From AI to Absorption: Office Demand, AI Talent Concentrations, and What it Means for Data Centers

Open AI’s GPT-3 used more power than other leading models to train Large Language Models Use a lot of Power to Train

Energy consumption when training LLMs in 2022 (in MWh)

• Energy consumption in training is high; frequent retraining is required to maintain data relevance. Lifetime energy consumption is even higher than initial training usage. • Energy savings from

1,400

1,200

1,000

800

600

AI will be big. Mobile phone operators alone expect AI to reduce power consumption by 10-15%.

MWh Used in Training

400

200

0

GPT-3

Gopher

Bloom

OPT

Note: San Francisco includes pre-OpenAI lease total

Source: Cushman & Wakefield Research, Cornell University, 2023

29