WorksComputer Science
Measured from the feed: the unsustainable growth of AI video generation
Abstract
Text-to-video generation runs at commercial scale, but its broader environmental impacts remain, to our knowledge, understudied. Providers rarely disclose video volumes or energy needs; we measured the yearly video output of two commercial platforms and built a prospective model of such services' absolute footprint from training to storage. Our projections suggest that reducing emissions depends on both limiting growth in video generation and achieving substantial efficiency gains. Under continued growth, emissions rise nearly seventy-five-fold by 2030 despite technological improvements. When usage remains at 2026 levels, the same improvements reduce emissions by about two-fifths. These estimates are a conservative lower bound: they cover one AI service among many; other services and uses would push the total higher.
The rapid growth of artificial intelligence raises concerns about the environmental footprint of the infrastructure that sustains it: this footprint is increasingly documented in aggregate, but its breakdown by usage type remains invisible. AI services run in data centres whose electricity demand is rising fast, whose cooling consumes large quantities of water, and whose dedicated hardware is itself a growing manufacturing burden: the carbon footprint of producing one graphics card has roughly quadrupled over the last decade, to about 300 kg eq [1]. Within this growth, one modality stands out for its resource cost: video generation. Delavande et al. found that generating a short video from text consumes roughly 90 Wh: nearly 30 the energy of image generation and 45,000 that of text classification [2].
Video generation services do not run on the user’s device: each request runs on racks of AI accelerators inside dedicated data centres. The resulting pressure is localised: nearly two-thirds of the new US data centres built since 2022 lie in high water-stress regions [3], where water withdrawals and thermal discharges can threaten freshwater biodiversity [4]. As AI video generation grows [5], such facilities may put an ever bigger strain on local energy grids and water resources.
Once experimental, AI video generation now serves entertainment [6, 7], personalized targeted advertisements [8], education videos [9], as well as mass propaganda and misinformation [10, 11]. In early 2026, the leading platform Kling AI surpassed 12 million monthly active users [12]. Professional production is following: Netflix reported in July 2026 that roughly 300 of its titles use generative-AI workflows [13], and Runway reports enterprise adoption across hundreds of companies [14].
While current studies estimate the energy consumption and carbon footprint of generating individual videos [15, 2, 16], these per-unit figures have limits. First, they are intensity indicators: video streaming shows that increasing the resolution increases the total energy consumption while lowering the energy per gigabyte. The efficiency indicator can thus improve while the service’s absolute impact increases [17]. Second, they describe a single generation, not a deployed service: they say nothing about the number of videos a service produces, nor about the infrastructure that number requires. Third, they cover few impact categories: a per-video energy figure says nothing about the cooling water or the hardware production these services require. The broader environmental impacts of AI video generation as a technology, i.e., its absolute footprint at worldwide scale and its evolution over time, are, to our knowledge, still understudied.
This study pursues two goals: estimating the direct environmental impacts of AI video generation as a deployed technology, at worldwide scale and for the present day, and exploring how these impacts could evolve to 2030 under contrasted scenarios of demand and video quality.
To make the problem tractable, we limit the scope to specialized text-to-video (T2V) generation services: platforms that produce a complete video clip from a text prompt. We therefore exclude other applications of generative AI to video, such as editing or enhancing existing footage [18, 19], avatar-based talking-head videos [20], and video generation embedded in a broader creative suite [21]. Impacts beyond the AI provider, such as video streaming, are also excluded. Since the effects we exclude would only add to the computing-related impacts of this service, our estimates form a lower bound for those impacts.
We build a bottom-up model of the serving infrastructure: starting from the energy and latency of a single generation, we scale these to the yearly volume of generated videos to size the required hardware, and from that hardware we derive the operational electricity use and the environmental impacts. We anchor this volume in a measurement campaign: we crawled Sora’s public feed for 54 days and combined it with HeyGen’s public activity counter, which gives a measured floor for the total number of videos generated per year. To account for Corresponding author: Tristan Coignion, [email protected] uncertainties, we quantify the spread of our estimates with a Monte Carlo analysis, and report which assumptions drive the results.
Here we provide, to our knowledge, the first estimate of the worldwide direct footprint of AI video generation as a deployed technology. We combine our two-platform measurements with the generation counts disclosed by six providers into a 2026 volume of 6.2 to 9.5 billion single-output videos per year. The model reports four indicators together: operational electricity, greenhouse gas emissions, operational water, and embodied abiotic resource depletion. In our median 2026 estimate, video generation consumes 724 GWh of electricity and emits 124.3 kt COeq, a footprint dominated by inference rather than by training. Under our Reference scenario, emissions grow nearly seventy-five-fold by 2030; freezing demand and video quality would cut them by about two-fifths. Holding emissions flat would instead require improving the inference stack by about 75% per year; the fastest rate we test, 30% per year, leaves 2030 emissions more than twenty-five times above their 2026 level, so efficiency gains alone cannot offset demand growth.
Results
How many AI videos are generated per year?
To size the annual volume of generated videos, we combine our own measurements on two platforms with the figures disclosed by some of the largest providers (§ 4-C). Our own measurements cover two services. Sora, OpenAI’s text-to-video model, maintained a public feed that listed every public video with its resolution, duration and prompt; HeyGen, whose text-to-video service displayed a real-time counter of generated videos on its homepage.
We crawled Sora’s public feed for 54 days between February and April 2026, recording 3.3M public videos, which annualizes to 22M per year; OpenAI discontinued Sora’s web and app experiences shortly after the crawl ended [22]. Likewise, from a 90-day monitoring of HeyGen’s counter, we deduced a rate of 111M videos per year, including private ones. Together they give a measured floor of roughly 133M videos per year. To our knowledge, six other providers publish generation counts, whose annualised rates sum to an observable total of about 6.2 billion single-output videos per year: Kling, about 360M per year [5]; Veo in Flow, about 660M [23]; MiniMax, about 430M; Vidu, about 267M; Higgsfield, about 560M; and Grok Imagine, about 3.8B at single-output equivalents, which alone accounts for most of the total volumetry in our study. Both lines of evidence are partial, so our model draws its 2026 volume from a one-sided triangular distribution spanning 6.2–9.5 billion videos, with its low end and mode anchored on 6.2B.
Brief overview of our model
Service under study. We study one year of operation of an AI video-generation service and estimate the infrastructure and energy it requires. The model covers three accounting components (training, inference, and storage) and excludes the diffusion of videos to viewers (content delivery, playback, social-media storage). The precise boundary, including the audio exclusion and the operational/embodied split, is given in the Methods (§ 4-A). The model runs over the years 2026–2030 and reports per year the required GPU fleet, the stored data, operational electricity use (consumption and peak power), and three environmental impacts: greenhouse-gas emissions, operational water use (facility water input per unit of IT energy, from a data-centre WUE covering direct potable and non-potable water and excluding indirect water use), and embodied abiotic depletion potential (ADP, kg Sb-eq).
Within this service, we represent all providers’ output with one reference generation model, treated as a black box. One inference is then one video produced end-to-end by one GPU, with durations and resolutions representative of the service’s output (Supplementary Methods S2.2).
Training covers the sector’s yearly retraining runs: the 20 to 40 runs per year represent every laboratory that trains AI video-generation models, at one to two trainings per laboratory (Supplementary Methods S2.6), plus a development overhead for preliminary and failed runs.
Storage covers the videos that providers keep on their drives over a multi-year retention window, in replicated copies. The Methods give the full boundary and the equipment model (§ 4-A and III-D).
Uncertainty. Every numeric input can carry an uncertainty distribution; we run 10,000 Monte Carlo simulations and report medians with p5–p95 ranges, and a one-at-a-time sensitivity analysis ranks the inputs by influence.
Figure 1. Annual volume of AI-generated videos. The shaded band is the model’s 2026 prior on video volume (one-sided triangular, 6.2–9.5 billion, low end and mode 6.2 billion). Provider rows are grouped in two tiers: our measurements, and the disclosed observable rates.
Impacts of AI video generation in 2026
Impact analysis. Figure 2 reports our estimates for 2026, the first simulated year, together with their uncertainty. The 7.16 billion videos estimated for 2026 (the median of the Monte Carlo runs) require a fleet of 72.6k GPUs, consume 724 GWh of electricity, and emit 124.3 kt COeq. Under an equal per-capita distribution of the planetary climate boundary [24], these emissions correspond to the targeted annual emissions of about 126,000 people.
Figure 2. Estimated 2026 footprint of specialized text-to-video generation, with sensitivity. Each row is normalised to its own median so that spreads can be compared; absolute medians are printed at the right. Dot: median. Thick bar: p25–p75. Thin bar: p5–p95, over 10,000 Monte Carlo samples.
Dividing total electricity by the number of videos gives an all-in cost of about 100 Wh per video, i.e., 4.1× the 24.4 Wh of GPU energy alone that the model charges to a single generation (a fleet-weighted average over the 2026 resolution and duration mix). Idle fleet time, facility overheads, training and storage multiply the single-inference figure more than fourfold.
Distribution of the impacts. Figure 3 breaks down global-warming potential by model component and by operational versus embodied origins; operational electricity and embodied ADP, each confined to a single scope (use-phase electricity, hardware manufacture), are decomposed in the text instead. One observation stands out: inference dominates every indicator in 2026. Inference is the largest single component of the footprint: 68.7% of the summed GHG component medians, 66.5% of operational electricity, and 64.0% of embodied ADP. Training is the second-largest component: 31.3% of the summed GHG component medians (27.4% operational, 3.9% embodied), 33.4% of operational electricity, and 36.0% of embodied ADP. Storage is negligible (under 0.2% of every indicator). The weight of training reflects its representation: each retraining run carries an energy cost independent of the yearly volume, and the model draws 20–40 retraining runs per year. The 2026 balance is robust to the volume prior: inference dominates even at the 6.2-billion-video observable total that anchors the prior’s low end, and halving the retraining count lowers 2026 GHG emissions by 16.6% without changing the ranking (Supplementary Results S1.2). Inference and training are still calibrated asymmetrically: inference is anchored on a production-scale commercial model, whereas training rests on a few published runs of small open-source T2V-models scaled by an R&D overhead. These shares depend on the assumptions and evidence currently available for training and inference. The inference-dominated balance agrees with a service-level LCA of Stable Diffusion, which likewise finds global-warming and primary-energy impacts dominated by inference cost [25].
Figure 3. Operational and embodied components of the global-warming potential (GWP₁₀₀, kg CO₂eq) of specialized text-to-video generation in 2026. Medians over 10,000 Monte Carlo samples; shares are of the summed component medians, which under-state the median total (124.3 kt CO₂eq) by about 5% because medians are not additive. The storage share (0.01%) is too thin to display.
Sensitivity analysis. We vary each input alone across its sensitivity distribution with all other inputs at their medians. In Figure 4, the energy per training run is the largest single contributor in two of the five columns, required GPUs and water, with the R&D overhead and the number of retrainings close behind. The facility parameters lead where they act directly: PUE produces the largest spread in electricity, and the water-use-efficiency coefficient the second-largest in water. Greenhouse-gas emissions are the exception: there, grid carbon intensity produces 12× the spread of video volume, so where the electricity comes from matters more than how much electricity is used. ADP is the one column where a hardware parameter leads: GPU lifespan produces 4.4× the spread of video volume, so hardware renewal cadence appears as a strong lever to reduce ADP.
Figure 4. One-at-a-time sensitivity analysis of the nine main input parameters for the reference year 2026. Each input is varied alone across its sensitivity distribution while all other inputs stay at their medians; bar length is the resulting p5–p95 output spread, normalised within each column to its largest value. Bars are comparable within a column, not across columns. This ranks the contribution of each assumption to output uncertainty.
Two more patterns emerge: the conversion coefficients keep their expected columns (carbon intensity first in GHG emissions; WUE (the EU-hyperscaler average of direct potable and non-potable water use that our WUE prior is centred on [26]) second in water, behind the energy per training run; and PUE first in electricity); and the three training-related parameters appear in all five columns and lead two of them, consistent with training accounting for about a third of the 2026 footprint.
