[theOGognf]

Some anecdotal data, but we recently estimated the cost of running a LLM at $WORK by looking at power usage over a bursty period of requests from our internal users and it was on the order of $10s/mil tokens. And we arent a big place, nor were our servers at max load, so I can see the cost being much lower at scale

[exceptione]

This is only the power usage?

[theOGognf]

Right, this is only power usage. Factoring in labor and all that would make it more expensive for sure. However, it’s not like it’s a complex system to maintain. We use a popular inference server and just run it with some modest rate limits . It’s been hands-off for close to a year at this point

[exceptione]

Ok! What hardware do you run? I had thought that would be the most expensive part.

[dist-epoch]

Hardware spend also need to be amortized (over 1 year? 2 years?) Unless you cloud rent.

[jenny91]

5 year amortization is pretty realistic I'd say. A100s (came out 2020Q1) are still in heavy use. (I think V100s from 2017Q3 are starting to be phased out a fair bit.)

[theOGognf]

That is true too