AI demand just pushed gpu rent up 58% in a year. Whats next?
retained text
michael dell posted that the h100, a chip they started shipping in 2022, rented for $1.96/hour a year ago and $3.10 today. +58% on hardware that was supposed to be obsolete by now. coreweave said prices on their older nvidia chips went up 30 to 50% last quarter gpu rent is the floor under every api price. so far the labs have been eating the difference. but anthropic prices its ipo wednesday, so from thursday there are public shareholders asking about gross margin every quarter, and eating costs gets a lot harder to explain which brings it to us. most ai apps, mine included, sell a flat plan on top of a bill that moves. that only works if the bill keeps going down we run a small ai writing tool on flat $19 plans. reran the numbers with model costs 30% higher and about 1 in 6 accounts went negative. those are the heavy users, and right now everyone else is paying for them so the plan is a monthly credit allowance with a hard cap per account, and overage only if they opt in. the pricing part took an afternoon. the billing part is where it gets fiddly, credits that expire, caps that have to be checked before the call and not at invoice time. stripe meters covers the simple version, we’re also looking at some open source options that does credits and caps together, pls suggest if any not saying api prices jump next month. but i’d stop pricing like inference only ever gets cheaper if you’re building on an api, what’s your plan if cost per call goes up instead of down?
document fetched 2026-11-18 00:28 · canonical https://reddit.com/r/OpenAI/comments/1xyr7vs/ai_demand_just_pushed_gpu_rent_up_58_in_a_year