The same asymmetry appears in the 2026 distributions (Figure 2): electricity spans a factor of 2.4 between the 5th and 95th percentiles, against 17.1 for GHG emissions.
Absolute one-at-a-time spreads
Future impacts of AI video generation
We contrast three scenarios that define what may change by 2030; their full year-by-year parameters are given in the Methods (Section III-E). All scenarios share the same technology trajectory, modelling the efficiency improvements of a real inference stack: an inference-optimization factor falls by a cumulative ∼35% over 2026–2030, about 10% per year compounded. Because this inference optimization factor evolution is arbitrary, we made an inversion analysis below and a sensitivity analysis in the Supplementary Results S1.6.
GPU utilisation, the fraction of time the hardware spends generating rather than idle, rises by a cumulative ∼18%; and the GPU fleet renews from 100% H100 in 2026 towards 95% B200 by 2030, a newer and more energy-efficient chip.
The scenarios differ in how demand, video quality and training energy evolve: the Reference scenario, in which video volume, resolution and duration all grow and the energy of each training run grows by 15% a year; Usage stagnation, which freezes volume, resolution and duration at their 2026 values and holds training energy constant; and Mass adoption, which shares the technology and quality trajectories of the Reference scenario but assumes a demand path far above it (detailed below). Figure 5 and Figure 6 trace the three scenarios over the period.
Figure 5. Divergence of the three scenarios, 2026–2030. Each line is the median of one indicator, indexed to its 2026 value. Colour identifies the indicator bundle and dash style identifies the scenario. Solid: Reference scenario. Dashed: Usage stagnation. Dotted: Mass adoption.
Figure 6. Greenhouse-gas trajectories under the three scenarios, 2026–2030. Lines are medians and bands span p5–p95 over 10,000 Monte Carlo samples; all three scenarios share the same 2026 distribution.
The volume trajectory of the Reference scenario is deliberately modest: video volume doubles in each of 2027, 2028 and 2029, then grows by 60% in 2030, for a cumulative 12.8× over four years (average yearly multiple about 1.9), which stays well below the growth Google reported for its AI products, whose monthly processed tokens rose from 9.7 trillion in May 2024 to more than 3.2 quadrillion in May 2026 [27]. The scenario is therefore a lower bound on demand rather than a central estimate. It remains plausible because video serving is expensive enough to give providers a brake that text and image generation arguably lacked: OpenAI discontinued Sora in 2026 [22], showing that usage surges in this modality can contract quickly.
Impact analysis in 2030 by scenario. Under the Reference scenario, the number of videos grows 12.8× between 2026 and 2030, but electricity grows 74.1× and GHG emissions 74.0×, nearly six times as fast as demand. The gap comes from video quality rather than volume: energy per video rises from 24.4 to 214 Wh (8.8×) as resolution and duration increase, in line with the near-quadratic scaling reported for current T2V models [2, 15]. This figure illustrates two opposing forces: quality growth alone (more pixels, longer clips) would have driven energy per video up 20.9× by 2030; the modelled technology efficiency trajectory mitigates this to the observed 8.8× factor.
The same mechanism shows up in the infrastructure. Effective latency rises 6.15×, so a fleet 44.3× larger serves only 12.8× more videos. Indicators also diverge as they grow: ADP rises 44.7× against 74.0× for GHG emissions, because ADP tracks fleet size whereas carbon tracks both fleet size and per-video energy. In absolute terms, 2030 sees 53,611 GWh consumed, 8,084 MW of peak power, and 9,202 kt CO₂eq emitted — on the same per-capita distribution of the planetary
climate boundary, the annual emissions of about 9.3 million people [24].
Under Usage stagnation, with demand, resolution and duration frozen, impacts still fall by about two-fifths by 2030 (GHG ×0.60, electricity ×0.61): energy per video falls to 10.2 Wh (×0.42), but the fleet keeps about half of its 2026 size (×0.52), because the training GPUs, whose count does not depend on output volume, offset the shrinking inference fleet. By 2030 training also dominates this shrinking footprint (about half of its GHG emissions and electricity), so under frozen demand the remaining footprint is set by the training race. Storage is the only quantity that still grows in this scenario (×2.0), since retained data keeps accumulating towards the retention-window steady state even at constant output. This two-fifths reduction is the most favourable outcome in our scenarios, and it holds only while usage, resolution and duration stay frozen at their 2026 levels; the Discussion weighs how likely that condition is.
The Reference scenario is a lower bound on demand, so a third scenario, Mass adoption, tests how the 2030 impacts respond to a demand path above it: video volume doubles in 2027, grows 2.5× in 2028, 4× in 2029 and 3× in 2030, a cumulative 60× against 12.8× for the Reference scenario, with technology and quality trajectories shared, so energy per video is identical and every impact scales with volume. In 2030, Mass adoption reaches 429.8 billion videos, 249,640 GWh of electricity, and 42.9 Mt CO₂eq of emissions (5th–95th percentiles: 7.8–135.7 Mt), with a GPU fleet of 14.9 million units and a peak power of 37.5 GW (Table S2); on the per-capita distribution of the planetary climate boundary, these emissions correspond to those of about 44 million people [24]. We read the scenario as an illustrative sensitivity case rather than a forecast: whether such accelerating rates transfer to video depends on forces the model does not represent, such as supply constraints, pricing, and provider strategy, which the recent Sora discontinuation illustrates [22].
| Unit | Reference scenario | Usage stagnation | Mass adoption | |||||
| Number of videos | billion videos | 2026: 7.16 | 2030: 91.7 | Change: × 12.8 | 2030: 7.16 | Change: × 1.0 | 2030: 429.8 | Change: × 60.0 |
| GHG Emissions | kt COeq | 124.3 | 9,202 | × 74.0 | 74.9 | × 0.60 | 42,868 | × 345 |
| Operational electricity used | GWh | 724 | 53,611 | × 74.1 | 439 | × 0.61 | 249,640 | × 345 |
| Operational water consumed | ML | 343 | 25,567 | × 74.5 | 208 | × 0.61 | 119,026 | × 347 |
| Embodied ADP | kg Sb-eq | 860 | 38,393 | × 44.7 | 450 | × 0.52 | 177,899 | × 207 |
| Peak Power | MW | 130.8 | 8,084 | × 61.8 | 76.7 | × 0.59 | 37,538 | × 287 |
| GPUs Required | k GPUs | 72.6 | 3,218 | × 44.3 | 37.6 | × 0.52 | 14,895 | × 205 |
| Stored Data (AI provider) | PiB | 48.4 | 3,295 | × 68.1 | 96.6 | × 2.0 | 12,520 | × 259 |
| Average Video Size | MiB/video | 2.42 | 8.33 | × 3.45 | 2.42 | × 1.0 | 8.33 | × 3.45 |
| Energy per Video | Wh/video | 24.4 | 214 | × 8.8 | 10.2 | × 0.42 | 214 | × 8.8 |
| Effective Latency | s/video | 125 | 771 | × 6.15 | 36.8 | × 0.294 | 771 | × 6.15 |
Table S2. Demand, environmental, and infrastructure quantities of text-to-video generation in 2026 and in 2030 under the three scenarios.
Required optimization rate (inversion analysis). We finally invert the model: under the Reference scenario, even a sustained 30% efficiency improvement per year, the fastest rate we test, leaves 2030 GHG emissions 27.5× above their 2026 level. Holding emissions flat at their 2026 level under that scenario’s demand would instead require improving the inference stack by about 75% per year compounded, about seven times the roughly 11% per year our technology trajectory models. Holding emissions flat would drive the optimization factor from 0.279 in 2026 down to about 0.0009 by 2030, below the best efficiency measured for a production model (Veo 3, at 0.084 [15]) in every year from 2027, so no software improvement of known magnitude can hold emissions flat. Software optimization alone therefore cannot hold the footprint steady; the remaining levers act on demand or on training. In short, at every optimization rate we test, efficiency gains alone leave the 2030 footprint far above its 2026 level, so within our scenarios the trajectory bends only when demand or video quality stop growing.
Discussion
Our scenarios show that the trajectory of text-to-video generation over the next five years is set jointly by video volume, quality, and hardware efficiency: under the Reference scenario, GHG emissions grow almost six times as fast as the number of videos (74.0× against 12.8×); under Usage stagnation, the same technological levers, with demand held fixed, trim the footprint by about two-fifths instead; and under Mass adoption, with demand above the Reference path, the 2030 impacts scale with volume. In this comparison, demand growth is the driver and quality growth the mechanism: quality sets how fast the energy of one video rises, and demand decides how many videos carry that energy.
Demand outweighs efficiency
The 2030 outcomes under the Reference scenario and Usage stagnation are more than a hundredfold apart in GHG emissions. Within the uncertainty we represent, the industry’s trajectory still separates the outcomes, though by a margin comparable to the spread of the model’s own uncertain inputs.
Neither trajectory is neutral against climate goals. The digital sector’s own target (Recommendation ITU-T L.1470, aligned with the Paris Agreement [28]) calls for cutting emissions by 45% between 2020 and 2030. Our Reference scenario runs in the opposite direction, multiplying operational electricity and GHG emissions nearly seventy-five-fold within four years. Sector-level studies reach the same direction: the Shift Project finds that even with grid decarbonisation, data-centre emissions keep growing at about 9% per year, and it derives absolute caps on data-centre electricity use [29]. The model already includes the main technical levers the sector proposes (fleet renewal and inference-stack optimisation) and growth still multiplies impacts, so the gap with goal trajectories follows from the growth scenario itself rather than from our choices.
Usage stagnation shows the right direction of travel: freezing volume and video quality while hardware improves would cut the footprint by about two-fifths, a direct reduction of the sector’s total impacts. Whether such a freeze is achievable is an economic question: across semiconductors, mobile access and data centres, efficiency improvements have been absorbed by rising technology affluence [30], and at the global scale by growing demand [31, 32]. The comparison that matters is therefore between efficiency gains and demand growth. Under the Reference scenario, holding emissions flat would require an optimization factor below the best known for a production model. This suggests that software optimization alone cannot hold emissions flat under this scenario.
The Mass adoption scenario makes the demand side of this comparison concrete. Its volume path grows between two and four times per year, within the expansion that large providers reported for text- and image-generation services over 2024–2026 [27], yet it multiplies the 2030 footprint nearly fivefold beyond the Reference path; because the two scenarios share their technology, quality and training trajectories, that additional footprint traces to demand alone. As a sensitivity case, the scenario isolates demand by design: technology, quality and training are held identical, so the sustainability question turns on demand as much as on technology efficiency.
Growth shifts burden to hardware
Longitudinal estimates show machine-learning training impacts growing exponentially through 2025 despite major hardware and algorithmic gains, and even under simulated carbon-aware training, a rebound attributed to ever larger models [33], while the embodied footprint of producing AI GPUs keeps rising steeply in parallel [1]. Efficiency on an expanding infrastructure therefore does not remove the burden; it shifts it toward the embodied cost of producing new hardware and toward rebound-driven growth in demand.
Under Mass adoption, video volume grows by a cumulative over 2026–2030 and the GPU fleet grows with it, so the material footprint of manufacturing that hardware would reach levels far beyond those observed today; supply constraints of that kind are not represented in the model. Supply-chain evidence indicates where such constraints would bite first: in 2025, the four largest AI chip designers (NVIDIA, Google, AMD, and Amazon) consumed about 90% of global advanced-packaging capacity and of high-bandwidth-memory supply, but only about 12% of advanced logic-die production [34].
Because those four designers already take up nearly all packaging and memory supply, sustaining hardware growth on the scale of the Mass adoption path would rest on expanding capacity that AI demand already dominates, and that expansion would proceed through new manufacturing facilities on multi-year lead times; advanced logic, where AI demand held a smaller share in 2025, looks like a softer constraint so far, though the industry’s move to 3 nm processes is likely to raise that share considerably.
