Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="trappedslider" data-source="post: 9835814" data-attributes="member: 41932"> <table style='width: 841px'><tr><th>AI Task</th><th>Per Unit</th><th>30 Minutes</th><th>60 Minutes</th></tr><tr><td>ChatGPT Text Generation (1 prompt/min)</td><td>~0.3 Wh</td><td>~9 Wh</td><td>~18 Wh</td></tr><tr><td>Image Generation (e.g., DALL·E)</td><td>~2.9 Wh/image</td><td>~58 Wh</td><td>~116 Wh</td></tr><tr><td>Audio Generation (e.g., MusicLM)</td><td>~5 Wh/minute</td><td>~150 Wh</td><td>~300 Wh</td></tr><tr><td>Video Generation (estimated)</td><td>3–10 Wh/minute†</td><td>~180–600 Wh</td><td>~360–1200 Wh</td></tr><tr><td>YouTube Video Streaming (HD)</td><td>~12 Wh/5 min</td><td>~144 Wh</td><td>~288 Wh</td></tr></table>† Video generation estimates remain variable depending on platform, resolution, and model complexity. Current figures are drawn from public estimates and indirect measurements such as <a href="https://aws.amazon.com/solutions/case-studies/synthesia-case-study/" target="_blank">Synthesia Case Study</a> and studies on frame-level inference scaling.Note: Figures are for data center/server-side energy, and do not include device, network, or edge delivery (unless noted).<ul> <li data-xf-list-type="ul">Training: This is the heavy-lifting stage where a model like GPT-4 is created. It requires massive amounts of data, compute, and time. Training a single large-scale model can consume millions of kilowatt-hours and take weeks or months of 24/7 GPU usage across thousands of servers. Training is capital- and energy-intensive, but it is an infrequent and centralised operation. In some enterprise scenarios, however, models are fine-tuned, refreshed, or retrained regularly — adding incremental training costs that shouldn’t be ignored.</li> <li data-xf-list-type="ul">Inference: This is what happens when you use the model — e.g., asking a question, generating text, or solving a problem. Inference is lightweight by comparison, typically consuming just a fraction of a watt-hour per query. It is decentralised and real-time, and the energy cost is proportional to how many interactions take place.</li> </ul>The following figures represent estimated energy use per individual user over typical engagement durations: <table style='width: 841px'><tr><th>Platform/Service</th><th>Per Interaction</th><th>30 Minutes</th><th>60 Minutes</th><th>Key Notes</th></tr><tr><td>TikTok (1 short video ≈ 15s)</td><td>~10.4 Wh (per video)</td><td>~1250–2500 Wh (1.25–2.5 kWh)</td><td>~2600–5000 Wh (2.6–5 kWh)</td><td>High-resolution short video, autoplay, high engagement loop</td></tr><tr><td>YouTube HD Streaming (5 min)</td><td>~12 Wh (per 5 min video)</td><td>~180–360 Wh</td><td>~360–720 Wh</td><td>Varies by resolution and device</td></tr><tr><td>Facebook/Instagram Browsing</td><td>~3.3–5.5 Wh (per scroll/post)</td><td>~60–100 Wh</td><td>~120–200 Wh</td><td>Includes video, image loading, and backend AI feeds</td></tr><tr><td>ChatGPT (1 prompt)</td><td>~0.3 Wh</td><td>~9 Wh</td><td>~18 Wh</td><td>Turn-based, ephemeral inference only</td></tr></table>Note: YouTube and TikTok energy figures are based on 2019–2021 estimates; actuals may now be lower due to ongoing efficiency improvements in streaming and content delivery.While energy use per unit of content might be lower (especially with caching and CDN delivery), the scale and continuous nature of usage is what drives up the carbon footprint of these platforms.Even a five-minute YouTube video — viewed 1 billion times — results in hundreds of millions of watt-hours consumed globally.Yes, AI uses energy — and yes, it should be designed, deployed, and scaled responsibly. But let’s not fall into the trap of isolating AI as the villain while ignoring:<ul> <li data-xf-list-type="ul">The persistent energy drain of social platforms</li> <li data-xf-list-type="ul">The scale multiplier of viral content</li> <li data-xf-list-type="ul">The invisible cost of idle digital infrastructure</li> </ul></blockquote>

Verification