TL;DRAI has an energy footprint — training and serving at scale consume real power.
Training frontier models and serving billions of queries use significant electricity and water (for cooling), concentrated in data centers. Inference at scale can outweigh one-time training over a model's life. The field is responding with efficiency (better hardware, quantization, MoE, smaller task-fit models) and cleaner energy sourcing. It's a real cost and externality to weigh, not a reason for paralysis — efficiency is both green and cheaper.
Treating efficiency as purely environmental (it’s also cheaper)
Try it: Explain why inference energy can exceed training energy over a model’s life.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>train once (big) + serve billions of queries (ongoing)
→ inference energy can dwarf training over time
efficiency = greener AND cheaper</pre></body></html>