Estimates are lower bounds
Sizing the trajectory the industry follows is difficult because usage is opaque: to our knowledge, only a handful of providers have published rough generation counts, our own measurements cover two more platforms, and every other line of evidence is disclosed at the provider’s convenience, over ill-matched time windows. Disclosure of inference conditions (data centres used by each generation, power-usage effectiveness, and the hardware fleet) would let future estimates replace these triangular priors with measured, location-weighted values. It would also make generation location visible, which grid operators and regulators need.
Our estimates should be read as a lower bound, for several reasons. The model holds the generation architecture fixed across 2026–2030, while real models are likely to grow more capable and more expensive to run over the same period. The scope excludes the diffusion of videos to viewers (content delivery, playback, and social-media storage), and audio generation remains excluded. Text-to-video is also only one usage among many; our results describe it in isolation and should not be read as a proxy for the footprint of AI as a whole.
Training is the second-largest component of the 2026 footprint: 31% of the summed GHG component medians and about one-third of operational electricity. That share, however, reflects what is publicly known rather than what training actually costs: it follows from the assumed retraining cadence (20–40 runs per year in our prior) and the assumed per-run energy, not from a measurement. No provider we study discloses the compute or energy behind its closed-source model, so our training-cost model is calibrated on a few published runs of smaller, open-source text-to-video models, scaled by a development overhead. Spending data point in the same direction: at three AI companies where the split is published or estimated, compute for research and development alone takes 40–52% of total operating expenses [35], which is consistent with training-related compute being larger than documented runs suggest.
Other simplifications shape how the results should be read. The GPU fleet is treated as dedicated to inference, with no sharing across services; the generation model is a black box with no GPU-memory accounting; one GPU serves one inference at a time, with no credit for batching (per-video energy varies within only 3.3% across batch sizes 1 to 4 [15]); and server overhead is applied uniformly across GPU generations despite being measured on a single reference architecture.
We deliberately excluded several adjacent effects from the model. Following Kaack et al., our estimates cover the computing-related impacts of machine learning, quantified operationally and at the embodied level for training, inference, and storage, while the immediate impacts of the applications it enables and the system-level impacts of their wider use lie outside our boundary [36]. Rebound and substitution effects belong to those wider, indirect categories.
Effects outside the boundary
A more efficient pipeline could lower the per-video impacts we compute, but also lower the barrier to generating and drive total demand up faster than efficiency brings it down, a rebound our scenarios do not capture [32]. Provider strategies point the same way: efficiency is deployed to enable larger activity volumes, an “intended enablement” documented for 5G and model training rather than measured for video generation [37]. Likewise, AI video generation that takes the place of traditional filming displaces burdens that sit outside our boundary (studios and sets, location generators, travel by cast, crew and equipment), about 2,840 t e for an average large-budget film production [38]. Replacement, however, is not the only outcome: generative clips are often added to footage shot conventionally rather than replacing it, as when Netflix reported hundreds of titles using generative-AI workflows in post-production [13]; the traditional burdens then persist, and the AI burden comes on top instead of displacing them. We do not evaluate that balance, and substitution carries a social counterpart (creative-sector employment, already a contested point in professional animation production [39]) that we equally leave unquantified.
Economy-wide modelling reinforces this caution: in a global energy–economy model, AI productivity gains raised rather than lowered fossil-fuel emissions [31]. Read against this record, the Reference scenario and Mass adoption are the less conditional of the three: the roughly two-fifths reduction under Usage stagnation holds only while usage stays frozen, a condition the economy-wide evidence does not support. We leave the direction and size of these effects, and of any regulatory lever able to shape them, to future work grounded in usage data that does not yet exist.
Methods
In this section, we present how we estimate the yearly resource needs and environmental impacts of a text-to-video generation service. The service represents the combined output of existing providers, rather than one user’s individual generation. The model follows the service from model training to inference and storage. Detailed equations, parameters, sources, and implementation rules appear in the Supplementary.
Scope and study boundaries
The scope of our study is the deployment of a text-to-video generation service. It includes training and retraining, inference, and the storage of generated videos. Training and retraining cover the sector’s model-development activity: each laboratory that trains AI video-generation models contributes one to two trainings per year, so the model draws 20 to 40 retraining runs per year in total (Supplementary Methods S2.6). It excludes content delivery, video playback, and storage on social-media platforms. It also excludes audio generation, which a small share of newer models performs alongside the video. We report operational electricity and operational water as operational impacts. The water metric is the facility water input per unit of IT energy, from a data-centre WUE: direct potable and non-potable water, with indirect water excluded. We report greenhouse-gas emissions as operational and embodied impacts. We report the embodied abiotic depletion potential (ADP) for the embodied phases only as an embodied impact only.
Model overview
Figure 7 shows the model’s structure for one simulated year. The model runs year by year: each year, it combines yearly inputs into per-video latencies, fleet and energy quantities, and yearly impacts, and passes the hardware stock to the following year. Each step is detailed below.

Figure 7. Model pipeline for one simulated year . Solid arrows carry energy and volume quantities, dashed arrows carry the hardware stock behind the embodied impacts, and the dotted arrow shows the year-to-year recursion.
The model first estimates per-video latency from video quality, duration, frame rate, and GPU model, following the compute-bound DiT latency model of Delavande et al. [2]. It weights these latencies by the annual resolution and duration mix and the GPU fleet mix. It then converts the resulting workload into active and idle GPU-hours, energy, and required GPUs. Training adds a per-run energy term and a development overhead. Storage adds the retained data volume, replication, and storage energy. The model applies server, IT-equipment, and data-centre overheads before it derives facility electricity, water, and operational emissions. It amortises embodied GPU, storage, and training impacts over the relevant hardware lifetime. The model passes the fleet and stored data from year to year .
The latency model is fitted on measurements of WAN2.1-T2V-1.3B on a single H100, and we transfer it to other GPUs by their peak-throughput ratio (Supplementary Methods). The fitted latency therefore describes an unoptimized research configuration, not a deployed endpoint. Deployed services use kernel optimization, distillation, quantization, caching, and similar techniques to accelerate generation, so we apply one inference-optimization factor that reduces the fitted latency and the resulting energy. We calibrate against the per-clip energy that Jegham et al. [15] report for Seedance 1.5, on our reference workload of one 8-second, 720p clip on a single H100; this yields for the base year.
Annual volume
We estimate annual video volume from two platform measurements and provider disclosures. We measure public Sora output from its public feed and HeyGen’s text-to-video output from its homepage counter. Provider disclosures provide an independent scale check for services outside these two platforms. The Supplementary give the crawl windows, counter processing, disclosure treatment, and utilisation calculation.
The model uses the Sora resolution and duration distributions to represent a generic short-clip service in 2026.
Parameters and data sources
The principal inputs of the model are annual video volume, video quality, GPU fleet composition, GPU utilisation, inference optimisation, training energy, storage retention, replication, PUE, WUE, and grid carbon intensity. We model H100 and B200 GPUs in an 8-GPU node configuration. Each GPU model is characterised by its rated power (TDP), its measured idle draw, and its embodied impacts, which include an amortised share of the host server. We report embodied resource use as the abiotic depletion potential (ADP), which aggregates the mineral and the fossil-resource components of the underlying hardware datasets. The model amortises these embodied impacts over the GPU lifespan as the fleet renews year by year. Storage hardware is modelled from the stored volume: a per-video file size fitted on our Sora scrape, accumulated over a rolling retention window and multiplied by a replication factor, then charged with an operational energy per unit of stored data and embodied impacts per unit of storage hardware. The Supplementary list all input and output variables, their units, distributions, sources, and fixed values. The replication package contains the source data and scenario definitions.
Uncertainty and scenarios
We represent uncertain numeric inputs with probability distributions and draw one coherent sample for each field in each Monte Carlo run. The final analysis uses 10,000 samples and seed 42 for the three scenarios. We also run a one-at-a-time sensitivity analysis with the same seed and sample count.
All scenarios share the same technology trajectory. The Reference scenario increases video volume, resolution, and duration. The Usage stagnation scenario freezes volume and video quality at their 2026 values. The Mass adoption scenario applies a faster volume path (yearly multiples of 2, 2.5, 4, and 3) with the same quality and technology trajectories. All scenarios renew the inference fleet and improve utilisation and inference optimisation.
Three assumptions limit the model. It treats generation models as black boxes without explicit GPU-memory requirements. It assigns one inference to one GPU and gives no batching credit. It represents training energy and training GPU needs without an explicit training purchase schedule or bursty concurrency. These choices make the estimates a lower bound for real deployments of this service.
Limitations
The estimates are a lower bound for the specialized text-to-video ecosystem. The scope excludes audio generation, the diffusion of videos to viewers, and generation usages outside specialized T2V services, and it holds the generation model’s size and inherent energy cost constant over the simulated years. Each of these scope decisions removes real demand or future growth from the estimate, so each one pushes the results down rather than up.
Several inputs describe the market rather than measure it. The resolution and duration mixes, and the storage file-size model, come from a single platform (Sora) and may not represent the clips generated worldwide. The fleet is restricted to two GPU models, H100 and B200, in an 8-GPU node configuration. Facility overheads, water use, and grid carbon intensity are drawn from European distributions and are not weighted by where generations actually run; for carbon intensity, the location of most generation remains an open question.
Training is the least measured part of the model. It is calibrated on a few published runs of small open-source models scaled by a development overhead, because no provider we study discloses the compute behind its closed-source model. The model also spreads each year’s training GPU-hours uniformly across the year; this keeps the provisioned fleet busy at all times and needs fewer GPUs, at lower peak power, than a real campaign that concentrates the same work into weeks. Finally, the generation model’s size and training time do not grow over the years, so the increase of the annual training cost is artificially low. All of these choices push the training estimates down.
Three further calibrations rest on a single reported data point each: the inference-optimization factor, the per-run training energy, and the transfer of latency from H100 to B200 GPUs. The embodied GPU figures additionally carry at least a factor-of-two uncertainty, as two public data sources disagree by that margin. These inputs would benefit from replacement by measured distributions as more data becomes public.
Supplementary Information
The full numerical inventories appear in the Supplementary Results, and the detailed equations, parameter tables, data sources, and uncertainty procedures in the Supplementary Methods.
Acknowledgements
This research was conducted with the support of the Observatory on AI’s Environmental Footprint led by the AI and Society Institute at ENS.
Author Contributions
T.C. designed the model and wrote the code; A.B., G.G. and A.-L.L. supervised the work. All authors reviewed the manuscript.
Competing Interests
The authors declare no competing interests.
Data Availability
The model inputs, archived simulation results and crawl data are available in the replication package at https://doi.org/10.5281/zenodo.24163581.
Code Availability
The simulator code and analysis notebooks are available in the replication package at https://doi.org/10.5281/zenodo.24163581.
Supplementary Results
This part of the paper gives the numerical details behind the figures of the Results section. All intervals correspond the 5th and 95th percentiles over 10,000 Monte Carlo samples of the seeded reference run.
Full 2026 inventory
Table S1 lists all quantities reported by the model for 2026, the first year of the simulation. It complements Figure 2 The 2026 values are identical under all three scenarios, because the scenarios diverge only after 2026.
| Unit | Median [p5–p95] | |
| Demand input | ||
| Number of videos | billion videos | 7.16 [6.28–8.79] |
| Environmental outputs | ||
| GHG Emissions | kt COeq | 124.3 [23.9–409.4] |
| Operational electricity used | GWh | 724 [477–1,161] |
| Operational water consumed | ML | 343 [237–526] |
| Embodied ADP | kg Sb-eq | 860 [515–1,602] |
| Infrastructure outputs | ||
| Peak Power | MW | 130.8 [84.1–215.1] |
| GPUs Required | k GPUs | 72.6 [52.2–110.0] |
| Stored Data (AI provider) | PiB | 48.4 [42.4–59.3] |
| Per-video unit factors | ||
| Video Size | MiB/video | 2.42 |
| Energy per Video | Wh/video | 24.4 |
| Effective Latency | s/video | 125 |
Table S1. Full 2026 inventory of the demand input, the environmental and infrastructure outputs, and the per-video unit factors that the model reports for text-to-video generation.
