warming Potential (GWP, in $\\mathrm{kgCO_2\\,eq}$) and Abiotic Depletion Potential (ADP, in $\\mathrm{kgSb\\,eq}$), computed with a machine-learning life-cycle assessment model. The hardware side uses a curated set of 167 workstation graphics-card models plus other training accelerators, tracking die area, technological node, memory size, and thermal design power to estimate per-card production impacts. The training side uses a database of notable AI systems; for systems without a reported training duration, the authors convert FLOP counts to GPU-hours using a linear regression calibrated on the systems that report both, then multiply by electricity mixes of the producing countries. The interpretive mechanism is the rebound effect: efficiency gains in compute and hardware do not lower total impacts because they enable larger models and more frequent hardware replacement, which shifts impacts into manufacturing. Scenarios that reduce carbon intensity by up to 25% per year test whether location-shifting and grid decarbonization can outpace this growth.","core_discovery":"Using a curated dataset of graphics cards released between 2013 and 2023 and a public database of notable AI systems, the authors estimate the production-phase impacts of the hardware and the full training-phase impacts of each system in terms of Global Warming Potential (GWP, in $\\mathrm{kgCO_2\\,eq}$) and Abiotic Depletion Potential (ADP, in $\\mathrm{kgSb\\,eq}$). Their central discovery is that both metrics grew exponentially for model training over 2012–2024, and that this growth persists under a simulated 25%-per-year reduction in the carbon intensity of electricity from 2019 onward. They also find that the production impact of individual graphics cards rose continuously, and that hardware production accounts for essentially all ADP and a meaningful share of GWP, so efficiency gains in energy use do not reduce the full environmental burden. The paper interprets the gap between efficiency gains and total impacts as a rebound effect: cheaper and more efficient compute encourages larger models, and more frequent hardware replacement shifts impacts into manufacturing.","pith_inferences":["A direct sensitivity test is suggested by the paper's own data coverage: reconstructing training impacts for the roughly three-quarters of notable systems without complete hardware records would show whether selection bias inflates the exponential trend.","Extending the same accounting to inference workloads and fine-tuning runs, which the paper excludes, would likely show a steeper sector-wide curve because inference demand is growing faster than training in many deployments.","Extrapolating the fitted trend would yield a crossover date beyond which even a 25%-per-year decarbonization rate no longer offsets the median model's emissions; the paper leaves that date unstated.","Applying a year-by-year datacenter-efficiency correction would probably steepen the early trend, since the paper assumes near-optimal infrastructure efficiency for all years."],"forward_implications":["Cutting the carbon intensity of electricity by up to 25% per year from 2019 does not flatten the exponential rise in training emissions; location-shifting alone cannot be the sector's solution.","Metallic-resource depletion from training is almost entirely tied to hardware production, so decarbonizing electricity leaves a large share of environmental impact untouched.","Hardware production impacts themselves are rising because cards have larger dies, finer nodes, and more memory, so frequent hardware refresh shifts emissions from use to manufacturing.","If the rebound effect visible in training extends to inference, then more efficient chips and models will tend to increase, not decrease, total AI energy demand.","Because real electricity mixes bottom out near 15–20 gCO2eq/kWh, even a fast global grid decarbonization has a hard physical ceiling against exponential energy growth."],"supporting_citations":[{"why":"Supplies the database of notable AI systems (compute, hardware, duration) from which every training-run impact estimate is drawn.","marker":"Epoch AI 2022"},{"why":"Provides the MLCA life-cycle assessment method used to convert hardware specs and energy use into GWP and ADP impacts.","marker":"Morand, Névéol, and Ligozat 2024"},{"why":"The industry projection this paper tests: that efficiency gains and grid decarbonization would plateau then shrink training carbon footprints.","marker":"Patterson et al. 2022"},{"why":"Documents the exponential growth of training compute, the trend whose environmental consequences the paper quantifies.","marker":"Sevilla et al. 2022"},{"why":"Supplies the digital-rebound concept used to interpret why efficiency gains did not reduce total training impacts.","marker":"Coroama and Mattern 2019"},{"why":"The first high-profile estimate of NLP training carbon emissions and the anchor date (2019) for the decarbonization scenarios.","marker":"Strubell, Ganesh, and McCallum 2019"},{"why":"Recalibrates datacenter energy-use growth and supports the assumed PUE and hardware-utilization values.","marker":"Masanet et al. 2020"}],"fun_headline_variants":["AI training impact keeps climbing despite greener grids","Efficiency alone can't flatten AI's exponential footprint","Rebound effect dooms AI efficiency gains","Greener power can't curb AI training's rising impact","Manufacturing shift hides AI's true environmental cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exponential trend is measured on a subset of the notable-AI-system database that has both hardware and compute information, and the paper does not test whether the missing systems would show the same growth.","fun_headline_variants_meta":{"raw":{"variants":["AI training impact keeps climbing despite greener grids","Efficiency alone can't flatten AI's exponential footprint","Rebound effect dooms AI efficiency gains","Greener power can't curb AI training's rising impact","Manufacturing shift hides AI's true environmental cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000327,"raw_usage":{"total_tokens":1860,"prompt_tokens":1010,"completion_tokens":850,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":778}},"tokens_in":626,"tokens_out":850,"duration_ms":7045,"temperature":1.0,"reasoning_tokens":778,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:31:07.336479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A complete accounting of all notable AI systems released by 2024 that shows training energy or carbon footprint per model plateauing after 2019 would disprove the exponential claim, as would datacenter-level measurements showing AI training electricity grew linearly while model counts grew exponentially.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the database of notable AI systems (compute, hardware, duration) from which every training-run impact estimate is drawn."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MLCA life-cycle assessment method used to convert hardware specs and energy use into GWP and ADP impacts."},{"cited_title":"R.; Texier, M.; and Dean, J","cited_arxiv_id":null,"evidence_quote":"The industry projection this paper tests: that efficiency gains and grid decarbonization would plateau then shrink training carbon footprints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the exponential growth of training compute, the trend whose environmental consequences the paper quantifies."},{"cited_title":"C.; and Mattern, F","cited_arxiv_id":null,"evidence_quote":"Supplies the digital-rebound concept used to interpret why efficiency gains did not reduce total training impacts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The first high-profile estimate of NLP training carbon emissions and the anchor date (2019) for the decarbonization scenarios."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recalibrates datacenter energy-use growth and supports the assumed PUE and hardware-utilization values."}],"review_version":1}