I don't know about near 100% margin, but definitely true on lowest priority access. Plenty of times I get model overloaded, outages, etc on my consumer subs, then switch to my work API sub and everything is perfectly fine.
what costs do you think are attributable to subscriptions that api customers are already paying for?
For one, GPUs use lower power when not at max utilisation; so electricity (and the cooling of that heat) is a factor.
what % of anthropic’s cost basis do you think is subscription power charges, keeping in mind they have a mix of triple net and and colossus which is capacity based pricing?