Median and 5th–95th percentile over 10,000 Monte Carlo samples. The three per-video quantities at the bottom are deterministic functions of the fixed inputs (the resolution and duration mix, the hardware model, and the optimization factor), so they carry no sampled spread; they are reported as plain values. Energy per Video is the GPU-active term only, excluding idle draw, facility overheads, training, and storage.
The 2026-2030 scenarios
Table S2 gives the 2030 medians of the three scenarios beside their common 2026 starting point, with the change factor for each.
The retraining count is held at the prior of Supplementary Methods S2.6 in all three scenarios; as a diagnostic sensitivity of this training-constancy assumption, a stagnation variant with the count halved to in every year brings the 2030 median GHG emissions down from 74.9 to 55.2 kt COeq, a reduction of about 47% against its own 2026 level instead of about two-fifths (the variant is included in our replication package).
Comparison with video-platform upload fluxes
To put the modelled volumes into perspective, we compare them with the upload flux of established video platforms.
YouTube states that, on average, over 20 million videos are uploaded to the platform daily, as of April 2025 [40]; the same figure is displayed on its press page [41]. This corresponds roughly 7.3 billion videos per year, very close to the 2026 median of the model of 7.16 billion videos (about 20 million per day). The closeness of the two counts is, however, a poor guide to volume, because the two flows differ greatly in video length: the generated clips average about 6.5 seconds in 2026 (76% of them 5 seconds long, none longer than 20 seconds) and about 10.5 seconds in 2030, against a mean of 615 seconds (median 126 seconds) in a random sample of YouTube uploads [42]. In seconds of footage, the unit we keep for the rest of this section, the 2026 median of the model is therefore about 1% of the YouTube upload flux, the 2030 Reference-scenario median about a fifth of it, and the Mass-adoption median about equal to it. For reference, the same comparison in raw video counts gives about 98% of this flux for 2026, thirteen times it for the 2030 Reference-scenario median (about 251 million videos per day), and sixty times it for the 2030 Mass-adoption median (about 1.2 billion videos per day); counting videos rather than seconds overstates the footage share by a factor of about 95 in 2026 and about 60 in 2030.
TikTok publishes no upload volumes; its transparency disclosures cover moderation actions and user counts only. The best available figure comes from a near-complete census of the platform, which estimates 269 million posts created over one measured 24-hour window in April 2024 [43], at a mean post length of 20.7 seconds. In seconds of footage, the 2026 median of the model is about 2% of this one-day flux, the 2030 Reference-scenario median about half of it, and the Mass-adoption median about 2.2 times it. For reference, the same comparison in raw video counts gives about 7.3% of this flux for 2026, about 93% for the 2030 Reference-scenario median, and more than four times the flux for the 2030 Mass-adoption median.
Three caveats bound these comparisons. First, the platform figures are unaudited company statements or sampling-based estimates with undisclosed definitions: YouTube does not state whether private or unlisted uploads are counted, and the TikTok figure is a one-day extrapolation that includes posts later deleted or privatised (about 46% of the posts sampled in the census were no longer publicly accessible) [43]. Second, the footage ratios above inherit the uncertainty of the length estimates on both sides. On the model side, the 2026 mean comes from measured public Sora outputs, but those outputs stand in for many providers whose clip lengths may differ, and the 2030 mean follows the scenarios’ assumed drift towards longer videos. On the platform side, the means come from single samples; the YouTube mean is dominated by a minority of long uploads (the sample median is 126 seconds), so it is sensitive to sample composition, and the YouTube footage flux is dominated by long-form content (Shorts, music, long-form uploads), whereas the model counts generated short clips only. Finally, seconds of footage measure volume only: one generated second and one uploaded second are not comparable in production impact. Third, the comparison is a flux-to-flux one; for stocks, YouTube reported over 20 billion uploaded videos over its first twenty years as of April 2025 [40], and a peer-reviewed census estimated 9.9 billion publicly visible videos in late 2022 [42].
Medians over 10,000 Monte Carlo samples; the 2026 column is shared by all three scenarios. Energy per Video is the GPU-active term only.
This comparison has no consequence for the model itself: the volume prior is anchored on provider disclosures and on the measured floor, not on platform fluxes, so the two estimates are independent. Its role is to bound the reading of the results. In footage terms, the modelled volumes remain small shares of established platform fluxes in 2026: about 1% of the YouTube upload flux over a year, and about 2% of the TikTok one-day flux over a day. Even the 2030 Mass-adoption median amounts to roughly one year of YouTube uploads at the 2025 rate.
Absolute one-at-a-time spreads
Figure 4 normalises each output column to its longest bar, so it ranks the inputs but hides the size of the spread they produce. Table S3 gives the same spreads in the unit of each output.
Validation of our model against real-world data
We finally confront the model’s per-video predictions with market prices.
Our per-video cost is calibrated to the per-clip energy that Jegham et al. report for Seedance 1.5, a mid-tier commercial service: 102.1 Wh for an 8-second, 720p clip on an H100 [15]. Converting the model’s GPU-seconds into an infrastructure cost at market rates and comparing with commercial per-clip prices checks whether the throughput level is plausible: at 2026 rates for H100-class hardware (0.0006–0.0022 USD per GPU-second across on-demand, serverless, and contract offers [44], [45]), that calibration point (about 525 GPU-seconds on an H100 at the model’s measured power draw) costs 0.3–1.2 USD of raw GPU infrastructure per clip and falls inside the 0.072–1.50 USD band a cost-disclosing service charges for its 5–8-second catalogue [45].
Sensitivity to the software-efficiency trajectory
The technology trajectory assumes that the inference-optimization factor falls by a cumulative ~35% over 2026–2030, about 10% per year compounded. To test how the conclusions depend on that assumption, we rerun the Reference scenario deterministically, holding every input at its median except the optimization factor, which follows a constant annual improvement rate instead of the assumed decelerating profile. Table S4 reports the 2030 outcomes.
Across the tested range, the efficiency path moves the 2030 total by about 4× (13,740 vs 3,360 kt COeq), but no tested rate approaches the roughly 75% per year that holding 2030 emissions at their 2026 level would require.
Supplementary Methods
This part of the paper gives the detailed model definition that supports the short Methods section above. It provides the equations, parameter definitions, data sources, uncertainty procedure, and modelling boundaries. We refer to section III and Figure 7 for an overview.
Scope and impact inventory
The scope of the study is the deployment of a text-to-video generation service, sized to the combined generation volume of all existing services, rather than the generation of a single video by a single user. It covers three cost components: training (the sum of cost of trainings and re-trainings), inference (a cost proportional to the volume of generations), and storage (the cumulative cost of retaining generated videos). It excludes the diffusion of videos to viewers, i.e., content delivery, playback, social-media storage and sharing and storage on other platforms.
Audio generation is out of scope. Most current text-to-video models produce silent video; only a growing share of newer models, sometimes called text-to-video-audio models, also synthesize an audio track alongside the video [15]. We exclude this audio component for clarity and simplicity. The exclusion keeps our estimates conservative: any audio generated only adds to the impacts we report, so the figures remain a lower bound for services that generate audio as well as video.
Not every impact is estimated for every phase. Greenhouse gas emissions cover both the operational and the embodied phase. Operational electricity and operational water are operational only. Embodied abiotic depletion potential (ADP, kg Sb-eq) is embodied only. Table S5 summarises which phases each reported indicator covers.
| Indicator | Training op. | Training emb. | Inference op. | Inference emb. | Storage op. | Storage emb. |
| Operational electricity | ✓ | – | ✓ | – | ✓ | – |
| Operational water | ✓ | – | ✓ | – | ✓ | – |
| GHG (t CO₂eq) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Embodied ADP (kg Sb-eq) | – | ✓ | – | ✓ | – | ✓ |
Table S5. Lifecycle phases covered by each reported indicator (✓ included, – excluded; op. = operational, emb. = embodied).
| Required GPUs (k GPUs) | GHG emissions (kt CO₂eq) | Operational electricity (GWh) | Operational water (ML) | Embodied ADP (kg Sb-eq) | |
| Grid carbon intensity | 0 | 340.9 | 0 | 0 | 0 |
| Energy per training run | 42.0 | 66.3 | 382.5 | 184.5 | 484.5 |
| Data-centre energy overhead (PUE) | 0 | 64.2 | 417.3 | 0 | 0 |
| R&D overhead on training | 24.5 | 38.6 | 223.0 | 107.6 | 282.5 |
| Number of videos generated | 15.6 | 27.4 | 160.1 | 77.2 | 180.0 |
| Retrainings per year | 16.1 | 25.4 | 146.7 | 70.8 | 185.8 |
| GPU lifespan | 0 | 12.2 | 0 | 0 | 793.9 |
| GPU utilisation | 12.8 | 4.8 | 16.8 | 8.1 | 147.6 |
| Data-centre water intensity (WUE) | 0 | 0 | 0 | 134.4 | 0 |
Table S3. Absolute p5–p95 output spread produced by each uncertain input alone, 2026, in the unit of each output (Reference scenario; 10,000 Monte Carlo samples per input, all other inputs at their medians). Rows are sorted by decreasing GHG spread; a zero means the input does not affect that output. The video retention period produces no spread in any of the five outputs and is therefore omitted (it is also absent from Figure 4).
| Efficiency path | in 2030 | GHG in 2030 (kt CO₂eq) | GHG growth × 2026 |
| Assumed profile | 0.181 | 8,957 | 73.4 |
| None (/yr) | 0.279 | 13,740 | 112.5 |
| /yr | 0.227 | 11,207 | 91.8 |
| /yr | 0.067 | 3,360 | 27.5 |
Table S4. Sensitivity of the Reference scenario to the software-efficiency trajectory. Deterministic runs with all inputs at their medians; the inference-optimization factor either follows the assumed decelerating profile (cumulative ∼35% over 2026–2030) or shrinks at a constant annual rate . All runs share the 2026 starting point (122.1 kt CO₂eq at ), which differs from the Monte Carlo median quoted in the main text (124.3 kt) because a run with all inputs frozen at their medians does not reproduce the median of the Monte Carlo outputs.
Estimating the initial annual volume of generated videos
The initial volume is the model’s dominant input, and no provider publishes an authoritative generation count. The evolutions of are scenario assumptions (Supplementary Methods S2.7). We therefore triangulate two independent sources: our own scrapes of two platforms, and the figures that providers disclose. We combine those figures into the triangular distribution described at the end of this subsection.
Sora scrape. Sora (sora.chatgpt.com) maintained a public feed that listed every public video with its resolution, duration and prompt; OpenAI discontinued the service in April 2026. We crawled this feed over 54 days between February and April 2026 and recorded about 3.3 million videos, an average of 61,000 per day, or 22M per year. The crawl covers public videos only, so the figure is a lower bound on Sora’s output. The same corpus provides the resolution and duration mixes that weight per-video latencies (Supplementary Methods S2.4) and the power-law file-size model used for storage (Supplementary Methods S2.5); the clips are predominantly short and low-resolution, with 76% lasting five seconds and 80% generated at 480p.
HeyGen scrape. HeyGen displays a cumulative counter of generated videos on its homepage. We sampled the counter at regular intervals of about five minutes between March and June 2026, and computed the mean daily generation rate from consecutive counter readings: 304,000 videos, or about 111M per year. Unlike the Sora crawl, the counter includes private generations, but it measures a single platform. It tracks HeyGen’s text-to-video service only; the company’s avatar-based and template-driven talking-head products are not included in the count. However, that service is not tied to a single engine: the platform’s editor routes prompts both to HeyGen’s in-house models and to third-party closed models that HeyGen reaches through their vendors’ APIs, including Sora, Google Veo and Kling [46]. A share of the counted output is therefore generated outside HeyGen’s own infrastructure, so the 111M figure measures the service’s activity rather than the load on HeyGen’s fleet; it may also overlap with the Kling disclosure below if Kuaishou’s count includes API usage, an overlap bounded at 11M videos per year, about 2% of the observable total.
Counting units. The provider disclosures below do not share a common unit, so we normalise them to single-output equivalents: one video is one finished output, and multi-variation renders are divided by the number of variations. The division counts prompt-equivalents: each rendered variation consumes compute, so per-prompt-equivalent energy implicitly amortises discarded variations across the market aggregate.
