energy_consumption - Mora Project

# Energy consumption The energy cost of AI generation gets talked about a lot, and from what I can tell, the conversation tends to overshoot in both directions, either dismissing it entirely or treating it as the defining problem with these tools. The real picture is somewhere in the middle, and worth looking at clearly. For a project that uses these tools with intent, knowing where the energy goes and how it compares to the alternatives feels like an honest starting point. ^overview **Data centre electricity consumption** reached 415 TWh globally in 2024, which is roughly 1.5% of total world electricity use, according to the IEA.[^iea-2025] That figure is projected to nearly double, to around 945 TWh, by 2030, driven in large part by AI workloads. A typical AI-focused data centre draws as much power as 100,000 households. ### Image generation in context Across common AI tasks, image and video generations are among the most energy-intensive. Luccioni et al. (2024) measured per-task consumption on professional A100 GPUs: image generation averaged around **2.9 kWh per 1,000 inferences**, compared to 0.002 kWh per 1,000 for a simple text classification task, giving a ratio of roughly 1,450 to one.[^luccioni-2024] Generating a thousand images with a model like SDXL produces around 1,594 g of CO₂ equivalent at average grid intensity. These figures reflect cloud infrastructure with its own compounding overhead; consumer GPU inference draws a fraction of that. ### Local hardware in practice Running generation on a consumer workstation keeps the energy cost visible. Our dual RTX 3060 setup runs on a 700 W power supply, rarely maxed, which roughly comparable to a domestic oven.[^rtx3060-tdp] That consumption shows up directly on an electricity bill instead of being abstracted into computational credits. The size of the cost is immediately legible, and bounded: a session starts and stops, it doesn't accumulate silently with other users' queries. What local generation doesn't change is the source of that power. Whether it comes from wind, nuclear, gas or even coal depends on geography and the local grid, not on where the GPU sits. A kilowatt-hour from hydropower carries a very different weight than one from a coal grid, and most people don't get to choose that. The transparency benefit of running locally is about cost visibility, it can be about the electricity being cleaner but not in every case. ### Compared to traditional VFX rendering For a useful reference point: professional visual effects production has relied on **render farms** for decades, clusters of machines running 24/7, billed by GHz-hour.[^render-farms] A single photorealistic VFX frame in a film, containing fluid simulation, volumetric lighting, ray-traced reflections, can take minutes to hours of continuous compute per node. AI-assisted generation, even for video, can produces comparable output in seconds on a single desktop workstation. Per finished image or short clip, the per-output energy cost can be lower by a large margin if correctly setup. The bottleneck has moved from compute time to creative direction. ## What this means for MORA This project runs inference locally and will also train [[LoRA|LoRAs]], lightweight specialized fine-tuning adapters that run on consumer hardware and cost a fraction of full model training. But it's worth being honest about scope: MORA doesn't train base models from scratch, yet it uses models that were. [[Strubell et al. (2019)]] estimated a single large [[NLP]] training run at roughly 626,000 lbs of CO₂ equivalent.[^training-cost] Every model used here carries that upstream cost, shared across all users who run it. That debt doesn't disappear because inference is local. What local, bounded usage does change is the ongoing cost, the accumulation. Inference sessions start and stop. LoRA training runs are short. The footprint from this project is visible and proportionate, not passively growing in a cloud somewhere. That's a meaningful distinction even if it isn't a clean conscience. --- [^iea-2025]: International Energy Agency — *Energy and AI* (2025). The IEA's dedicated report on AI's role in global energy demand, covering data centre consumption trends, projected growth to 2030, and regional infrastructure breakdowns. https://www.iea.org/reports/energy-and-ai [^luccioni-2024]: [[Luccioni et al. (2024)]] — *"Power Hungry Processing: Watts Driving the Cost of AI Deployment?"* ACM FAccT (2024). arXiv:2311.16863. One of the first systematic comparisons of inference-phase energy consumption across AI task types, measured on production-grade hardware. https://arxiv.org/abs/2311.16863 [^rtx3060-tdp]: NVIDIA GeForce RTX 3060 — official specifications. https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3060-3060ti/ [^render-farms]: Wikipedia — *Render farm* (accessed 2026-03-10). Overview of how distributed rendering infrastructure is structured and operated in film and VFX production. https://en.wikipedia.org/wiki/Render_farm [^training-cost]: [[Strubell et al. (2019)]] — *"Energy and Policy Considerations for Deep Learning in NLP"* (ACL 2019). An early study quantifying the energy and carbon cost of training large language models, which helped establish the distinction between training and inference as separate cost categories. https://arxiv.org/abs/1906.02243