Hacker News Reader

Comment by Dead_Lemon | Hacker News Reader

All LLM are heavily subsidized by investors. No one is paying the true cost of tokens at the moment. The rug pull is to get as many people defendant on the tool before they have to massively increase prices.

View on HN

What about people who are running open models on their own GPUs? They are being subsidised too?

That's the point gp was making. If the big labs all increase prices 10x, the cost of running your own open models stays the same

Sure, open models will definitely be an alternative once the frontier models hike prices to match costs. But that will be a minority, and people will start to switch, but people use the frontier models for a reason and these AI labs will extort that, once the easy investor money dries up.

My company pay a copilot subscription to everyone, plus my team have a bedrock subscription that cost us 7k last month. In the meantime, some guys created a new internal team and are testing open source models (for our red team) on our own hardware (we didn't sold our 'old', 3 year old servers RAM/CPU, so the only investment was GPUs). My guess is that within the next 6 months our red team will use open source models exclusively, and my team will probably follow within two years. Especially if the GPU prices go down

> Especially if the GPU prices go down

I rather expect them to go up. I bought a bunch three years ago and a lot more recently. All of them I could sell for 2x what I paid.

How long are those models going to be safe to operate with internet access and tool use without ever being updated? Imagine if "do it for my grandma" is still as effective today because the models haven't changed and that every open-source maintainer upset with AI started polluting their libraries' logs with prompts that have been proven to work with no known reliable fix ever coming in this hypothetical future.

You're assuming that people will be able to buy GPUs in the future. Which currently isn't the case and there is no timeline when they will be available again (at their pre-AI affordable price). I wouldn't be surprised if the labs keep buying up all the GPU capacity forever just to keep home models at bay (which I'm sure is what they're currently doing with all computer hardware).

Evidence ?

Who are these investors subsidizing random inference providers on https://openrouter.ai/ ?

Also: who are the investors subsidizing electricity bills for the people who run LLMs on their own GPUs? I'm still waiting for my subsidy cheque over here, so if you can plz send it to me ASAP, thanks.

Question: did you train that LLM yourself on your own GPU?

Question back: how do you imagine them "raising the prices" of the tokens generated on my own GPU? That's the implication we're discussing here. Please explain how that works.

I think the crux of the issue is that frontier labs do not make any financial sense. They are spending investor dollars in order to sell tokens at pennies, and then a Chinese lab can build a competing model on those pennies, and the frontier labs can't then start selling the tokens at a price where they'd actually even break even because most of the volume will then shift to those cheaper models.

as much as training might be subsidized, I don't buy that inference is, if there was no training required and labs could run their current models forever it'd be a gravy train even at current prices

Reply on news.ycombinator.com