Other providers and public disclosures. To our knowledge, no other provider publishes a continuing generation count, but several disclose cumulative figures or recent increments from which annual volumes can be extrapolated. Table S6 lists the disclosures we use.
| Provider | Disclosure | Annualised |
| Kling (Kuaishou) | 600M cumulative over 19 months (Dec 2025); ≈30M/mo end-2025 | ≈360M/yr |
| Veo in Flow (Google) | 275M over 5 months (2025) | ≈660M/yr |
| MiniMax (Hailuo) | >600M cumulative by Dec 2025 (≈17 months after the Aug 2024 launch) | ≈430M/yr |
| Vidu (Shengshu) | >400M cumulative (Oct 2025, since Apr 2024) | ≈267M/yr |
| Higgsfield | >800M AI-media pieces (Jan–Jun 2026); video share ≈35% (300M/850M cumulative) | ≈560M/yr |
| Grok Imagine (xAI) | 1.245B rendered clips over 30 days (Feb 2026); ≈4 variations per prompt | ≈3.8B/yr at single-output equivalents |
| PixVerse | 0.8B → 2.1B cumulative, self-reported (Sep 2025 → Mar 2026); definition unstated | ≈0.8B/yr |
Table S6. Provider volume disclosures used to build the observable total and to justify the high end of the volume prior. Cumulative totals are annualised from the most recent disclosed increment, except MiniMax, which is annualised from its lifetime total. All rows except PixVerse enter the 6.2B observable total; PixVerse only justifies its upper tail.
Kuaishou reported more than 600M videos for its Kling model over its first 19 months, about 30M per month at the end of 2025 [5]. Google reported 275M Veo videos in Flow over five months, about 55M per month [23]. MiniMax stated that its video models had generated more than 600M videos cumulatively by the end of 2025, about 17 months after the August 2024 launch of its video generation service, a lifetime average of about 35M per month [47]. Shengshu reported more than 400M Vidu videos since the April 2024 launch, about 22M per month [48]. Higgsfield, a platform that also resells third-party video models, reported more than 800M pieces of AI media generated on its platform over January–June 2026 [49]; its own cumulative split of 300M videos out of 850M generations puts the video share at about 35%, which we adopt as about 560M videos per year [50]. Because the platform resells models that the provider rows also count, this figure may overlap with the rows above, and we treat it as an upper bound on the volume Higgsfield adds on top of them. Grok Imagine (xAI) reported 1.245B rendered clips over the 30 days before 2 February 2026 [51], and close to 50M videos per day at its public all-hands meeting later that month [52]. These counts cover rendered clips, about four variations per prompt, so we divide them by four: at single-output equivalents, the disclosed 30-day rate implies about 3.8B videos per year and the all-hands rate about 4.6B.
Several providers disclose no usable volume figure. ByteDance discloses no generation counts for Seedance, and OpenAI never disclosed Sora volumes before discontinuing the service in April 2026 [22]. To our knowledge, no Chinese provider discloses a monthly rate, only cumulative totals, so the Kling and Vidu annualisations assume that each provider’s last observed rate held over the base year. Because these platforms launched between 2024 and 2025, no multi-year historical trend exists for them: the disclosed rates describe current annualised output, not a trajectory that can be extrapolated onwards.
Measured floor and prior. Our two scrapes put the directly measured output at about 133M videos per year. Adding the observable disclosures of Table S6 gives an observable total of about 6.2 billion single-output videos per year Under our normalisation, the disclosed Grok Imagine rate alone exceeds the sum of all other observable providers, so a single service accounts for most of the observable total. We anchor the prior on this total: , a one-sided distribution. The low bound is a deliberate undercount: the Sora figure counts public videos only, several large services entered the market during 2026, and the disclosed rates are 2025 snapshots.
We justify the 9.5B high end item by item rather than by hand. Four items add about 3.3B on top of the 6.2B mode. First, the Grok Imagine upside: the all-hands rate reported above would add about 0.8B single-output videos per year over the disclosed 30-day rate; because the platform reached this scale during the base year and no post-February 2026 figure exists, we keep this increment out of the mode. Second, PixVerse: its counting definition is unstated, so we count its self-reported volume at a 70% discount, which adds about 0.8B videos per year; the discount is documented in our replication package. Third, ByteDance: no primary disclosure exists for Seedance, whose video generation is embedded in Doubao, Jimeng, and CapCut, so we count it at the midpoint of a Kling-to-Grok parity range, about 1.0B videos per year; this is a stated analogy, not a disclosure. Fourth, Google’s non-Flow surfaces: we count video generation through the Gemini app, free YouTube Shorts Remix access, and the Veo API at parity with Veo in Flow’s disclosed rate, which adds about 0.7B videos per year. The itemised sum is about 9.5B. Beyond this total we cannot quantify several possible additions: Meta Vibes, Midjourney’s unlimited-tier generations, and the further long tail. The Kling rate is also a 2025 figure; Kuaishou’s reported revenue growth for Kling (annualised revenue run rate up from USD100M in March 2025 to USD240M in December 2025) suggests the underlying volume grew further, but no updated count exists [5].
Average-to-peak utilisation. The utilisation of the inference fleet is the fraction of provisioned GPU time spent actively generating. We estimate it from the same scrapes as the ratio of average to peak generation rate, where the peak is the mean of the five busiest hourly bins: Sora yields 62.1% and HeyGen 45.8%. The two scrapes give point estimates for two services only, so we do not fix at either value: the 2026 value is drawn from , a prior whose mode sits on Sora’s measured 62.1%, the higher of the two. The lower and upper bounds are set by hand to span preferably higher-utilisation services; in particular, the low bound does not cover HeyGen’s 45.8%.
Impact estimation model structure
The model estimates the yearly environmental impact of a text-to-video generation service through a pipeline of calculations involving 32 input variables and 11 output variables, with results for year feeding year through, for example, the GPU fleet composition. Its dominant path runs from the annual video volume through per-video latencies, annual active GPU-hours, and inference, training, and storage energy, to facility electricity and the three reported impacts. This section gives the equations, coefficients, and per-step definitions; Table S7 summarises the notation.
| Symbol | Meaning | Unit |
| Simulation year | year | |
| Annual generated video volume | video/year | |
| Per-video latency for resolution and duration on GPU model | s/video | |
| , | Peak BF16 dense throughput of GPU model ; denoising steps | TFLOP/s, – |
| , , | Fitted latency coefficients (quadratic, linear, constant) | – |
| , | Resolution and duration shares | – |
| GPU fleet share for model | – | |
| , , | Numbers of GPU models, video durations, and resolutions | – |
| Inference-optimization factor | – | |
| , | Inference and training GPU utilisation rate | – |
| , | Active inference and training GPU-hours | GPU h |
| Active inference GPU-hours on GPU model | GPU h | |
| , | Required inference and training GPUs | GPU |
| , , | Annual inference, training, and storage energy | MWh |
| Active inference energy, idle draw excluded | MWh | |
| Average per-video inference energy | Wh/video | |
| , , | Training runs per year, energy per training run, development overhead | –, MWh, – |
| , , | Training GPU model; nameplate and static GPU power | –, W, W |
| Training power fraction of the GPU nameplate power | – | |
| Fleet-weighted average inference GPU power | W | |
| Provisioned inference fleet of GPU model | GPU | |
| GPU lifespan | years | |
| , | Per-GPU embodied GHG and ADP coefficients | kg COeq, kg Sb-eq |
| , , PUE | Server and IT-equipment overheads; power usage effectiveness | – |
| WUE | Water Usage Effectiveness | L/kWh |
| , | Stored video data; storage electricity intensity | TiB, MWh/GiB |
| Storage replication factor | – | |
| fps | Frame rate of generated videos | frames/s |
| , | IT-equipment energy; facility electricity | MWh, GWh |
| Peak facility power | MW | |
| , , | GHG, operational water, and embodied abiotic depletion potential | t COeq, kL, kg Sb-eq |
Table S7. Core notation used in the Supplementary Methods.
Three assumptions bound what the model can represent. First, generation models are treated as black boxes characterised only by video quality and cost, so GPU memory requirements are not modelled. Second, a GPU handles exactly one inference at a time and one inference occupies exactly one GPU: we found no published evidence of batched video generation at production scale, and splitting a model across GPUs improves latency but not throughput [15]. Third, only inference GPUs form an explicit fleet with a purchase and retirement schedule; training hardware is accounted for by amortising its GPU-hours, without modelling a purchase cadence or bursty peak concurrency. Training is likewise a sequence of discrete runs and not a continuous workload: video-generation models are released as versions, each from a bounded training
| Provider | Disclosure | Annualised |
| Kling (Kuaishou) | 600M cumulative over 19 months (Dec 2025); ≈30M/mo end-2025 | ≈360M/yr |
| Veo in Flow (Google) | 275M over 5 months (2025) | ≈660M/yr |
| MiniMax (Hailuo) | >600M cumulative by Dec 2025 (≈17 months after the Aug 2024 launch) | ≈430M/yr |
| Vidu (Shengshu) | >400M cumulative (Oct 2025, since Apr 2024) | ≈267M/yr |
| Higgsfield | >800M AI-media pieces (Jan–Jun 2026); video share ≈35% (300M/850M cumulative) | ≈560M/yr |
| Grok Imagine (xAI) | 1.245B rendered clips over 30 days (Feb 2026); ≈4 variations per prompt | ≈3.8B/yr at single-output equivalents |
| PixVerse | 0.8B → 2.1B cumulative, self-reported (Sep 2025 → Mar 2026); definition unstated | ≈0.8B/yr |
campaign, and the work between releases is covered by the development-overhead multiplier of subsection S2.4.
Headline relations
Seven groups of relations fix our model’s outputs.
Per-video inference latency. The model’s primary physical quantity is the per-video generation latency , the number of GPU-seconds of model needed to generate one video of resolution and duration . Latency follows a compute-bound model for DiT text-to-video diffusion taken from Delavande et al. [2]: per-video latency is quadratic in the DiT token length, and therefore grows with resolution, duration, and frame rate. The DiT token length of a video of resolution and duration is , where and are the pixel height and width of , each rounded to the nearest multiple of 16 pixels (the width follows from the height at a 16:9 aspect ratio), and fps is the frame rate. The latency of that video on GPU model is
where , , are the fitted coefficients, is the number of denoising steps, fixed at 50 in all scenarios, is the peak throughput of GPU model , its value for the H100 (the configuration the coefficients were fitted on), and an inference-optimization factor defined below. The quadratic term is self-attention, the linear term the MLP and cross-attention layers, and the constant a fixed per-generation overhead. The linear and constant coefficients are fitted on WAN2.1-T2V-1.3B measurements taken on H100 by Delavande et al. [2] and shared across generation models, while the quadratic coefficient is refitted per generation model; all our scenarios use WAN2.1-T2V-1.3B. The fitted latencies are transferred to other GPU models by scaling with the ratio of peak BF16 dense throughputs: a B200 at 2250 TFLOP/s is modelled as times faster than an H100. This transfer is sound only in a compute-bound regime and would break down if generation were bottlenecked by memory bandwidth.
The fitted latency describes an unoptimized research configuration, not a deployed endpoint. A single inference-optimization factor closes that gap, absorbing step and classifier-free-guidance distillation, kernel fusion, caching, sparse attention and quantization into one multiplier on latency, and since energy consumption during video generation is latency times power, on per-video energy as well. We call the prediction of (S1) at the unoptimized baseline: the WAN2.1-fitted coefficients applied as measured, with no production inference optimization. Factors above 1 are possible: they describe models whose serving configuration is less optimized than this baseline. We calibrate against one of the per-clip energy figures reported by Jegham et al. [15], who estimate text-to-video inference energy from architectural first principles and observable generation parameters. On our reference workload (one 8s, 720p clip at 24 fps on a single H100), the unoptimized baseline predicts 366 Wh, and matching Seedance 1.5’s 102.1 Wh gives for the base year. Table S9 reports the factor implied by every model in that survey. Both and the resolution and duration mix are per-year inputs, so carries an implicit dependence on that we leave out of the notation; the yearly trajectory of is quantified in Supplementary Methods S2.7.
