the.com/compute costs
the metered heartbeat of the internet age: thinking now has a price per hour.
means the money spent renting processing power, memory, and storage to run software, especially AI models, usually billed by usage rather than ownership.
from born with cloud computing in the 2000s, when amazon and google turned owning servers into renting them, then exploded when AI training turned electricity and silicon into the new oil.
gpu scarcitynvidia h100 chips once resold for triple retail price
training tabgpt-4 reportedly cost over 100 million dollars to train
inference foreverrunning a model costs more over time than training it
power problemdata centers may use 8 percent of us electricity by 2030
for instance
openai training runs — gpt-4 scale training estimated near 100 million dollars, 2023
aws ec2 pricing — charges per second for rented cloud servers since 2017
nvidia h100 shortage — demand outpaced supply through 2023 to 2024 ai boom
microsoft three mile island — reopening a nuclear plant in 2024 just to power ai