TL;DR
This study provides the first global estimate of the environmental footprint of AI video generation, revealing that inference costs—not training—dominate the total impact. The authors warn that without curbing demand growth, even aggressive technological efficiency gains will be insufficient to prevent a massive 75-fold increase in carbon emissions by 2030.
What Problem Does This Paper Address?
While the environmental impact of AI is increasingly documented in aggregate, the footprint of specific modalities remains largely "invisible." Video generation is a particularly resource-intensive outlier: generating a single short video consumes roughly 90 Wh—nearly 30 times the energy of image generation and 45,000 times that of text classification.
Because commercial providers rarely disclose video volumes or energy needs, researchers have struggled to estimate the "absolute footprint" of these services at scale. This paper tackles that gap by building a bottom-up model to estimate the direct environmental impacts (carbon, water, energy, and resource depletion) of text-to-video (T2V) generation as a deployed technology, both for the present day and projected through 2030.
Methods
The researchers developed a serving infrastructure model that scales from the energy and latency of a single video generation up to yearly global volumes.
- Measuring Volume: To anchor their model in reality, they conducted a measurement campaign, crawling OpenAI’s Sora public feed for 54 days and monitoring HeyGen’s public activity counter for 90 days.
- Modeling Infrastructure: They combined these measurements with disclosed figures from other major providers (Kling, Grok Imagine, etc.) to estimate a 2026 baseline volume.
- Impact Indicators: The model tracks four key metrics: operational electricity, greenhouse gas (GHG) emissions, operational water use, and embodied abiotic resource depletion (ADP).
- Uncertainty Analysis: They performed 10,000 Monte Carlo simulations to quantify the spread of their estimates based on uncertain inputs like grid carbon intensity and GPU lifespan.
Datasets
- Sora Public Feed: 3.3 million public videos recorded, providing resolution and duration distributions.
- HeyGen Activity Counter: Real-time monitoring to deduce an annual rate of ~111 million videos.
- Provider Disclosures: Annualized rates from Kling, Veo, MiniMax, Vidu, Higgsfield, and Grok Imagine.
- Hardware Benchmarks: Latency models fitted on WAN2.1-T2V-1.3B measurements using NVIDIA H100 GPUs.
Key Insights and Ideas
- Inference is the Elephant in the Room: While public discourse often focuses on the massive energy cost of training models, this study finds that for a deployed service, inference accounts for ~69% of the total carbon footprint.
- The "Quality Trap": Energy consumption scales near-quadratically with video resolution and duration. As services move toward higher definition and longer clips, the energy per video skyrockets, potentially wiping out efficiency gains from better hardware.
- Efficiency Can't Outrun Growth: Under a "Reference" growth scenario (where volume grows ~13x by 2030), emissions rise 75-fold. To keep emissions flat at 2026 levels, the industry would need a staggering 75% annual improvement in inference efficiency—far beyond the ~10% annual gains currently modeled.
- Hardware Renewal Shifts the Burden: Improving operational efficiency often requires faster hardware cycles, which increases the "embodied" footprint (the environmental cost of manufacturing the GPUs).
Key Results
- 2026 Baseline: The median estimate for 2026 is 7.16 billion videos, requiring a fleet of 72,600 GPUs, consuming 724 GWh of electricity, and emitting 124.3 kt CO2eq. This is equivalent to the annual emissions of roughly 126,000 people.
- The "All-in" Cost: While a single generation might take ~24 Wh of GPU energy, the "all-in" cost (including idle time, facility overheads, training, and storage) is approximately 100 Wh per video.
- 2030 Projections:
- Reference Scenario: Emissions rise to 9,202 kt CO2eq (a 74x increase).
- Mass Adoption Scenario: If demand follows the trajectory of previous AI products, emissions could reach 42.9 Mt CO2eq—comparable to the footprint of 44 million people.
- Usage Stagnation: Only by freezing demand and quality at 2026 levels do technological improvements actually reduce the total footprint (by about 40%).
Limitations and Future Work
The authors emphasize that their estimates are a conservative lower bound. They deliberately excluded:
- Audio generation, which is increasingly bundled with video.
- Downstream impacts like video streaming, playback, and social media storage.
- Non-specialized uses like AI-powered video editing or "talking head" avatars.
Future work is needed to replace current "triangular priors" with measured, location-weighted values as providers (hopefully) become more transparent about their data center locations and hardware utilization. The study concludes that software optimization alone cannot hold the footprint steady; sustainability in this sector depends heavily on managing demand and video quality.
Summary updated