Inference GPU-hours and GPU count. Let be the annual GPU-hours required to generate the yearly videos on GPU model alone. We consider video durations () with respective shares , and resolutions with respective shares ; each set of shares sums to one, , and, for the sake of simplicity, the resolutions and durations are assumed independent. is computed from the per-video latency of (S1), weighted by these mixes. Putting this together, we obtain the annual GPU-hours as:
where the double sum represents the average per-video latency under the current year’s resolution and duration mixes for a given GPU model. The factor converts GPU-seconds to GPU-hours. Taking into account the GPU models with respective share (with ), we obtain the annual active GPU-hours for inference as:
The size of the inference fleet follows from the utilisation rate , the fraction of provisioned GPU time spent actively generating (Supplementary Methods S2.2); the remaining fraction is idle time, during which GPUs have no job. A lower inflates the fleet required to serve the same video volume. The required inference fleet is the ratio between the required annual GPU-time and the active GPU-time offered by one GPU over a year:
Here 8760 is the number of hours in a year and the ceiling reflects that a fractional GPU cannot be provisioned. Later years scale the 2026 value of by each scenario’s annual utilisation gain, compounded year on year; the gains are quantified in Supplementary Methods S2.7.
Inference energy consumption. In terms of energy consumption, each GPU model is characterised by its nameplate power at active hours, and its static power at idle hours. In our simulations, is the thermal design power (TDP) provided by the manufacturers. The active part (in MWh) is computed from the previous active GPU-hours spread over the mix of GPU models:
In (S5), the factor converts Wh to MWh. Adding idle consumption yields the total inference energy consumption of the GPU fleet:
Finally, we also compute the average per-video energy , in Wh per video, from the active inference energy alone. It is reported alongside the annual totals and used in the Results.
Training GPU count and energy. In contrast to the inference fleet, for training we assume a homogeneous fleet of GPUs of model , running at a fraction of the nameplate power . This fraction, like the utilisation rate below, is an explicit modelling assumption without external source (Supplementary Methods S2.5). Idle time is captured as before, through the training utilisation rate : the fraction of provisioned training GPU-time spent actively computing, in analogy with . We consider training runs per year, each requiring MWh on GPUs of model . Training is modelled per run because published evidence reports training campaigns as per-run totals: announced GPU-hour counts (converted to energy at nameplate power) or training-FLOP estimates. A per-run figure already integrates a campaign’s duration, hardware, and utilisation, so the run energy needs no fleet profile for a workload that runs in bounded bursts; the per-run energy is converted back into GPU-hours only to size the training fleet and its idle draw, at the assumed training power fraction rather than the campaign’s own utilisation. Research and development overhead is captured through a additional factor: one training run induces an additional cost equivalent to training runs, i.e., a total of runs.
The number of training GPU-hours is computed as the ratio between the estimated total energy demand for the training runs per year, including research and development overhead, and the useful energy a single training GPU can deliver in one hour:
The factor converts MWh to Wh. Taking into account idle time, the training GPU count is obtained as for the inference one, i.e.:
The GPUs Required indicator reported in the results is the sum of and , though only explicitly drives hardware purchases in our model. The sizing in (S8) spreads the year’s training GPU-hours uniformly across the 8760 hours: is a sizing construct for the reported fleet, not a description of a bursty training campaign. Annual training energy is the sum of the active training energy and idle time power consumption:
Storage. The stored volume (in TiB) is the product of each year’s video volume and that year’s average file size, accumulated over the rolling retention window and scaled by the replication factor (Supplementary Methods S2.5). The annual storage energy is then
in MWh, where the factor 1024 converts TiB to GiB and MWh/GiB (i.e., 6.6 Wh/GiB, converted from the source’s 6.132 kWh per decimal TB per year) is the annual electricity intensity of HDD storage systems [53].
Total energy consumption and peak power demand. GPU-board and storage energy are raised to facility level through nested overheads: a server factor, an IT-equipment factor, and the Power Usage Effectiveness (PUE) factor covering the remaining non-IT loads, following Epoch AI’s methodology [54]. Storage energy receives only PUE: its coefficient is already an IT-level figure that counts the storage servers themselves, so applying the server and IT-equipment overheads would double-count them. The IT-equipment energy and the facility electricity are then
in MWh and GWh respectively. Peak facility power is sized for the worst case in which the whole provisioned inference fleet and the whole provisioned training fleet draw full nameplate power at once:
in MW, where is the fleet-weighted nameplate power, and and are the server and IT-equipment overheads, and is the storage energy, which runs continuously and therefore peaks at its average. This is a lower bound on the true peak: spreads the year’s training GPU-hours uniformly across 8760 hours, so a bursty training campaign concentrating the same hours into a shorter window would need more simultaneous GPUs.
Embodied impacts. Embodied impacts are amortised as a stock, spread over the provisioned fleet and the GPU lifespan, rather than charged in full in the purchase year. This attributes each GPU’s manufacturing footprint to the years it is actually in use, avoiding a spike disconnected from the fleet’s operation. For GHG, the inference fleet contributes
in t CO, where is the per-GPU embodied GHG coefficient in kg CO, is the provisioned fleet of model , the GPU lifespan in years, and the factor 1/1000 converts kg to tonnes. Abiotic depletion potential follows the identical structure with , the per-GPU ADP coefficient, in place of . Training hardware is charged the same way but amortised over GPU-hours rather than GPU-years, as it is not provisioned through the fleet’s purchase and replacement path; because the training GPU-hours inherit the development-overhead multiplier, so does the training embodied impact. The model aggregates these terms into the three reported impacts: GHG emissions sum the operational emissions of facility electricity and the embodied terms of the GPU fleet, storage hardware, and training hardware. Embodied ADP sums the same three embodied terms, with per-GPU ADP coefficients in place of the GHG ones; it carries no operational component. Operational water multiplies the IT-equipment energy of (S11) by the Water Usage Effectiveness (WUE), the facility water input per unit of IT energy. The inventory counts direct potable and non-potable water only; indirect water, and the water embodied in hardware manufacturing, stay outside it.
Hardware, parameters, and data sources
We model two accelerators, H100 and B200, deployed in 8-GPU DGX-class nodes. These two bracket the hardware actually available to a 2026 service: the intermediate B100 was effectively cancelled over a design flaw and most cloud providers moved straight from Hopper to the B200 [55]. Embodied figures sum the bare accelerator board, taken from the Hubble GPU dataset [56], and a per-GPU (1/8) share of its host server, obtained from BoaviztAPI [57]; Table S8 gives the resulting per-GPU values. These are embodied figures should be read as carrying at least a factor-of-two uncertainty: BoaviztAPI’s own GPU-component model returns roughly twice the Hubble board figure for both accelerators, and we have not resolved the divergence.
| Model | TDP (W) | Idle (W) | Peak (TFLOP/s) | GHG (kg COeq) | Embodied ADP (kg Sb-eq) |
| H100 | 700 | 59 | 989 | 885.5 | 0.0577 |
| B200 | 1000 | 140 | 2250 | 1321.5 | 0.05896 |
Table S8. Per-GPU hardware specifications (board plus a 1/8 share of the DGX host). Peak throughput is BF16 dense and drives the cross-GPU latency transfer. Idle power feeds the inference and training idle-energy terms.
What the ADP figure contains. The two datasets publish different ADP variants: the Hubble board values are the abiotic depletion potential for mineral and metal resources (ADPe), while the BoaviztAPI host-server and hard-drive values are the combined ADP, which BoaviztAPI documents as covering both a material and a fossil-resource component. For the H100, the host term accounts for 0.04875 kg Sb-eq of the total 0.0577 kg Sb-eq per GPU, and about a tenth of that host value is its mineral component, so roughly three quarters of the per-GPU figure reflects fossil-resource depletion rather than minerals.
We keep the combined ADP rather than substitute a mineral-only figure because a configuration-sensitive one exists for the board but not for the host server or the hard drive: BoaviztAPI returns the same value for every server configuration and none at all for storage hardware. The composition matters little at service level: storage stays below 0.1% of the total in every reported year, and the inference fleet supplies roughly 60% or more of it, so the indicator is dominated by the embodied ADP of the inference fleet.
Table S9 lists the per-clip energies that Jegham et al. report for each model of their survey and the optimization factor each one implies on our reference workload.
| Model | Per-clip energy (Wh) | |
| Veo 3 | 30.8 | 0.084 |
| Veo 3.1 | 39.5 | 0.108 |
| Seedance 1 | 80.0 | 0.219 |
| Seedance 1.5 | 102.1 | 0.279 |
| Wan 2.1-14B | 382.7 | 1.045 |
| HunyuanVideo | 411.4 | 1.124 |
| HunyuanVideo-1.5 | 478.5 | 1.307 |
Table S9. Inference-optimization factor implied by each model surveyed by Jegham et al. [15], on a reference workload of one 8 s 720p clip at 24 fps on a single H100. Our unoptimized baseline predicts 366 Wh for that clip. The base scenario adopts the Seedance 1.5 value.
Idle power is rarely published by manufacturers, so both figures come from independent measurements: 59 W for the H100 from an NVML-instrumented inference benchmark [58], and 140 W for the B200 from a reported nvidia-smi reading on an 8-GPU node [59]. GPU lifespan is modelled as years, spanning the tension between hyperscaler depreciation schedules (4–6 years) and the shorter service lives implied by observed failure rates and obsolescence.
Annual video volume follows the prior described in Supplementary Methods S2.2. Training energy per run is treated as uncertain across the range of published 2024-and-later text-to-video models, MWh. The mode is Seaweed-7B’s reported 665,000 H100 GPU-hours ( MWh at full TDP), one of the only primary-sourced GPU-hour figures among commercial models and the median of the models we surveyed; the low bound is an efficient 14B model of the Wan class and the high bound a 30B frontier model, with the spread driven almost entirely by parameter count [60]. Published GPU-hour counts are read as energy at full TDP and converted back to GPU-hours at the 80% training power assumed below, so the modelled GPU-hours exceed the published count by a factor 1/0.8. Energy is preserved; the provisioned training count and its idle draw follow the inflated hours. The yearly retraining count is , the product of the number of laboratories that train AI video-generation models and one to two trainings per laboratory per year; Supplementary Methods S2.6 derives this prior. The development overhead is , anchored on the OLMo [61] and Moshi [62] footprint studies: the former reports a development phase costing about half the final run for a mature pipeline, the latter up to forty times for a radically novel architecture. Training is assumed to run on H100s at 80% of TDP, with GPUs active 70% of their provisioned time.
Storage impact is driven by a per-video file size, itself a power law fitted on the same Sora scrape as the resolution and duration mixes, and by a stored-data volume accumulated over a rolling retention window and scaled by a replication factor. Storage energy uses Shehabi et al.’s 6.132 kWh/TB/yr operational intensity [53], and storage embodied impacts use BoaviztAPI’s per-drive HDD figure amortised over a 24 TB nearline drive and a 5-year service life; the reference scenario keeps three replicated copies () over a retention window drawn from years.
Facility overheads draw on an EU data-centre report [26] for PUE and WUE, and on Epoch AI’s GPU power methodology [54] for the server () and IT-equipment () multipliers; combined with our PUE, these leave GPU boards accounting for roughly 40% of facility power. PUE is drawn from a Weibull distribution fitted to the EU survey (mean 1.49) and WUE from triangular(0.5, 0.704, 0.9) L/kWh (mean 0.70), its mode anchored on the report’s hyperscaler average (0.704 L/kWh, Fig. 12). Carbon intensity is drawn from a distribution fitted to European grid data [63] with a mean of 192 g COeq/kWh and a median of 154 g, the gap reflecting a strong right skew across countries. Both the PUE and the carbon-intensity distributions are EU-wide and unweighted by generation location, which is a conservative choice for PUE and an open question for carbon.
