so, for an extremely unscientific demonstration, here (warning: AWS may try hard to get you to engage with Explainer[0]) is an instance of an aws pricing estimate for big handwave “some gpu compute”

and when I say “extremely unscientific”, I mean “I largely pulled the numbers out of my ass”. even so, they’re not entirely baseless, nor just picking absolute maxvals and laughing

parameters assumptions made:

• “somewhat beefy” gpu instances (g4dn.4xlarge, selected through the tried and tested “squint until it looks right” method)
• 6-day traffic pattern, excluding sunday[1]
• daily “4h peak” total peak load profile[2]
• 50 instances mininum, 150 maximum (let’s pretend we’re not openai but are instead some random fuckwit flybynight modelfuckery startup)
• us west coast
• spot instances, convertible spot reserves, 3y full prepay commit (yeah I know full vs partial is a big diff; once again, snore)

(and before we get any fucking ruleslawyering dumb motherfuckers rolling in here about accuracy or whatever: get fucked kthx. this is just a very loosely demonstrative example)

so you’d have a variable buffer of 50…150 instances, featuring 3.2…9.6TiB of RAM for working set size, 800…2400 vCPU, 50…150 nvidia t4 cores, and 800…2400GiB gpu vram

let’s presume a perfectly spherical ops team of uniform capability[3] and imagine that we have some lovely and capable active instance prewarming and correct host caching and whatnot. y’know, things to reduce user latency. let’s pretend we’re fully dynamic[4]

so, by the numbers, then

1y times 4h daily gives us 1460h (in seconds, that’s 5256000). this extremely inaccurate full-of-presumptions number gives us “service-capable life time”. the times your concierge is at the desk, the times you can get pizza delivered.

x3 to get to lifetime matching our spot commit, x50…x150 to get to “total possible instance hours”. which is the top end of our sunshine and rainbows pretend compute budget. which, of course, we still have exactly no idea how to spend. because we don’t know the real cost of servicing a query!

but let’s work backwards from some made-up shit, using numbers The Poor Public gets (vs numbers Free Microsoft Credits will imbue unto you), and see where we end up!

so that means our baseline:

• upfront cost: $4,527,400.00
• monthly: $1460.00 (x3 x12 = $52560)
• whatever the hell else is incurred (s3, bandwidth, …)
• >=200k/y per ops/whatever person we have

3y of 4h-daily at 50 instances = 788400000 seconds. at 150 instances, 2365200000 seconds.

so we can say that, for our deeply Whiffs Ever So Slightly values, a second’s compute on the low instance-count end is $0.01722755 and $0.00574252 at the higher instance-count end! which gives us a bit of a handle!

this, of course, entirely ignores parallelism, n-instance job/load/whatever distribution, database lookups, network traffic, allllllll kinds of shit. which we can’t really have good information on without some insider infrastructure leaks anyway. if we pretend to look at the compute alone.

so what does $1000/query mean, in the sense of our very ridiculous and fantastical numbers? since the units are now The Same, we can simply divide things!

at the 50 instance mark, we’d need to hypothetically spend 174139.68 instance-seconds. that’s 2.0154 days of linear compute!

at the 150 instance mark, 522419.05 instance-seconds! 6.070 days of linear compute!

so! what have we learned? well, we’ve learned that we couldn’t deliver responses to prompts in Reasonable Time at these hardware presumptions! which, again, are linear presumptions. and there’s gonna be a fair chunk of parallelism and other parts involved here. but even so, turns out it’d be a bit of a sizable chunk of compute allocated. to even a single prompt response.

[0] - a product/service whose very existence I find hilarious; the entire suite of aws products is designed to extract as much money from every possible function whatsoever, leading to complexity, which they then respond to by… producing a chatbot to “guide users”

[1] - yes yes I know, the world is not uniform and the fucking promptfans come from everywhere. I’m presuming amerocentric design thinking (which imo is probably not wrong)

[2] - let’s pretend that the calculators’ presumption of 4h persistent peak load and our presumption of short-duration load approaching 4h cumulative are the same

[3] - oh, who am I kidding, you know it’s gonna be some dumb motherfuckers with ansible and k8s and terraform and chucklefuckery