A large language model (LLM) needs more than data to run. Training it and using it to produce responses—called inference—require specialized hardware and substantial electricity. Its environmental impact is therefore a question about both computing and the physical resources behind it. section 11.3
A broad assessment looks beyond the electricity used by a running model:
- Hardware begins with resource extraction. Computer chips require materials including silicon, aluminum, and copper, followed by refining and manufacturing. section 11.3
- Training and inference both consume electricity on specialized chips called graphics processing units (GPUs). section 11.3
- Supply chains matter too. Advanced chips can depend on many vendors, while only a few suppliers can make certain components. section 11.3
Why hardware belongs in the picture
A reported GPU shortage in August 2023 saw startups and large corporations competing for chips. The rush makes clear why assessing LLMs only while they are running would leave out the hardware needed to operate them. section 11.3
Measuring these effects could help inform standards or limits on acceptable impact. That remains a governance aim, not a settled measure: claims about an LLM’s environmental cost depend in part on which stages of its life cycle are counted. section 11.3