Several inputs carry no external source and should be read as explicit modelling assumptions rather than fitted quantities: the GPU fleet composition and its evolution, the training usage and active-time factors, the replication factor and the retention window. The retraining count builds on a documented laboratory count, but its retained value stays an author-set floor (Supplementary Methods S2.6). The warmup fraction affects only the reported purchase schedule, not any environmental output, since embodied impacts amortise over the provisioned fleet rather than the purchase flow.
Deriving the yearly retraining count
The retraining count combines a count of developing laboratories with a per-laboratory training rate. We counted the laboratories that trained and released a video-generation model for commercial use between 2024 and late 2026; this strict count finds 27 laboratories, and stays between 24 and 34 under other boundary choices (for example, counting only products still active in late 2026, or also counting research previews and post-training efforts). The model retains laboratories. We set this floor by hand, below the strict count of 27, as a conservative choice: it absorbs the retirement of training programmes (two major products were discontinued in 2026), releases limited to early access, and laboratories whose continued activity we could not verify. The full laboratory list, with one source per entry, is in our replication package. Each retained laboratory trains one to two video-generation models per year, drawn as , and the count is held constant across the 2026–2030 horizon, so the retraining count follows .
Scenarios
All scenarios share a 2026 baseline and a common technology trajectory in which utilisation improves and inference optimizations compound year over year, and they differ in how demand and video quality evolve to 2030. Utilisation rises year on year by +10% in 2027, +5% in 2028, +2% in 2029, and +0.1% in 2030, and the inference-optimization factor falls by 20%, 10%, 5%, and 5% over the same years, each change applied multiplicatively to the previous year’s value.
We report three scenarios. The Reference scenario lets demand, resolution and duration all grow, with the fleet renewing towards B200s. Usage stagnation freezes volume and video quality at their 2026 values, so that only the technology evolves; it isolates what efficiency gains alone can achieve. Mass adoption keeps the technology, quality and training trajectories of the Reference scenario and accelerates only the volume path, to test how the 2030 impacts respond to demand above the Reference scenario; because the per-video quantities are identical in both scenarios, every impact indicator scales with volume. The yearly retraining count keeps the prior derived in Supplementary Methods S2.6 in all three scenarios, so the scenarios vary demand rather than the intensity of the training race; Supplementary Results S1.2 reports a diagnostic sensitivity run in which this assumption is relaxed.
The volume trajectory of the Reference scenario is deliberately modest. Its 12.8× growth over four years is an average yearly multiple of about 1.9: volume doubles in each of 2027, 2028 and 2029, then grows by +60% in 2030. Growth at Google ran faster: the company reported that the tokens processed monthly across its AI products, a broader unit of activity than video generations, rose from 9.7 trillion in May 2024 to more than 3.2 quadrillion in May 2026 [27]. Against this curve, the expansion of the Reference scenario sits below the growth rates reported by large operators, and we read it as a conservative baseline rather than a central estimate. Usage can also contract: OpenAI discontinued Sora in 2026 [22].
The accelerating rates of Mass adoption (yearly multiples of 2.0 in 2027, 2.5 in 2028, 4.0 in 2029 and 3.0 in 2030) remain at or below the expansion reported for text- and image-generation services over 2024–2026: the Google token growth cited above breaks down into a fifty-fold increase between 2024 and 2025 and a seven-fold increase between 2025 and 2026, while ChatGPT reached about 10% of the world’s adult population within three years of its launch [64]. Whether even these lower rates transfer to video is an open question, so we read the scenario as an illustrative sensitivity case rather than a forecast.
Uncertainty analysis
We quantify output uncertainty with a Monte Carlo procedure over scenario inputs: most numeric fields can be annotated with a distribution (normal, lognormal, uniform, triangular, or a generic distribution with optional bounds) in place of a fixed value, and each run draws one sample of every annotated field before executing the pipeline; the per-element shares of the quality and fleet mixes are the exception and stay fixed. All reported results use 10,000 Monte Carlo samples with a fixed random seed (42), for the three scenarios and for the one-at-a-time sensitivity runs, so every table, tornado panel, and distribution figure in this paper comes from the same seeded run. We complement this with a one-at-a-time sensitivity analysis, sampling a single uncertain field per pass while holding the rest at their median.
A final diagnostic run, not one of the three scenarios, isolates the role of video quality: it follows the Reference scenario’s demand and quality paths but holds the efficiency-related inputs at their 2026 values, with the GPU fleet kept at 100% H100 and the inference-optimization factor and GPU utilisation frozen. Quality growth alone would raise energy per video from 24.4 Wh in 2026 to 20.9× its 2026 level by 2030, against the 8.8× observed under the full technology trajectory.
Threats to Validity
This section lists the methodological threats that remain on top of the deliberate scope decisions recorded in the Limitations subsection of the Methods section. The items here are uncontrolled issues, not scope choices, and each is stated with its mitigation and its residual risk. Because the study is a simulation-based estimate rather than a controlled experiment, internal validity is read as the validity of the model mechanics: given the inputs, are the equations and the transfer steps the correct implementation of the mechanism they claim?
Construct validity
The 2026 volume prior. The dominant input is drawn from a one-sided triangular distribution of 6.2–9.5B videos whose low end and mode sit on the observable total of measured and disclosed rates, above our measured floor of roughly 133M videos per year. The disclosed rates are point-in-time, partial, and self-reported; the two scrapes behind the floor are only a floor. Mitigation: the prior is deliberately wide, anchored at both ends, and the floor comes from two independent measurements. Residual risk: high, on the absolute 2026 level of every output; lower on scenario comparisons, which share the same 2026 baseline.
The per-video energy anchors. The 24.4 Wh GPU-active energy per video is derived, not measured: a quadratic latency model fitted on one model (WAN2.1-T2V-1.3B), multiplied by a single optimization factor that no anchor alone can identify, and calibrated on one clip (Seedance 1.5, 102.1 Wh). Mitigation: the implementation reproduces the source benchmark to within 2% [2], the model’s unoptimized-baseline 366 Wh prediction for the reference clip falls inside the 30.8–478.5 Wh range of Jegham et al. [15], and the market-cost validation matches inside the fitted domain. Residual risk: every model in that survey implies a different optimization factor (0.084–1.307); the inference energy inherits this spread.
One platform’s resolution and duration mix. The resolution and duration mixes (76% last five seconds, 80% at 480p, treated as independent marginals) and the storage file-size model come from one platform’s public feed (Sora, 3.3M videos). Mitigation: the corpus is large, but the mix is a fixed input rather than an uncertainty, so its error does not appear in the reported ranges. Residual risk: if the global mix skews to longer or higher-resolution clips, per-video energy and storage are both understated; the risk runs in both directions.
European conversion coefficients. PUE, WUE, and grid carbon intensity are drawn from European distributions and applied without weighting by where generations actually run, so every non-electricity indicator inherits the location assumption. Mitigation: the PUE distribution is fitted to a European data-centre survey (mean 1.49) and is conservative for PUE. Residual risk: carbon intensity is the largest single driver of the GHG spread: operational electricity spans a factor of 2.4 between the 5th and 95th percentiles of the Monte Carlo samples, GHG 17.1.
Internal validity
The 2030 extrapolation stack. The Reference scenario and Mass adoption 2030 figures combine three unverified steps: the latency coefficients are evaluated outside the fitted token-length range, where the dominant quadratic term amplifies the error; a constant optimization factor is applied where measured optimizations strengthen with token length and therefore under-correct at high tiers; and the H100-to-B200 transfer assumes an empirical efficiency that was measured only on H100. Mitigation: the 2026 baseline stays inside the validated regime, and the market-cost validation shows the model diverges upward at higher tiers. Residual risk: the 2030 impact levels under the Reference scenario and Mass adoption are the least supported numbers in the paper, and on the per-video-energy side it is likely an overestimate; the lower-bound framing is solid for 2026, but the 2030 level carries this upward model error.
External validity
A snapshot of a fast-moving field. The 2026 inputs (fleet, prices, calibrations, grid mix) are a snapshot, and the field moves fast: new models, new accelerators, and grid decarbonisation shift between publication and replication. Mitigation: all results are reported as distributions from one seeded run, and the technology trajectory is an explicit scenario input. Residual risk: any number quoted from this paper decays; the 2030 claims additionally rest on a technology trajectory that is partly assumed, as stated in the Limitations subsection.
Conclusion validity
One-at-a-time sensitivity. The tornado panels rank each input varied alone while the rest are held at their medians, so compounding between volume, quality, and the optimization factor is invisible to the ranking. Mitigation: the per-output Monte Carlo spreads (2.4× for electricity, 17.1× for GHG, between the 5th and 95th percentiles) come from the full joint sampling and do not depend on the one-at-a-time design. Residual risk: the per-input ranking itself can be an artefact of the design; the directional claims (carbon intensity dominates GHG; the training-related parameters lead the GPUs and water outputs) are the ones that rest on it.
Reporting on skewed distributions. The tables report medians with min and max over 10,000 samples; the min and max columns are single-sample extremes, not distributional ranges, and the factor of 123 between the two 2030 scenarios is a ratio of medians. Mitigation: the growth claim is supported by a non-overlap check: the 2030 fifth percentile (1,681 kt) lies above the 2026 ninety-fifth percentile (409 kt). Residual risk: median comparisons can hide overlap where distributions cross; the paper makes no such crossing claim. The 123× scenario contrast is itself widened by the extrapolation stack above, since the Usage stagnation trajectory stays in the validated regime.
References
- [1]C. Morand, A. Néveol, and A.-L. Ligozat, “The Rising Unsustainability of AI Graphics Cards Production,” Jun. 2026, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2607.01258
- [2]J. Delavande, R. Pierrard, and S. Luccioni, “Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models,” Sep. 2025, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2509.19222
- [3]A. P. Privette, A. Barros, and X. Cai, “Data centers water footprint: The need for more transparency,” AGU Advances, vol. 7, p. e2025AV002140, 2026.DOI
- [4]H. Jager and H. S. Yoon, “Data centers: An emerging threat to freshwater biodiversity in the United States,” Water Biology and Security, p. 100585, 2026.DOI
- [5]Kuaishou Technology, “Kling AI annualized revenue run rate hits USD240 Million in December 2025,” Jan. 2026. [Online]. Available: https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-annualized-revenue-run-rate-hits-usd240-million/
- [6]L. Sun and X. Chen, “Automatic social media content generation and style control via multimodal generative models: A methodological study,” in Second International Conference on Communication, Information, and Digital Technologies (CIDT 2025), S. S. Agaian and X. Kong, Eds. Telok Blangah, Singapore: SPIE, Jan. 2026, p. 11.
- [7]H. Wang, G. Zhang, and K. Yan, “Based on Runway Gen-4: A dynamic video generation method for optimizing movie VFX workflows,” in 2025 3rd International Conference on Artificial Intelligence and Automation Control (AIAC), 2025, pp. 146–151.
- [8]A. Kapoor and M. Kumar, “Frontiers: Generative AI and Personalized Video Advertisements,” Marketing Science, vol. 44, no. 4, pp. 733–747, Jul. 2025.
- [9]T. T. Hai, D. T. T. Mai, and N. V. Hanh, “A rapid review of using AI-generated instructional videos in higher education,” Frontiers in Computer Science, vol. 7, p. 1721093, 2025.
- [10]S. Kumar, S. Sai, V. Chamola, A. Gaur, C. Agarwal, K. Huang, and A. Hussain, “Peeping into the Future: Understanding and Combating Generative AI-Based Fake News,” Cognitive Computation, vol. 17, no. 3, p. 103, Jun. 2025.
- [11]Samson Olufemi Olunipekun, “Computational propaganda and misinformation: AI technologies as tools of media manipulation,” World Journal of Advanced Research and Reviews, vol. 25, no. 1, pp. 911–923, Jan. 2025.
- [12]The Chosun Daily, “Kling Surpasses 12 Million Monthly Users,” Jan. 2026. [Online]. Available: https://www.chosun.com/english/industry-en/2026/01/23/CB6RRG6ADJB2TKOW63TD6S4YYA/
- [13]Netflix, Inc., “Netflix Second Quarter 2026 Earnings Interview,” Jul. 2026. [Online]. Available: https://ir.netflix.net/investor-news-and-events/investor-events/event-details/2026/Netflix-Second-Quarter-2026-Earnings-Interview-2026-YudK6GN8y2/default.aspx
- [14]Runway, “The AI Media Report,” Jul. 2026. [Online]. Available: https://runway.com/news/company-news/ai-media-report
- [15]N. Jegham, B. Gamazaychikov, and S. Luccioni, “Lights, Camera, Carbon: Architectural Scaling Laws for Video Generation Energy Consumption,” Jul. 2026, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2607.04553
- [16]B. Li, Y. Jiang, and D. Tiwari, “Carbon in Motion: Characterizing Open-Sora on the Sustainability of Generative AI for Video Generation,” ACM SIGEnergy Energy Informatics Review, vol. 4, no. 5, pp. 160–165, 2024.DOI
- [17]G. Guennebaud and A. Bugeau, “Energy consumption of data transfer: Intensity indicators versus absolute estimates,” Journal of Industrial Ecology, vol. 28, no. 4, pp. 996–1008, 2024.DOI
- [18]Runway, “Introducing Runway Aleph,” Jul. 2025. [Online]. Available: https://runway.com/research/introducing-runway-aleph
- [19]Adobe, “Generative Extend overview in Premiere,” Sep. 2026. [Online]. Available: https://helpx.adobe.com/premiere/desktop/edit-projects/edit-with-generative-ai/generative-extend-overview.html
- [20]Synthesia, “Free AI avatar generator — create fully customizable avatars,” 2026. [Online]. Available: https://www.synthesia.io/features/avatars
- [21]Canva, “Use Magic Media to create photos, graphic, and videos,” 2026. [Online]. Available: https://www.canva.com/help/using-magic-media/
- [22]OpenAI, “What to know about the Sora discontinuation,” Apr. 2026. [Online]. Available: https://help.openai.com/en/articles/2001152-what-to-know-about-the-sora-discontinuation
- [23]J. Gallegos, “Introducing Veo 3.1 and advanced capabilities in Flow,” Oct. 2025. [Online]. Available: https://blog.google/innovation-and-ai/products/veo-updates-flow/
- [24]S. Sala, E. Crenna, M. Secchi, and E. Sanyé-Mengual, “Environmental sustainability of European production and consumption assessed against planetary boundaries,” Journal of Environmental Management, vol. 269, p. 110686, 2020, per-capita planetary-boundary table: climate 985 kg CO2eq/person/yr (GWP100); mineral resources 3.18E-02 kg Sb-eq/person/yr. Verified 2026-09-24. [Online]. Available: https://doi.org/10.1016/j.jenvman.2020.110686
- [25]A. Berthelot, E. Caron, M. Jay, and L. Lefèvre, “Estimating the environmental impact of Generative-AI services using an LCA-based methodology,” Procedia CIRP, vol. 122, pp. 707–712, 2024, 31st CIRP Conference on Life Cycle Engineering (LCE 2024).DOI
- [26]European Commission, Directorate-General for Energy, EY, AIT, and Borderstep, “Assessment of the energy performance and sustainability of data centres in EU,” Jul. 2025. [Online]. Available: https: //op.europa.eu/en/publication-detail/-/publication/83be4c3e-5c79-11f0-a9d0-01aa75ed71a1/language-en
- [27]S. Pichai, “I/O 2026: Welcome to the agentic Gemini era,” May 2026. [Online]. Available: https://blog.google/innovation-and-ai/sundar-pichai-io-2026/
- [28]International Telecommunication Union, “Greenhouse gas emissions trajectories for the information and communication technology sector compatible with the UNFCCC Paris Agreement,” Feb. 2020. [Online]. Available: https://www.itu.int/rec/T-REC-L.1470
- [29]The Shift Project, “Intelligence artificielle, données, calculs : le rapport final du Shift,” Oct. 2025. [Online]. Available: https://theshiftproject.org/app/uploads/2025/11/RF-IA-UK.pdf
- [30]D. Bol, T. Pirson, and R. Dekimpe, “Moore’s law and ICT innovation in the Anthropocene,” in 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE), Feb. 2021, pp. 19–24.DOI
- [31]W. Alpine, N. Geldner, H. Alpine, and M. G. Chepeliev, “AI-driven productivity gains enable more CO₂ emissions than they avoid in a global energy–economy model,” npj Climate Action, vol. 5, no. 1, p. 71, Aug. 2026.DOI
- [32]A. S. Luccioni, E. Strubell, and K. Crawford, “From Efficiency Gains to Rebound Effects: The Problem of Jevons’ Paradox in AI’s Polarized Environmental Debate,” Jan. 2025, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2501.16548
- [33]C. Morand, A.-L. Ligozat, and A. Névóel, “The Environmental Impacts of Machine Learning Keep Rising Evidencing Rebound Effect,” Oct. 2025, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2510.09022
- [34]Epoch AI, “Advanced packaging and HBM, not logic dies, were the bottlenecks on AI chip production in 2025,” 2026, data insight; CC-BY. The page body is not machine-readable; its content was transcribed by the authors on 2026-10-19, and the cited findings rest on that transcription. [Online]. Available: https://epoch.ai/data-insights/ai-chip-supply-chain-constraints
- [35]——, “Compute accounts for the majority of expenses of AI companies,” 2026. [Online]. Available: https://epoch.ai/data-insights/company-spending-breakdown
- [36]L. H. Kaack, P. L. Donti, E. Strubell, G. Kamiya, F. Creutzig, and D. Rolnick, “Aligning artificial intelligence with climate change mitigation,” Nature Climate Change, vol. 12, no. 6, pp. 518–527, Jun. 2022.DOI
- [37]C. Morand, J. Combaz, A. Névóel, and A.-L. Ligozat, “When rebound effect is not a side effect: Analyzing sociotechnical contexts of digital technologies,” Mar. 2026, unpublished manuscript.
- [38]Arup, “A Screen New Deal — a Route Map to Sustainable Film Production,” Arup; British Film Institute; albert, Tech. Rep., Sep. 2020. [Online]. Available: https://www.arup.com/globalassets/downloads/insights/a-screen-new-deal-a-routemap-to-sustainable-film-production.pdf
- [39]Luminate Intelligence, “Animation + AI,” Sep. 2025. [Online]. Available: https://luminatedata.com/studies/animation-ai/
- [40]YouTube, “20 ways we’re celebrating two decades of YouTube,” Apr. 2025. [Online]. Available: https://blog.youtube/news-and-events/happy-birthday-youtube-20/
- [41]——, “YouTube for press,” 2025, press statistics page; the same stat on the 20th-birthday post is footnoted “As of April 2025”. [Online]. Available: https://blog.youtube/press/
- [42]R. McGrady, K. Zheng, R. Curran, J. Baumgartner, and E. Zuckerman, “Dialing for videos: A random sample of YouTube,” Journal of Quantitative Description: Digital Media, vol. 3, 2023. [Online]. Available: https://journalqd.org/article/view/4066DOI
- [43]B. Steel, M. Schirmer, D. Ruths, and J. Pfeffer, “Just another hour on TikTok: Reverse-engineering unique identifiers to obtain a complete slice of TikTok,” Journal of Quantitative Description: Digital Media, vol. 6, no. 1, 2026. [Online]. Available: https://arxiv.org/abs/2504.13279
- [44]RunPod, “GPU Cloud Pricing,” Jul. 2026. [Online]. Available: https://www.runpod.io/pricing
- [45]FairStack Team, “AI Generation Cost: Transparent price breakdown (2026),” Feb. 2026. [Online]. Available: https://fairstack.ai/blog/ai-generation-cost
- [46]HeyGen, “Free ai video generator,” accessed October 2026. [Online]. Available: https://www.heygen.com/tool/ai-video-generator
- [47]MiniMax Group Inc., “Annual results announcement for the year ended December 31, 2025,” Mar. 2026. [Online]. Available: https://www.hkexnews.hk/listedco/listconews/sehk/2026/0302/2026030202837.pdf
- [48]ShengShu Technology, “Vidu launches Q2 reference-to-video, pioneering a new era of high consistency and creative control,” Oct. 2025. [Online]. Available: https://www.newswire.com/news-releases/vidu-launches-q2-reference-to-video-pioneering-a-new-era-of-high-consistency-and-creative-control-302590002.html
- [49]Higgsfield, “The state of AI video report,” 2026, first edition; accessed October 2026. [Online]. Available: https://static.higgsfield.ai/State%20Of%20%20The%20AI%20Video%20Report.pdf
- [50]——, “About Higgsfield,” accessed October 2026. [Online]. Available: https://higgsfield.ai/about
- [51]xAI, “Introducing Grok Imagine 1.0,” Feb. 2026, post on X. [Online]. Available: https://x.com/xai/status/2018164753810764061
- [52]——, “xai all-hands,” Feb. 2026, archived transcript of the meeting of 10 February 2026; accessed October 2026. [Online]. Available: https://elonmuskarchive.org/video/xai-all-hands-2026
- [53]A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomy, E. Masanet, and D. Sartor, “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Laboratory, Berkeley, California, Report LBNL-2001637, Dec. 2024.
- [54]Epoch AI, “GPUs account for about 40% of power usage in AI data centers,” Dec. 2025. [Online]. Available: https://epoch.ai/data-insights/gpus-power-usage-in-ai-data-centers
- [55]A. Garreffa, “Analyst: NVIDIA has “Effectively Canceled” B100 AI GPU over design flaw: B200A to replace it,” Aug. 2024. [Online]. Available: https://www.tweaktown.com/news/100083/analyst-nvidia-has-effectively-canceled-b100-ai-gpu-over-design-flaw-b200a-to-replace-it/index.html
- [56]E. Lees-Perasso, D. Eckhajzer, G. Roussilhe, and T. De Latour, “GPU embedded environmental impacts data,” 2026. [Online]. Available: https://github.com/hubbo-org/GPU-Embedded-Environmental-Impacts
- [57]T. Simon, D. Eckhajzer, A. Berthelot, E. Fourboul, S. Rince, and R. Rouvoy, “BoaviztAPI: A bottom-up model to assess the environmental impacts of cloud services,” ACM SIGEnergy Energy Informatics Review, vol. 4, no. 5, pp. 84–90, Dec. 2024.
- [58]M. Fadel Argerich, J. Fürst, and M. Patiño-Martínez, “Watt counts: Energy-aware benchmark for sustainable LLM inference on heterogeneous GPU architectures,” Apr. 2025, arXiv preprint. [Online]. Available: https://arxiv.org/abs/2504.09048
- [59]L. Li, “NVIDIA Blackwell B200 vs H100: Real-world benchmarks, costs, and why We Self-Host,” Apr. 2025. [Online]. Available: https://www.lightly.ai/blog/nvidia-b200-vs-h100
- [60]Epoch AI, “Data on AI Models,” Aug. 2026. [Online]. Available: https://epoch.ai/data/ai-models
- [61]J. Morrison, C. Na, J. Fernandez, T. Dettmers, E. Strubell, and J. Dodge, “Holistically evaluating the environmental impact of creating language models,” in The Thirteenth International Conference on Learning Representations (ICLR), 2025. [Online]. Available: https://openreview.net/forum?id=04qx93Viwjarxiv.org/abs/2503.05804
- [62]M. López-Rauhut, L. Landrieu, M. Aubry, and A.-L. Ligozat, “Environmental footprint on GenAI research: Insights from the Moshi foundation model,” Transactions on Machine Learning Research, 2026. [Online]. Available: https://openreview.net/forum?id=uurX0xsr8G
- [63]European Environment Agency, “Greenhouse gas emission intensity of electricity generation in Europe,” Jul. 2026. [Online]. Available: https://www.eea.europa.eu/en/analysis/indicators/greenhouse-gas-emission-intensity-of-1
- [64]A. Chatterji, T. Cunningham, D. J. Deming, Z. Hitzig, C. Ong, C. Y. Shan, and K. Wadman, “How people use ChatGPT,” National Bureau of Economic Research, Working Paper 34255, Sep. 2025. [Online]. Available: https://www.nber.org/papers/w34255DOI





