I want to let people know why I’m strictly against using AI in everything I do without sounding like an ‘AI vegan’, especially in front of those who are genuinely ready to listen and follow the same.
Any sources I try to find to cite regarding my viewpoint are either mild enough to be considered AI generated themselves or filled with extremist views of the author. I want to explain the situation in an objective manner that is simple to understand and also alarming enough for them to take action.
Check out wheresyoured.at for some “haters guides.”
My general take is that virtually none of the common “useful” forms of AI are even remotely sustainable strictly from a financial standpoint, so there’s not use getting too excited about them.
The financial argument is pretty difficult to make.
You’re right in one sense, there is a bubble here and some investors/companies are going to lose a lot of money when they get beaten by competitors.
However, you’re also wrong in the sense that the marginal cost to run them is actually quite low, even with the hardware and electricity costs. The benefit doesn’t have to be that high to generate a positive ROI with such low marginal costs.
People are clearly using these tools more and more, even for commercial purposes when you’re paying per token and not some subsidized subscription, just check out the graphs on OpenRouter https://openrouter.ai/rankings
None of the hyperscalers have produced enough revenue to even cover operating costs. Many have reported deceptive “annualized” figures or just stopped reporting at all.
Couple that with the hardware having a limited lifespan of around 5 years, and you’ve got an entire industry being subsidized by hype.
Covering operating costs doesn’t make sense as the threshold for this discussion though.
Operating costs would include things like computing costs for training new models and staffing costs for researchers, both of which would completely disappear in a marginal cost calculation for an existing model.
If we use Deepseek R1 as an example of a large high end model, you can run a 8-bit quantized version of the 600B+ parameter model on Vast.Ai for about $18 per hour, or even on AWS for like $50/hour. Those produce tokens fast enough that you can have quite a few users on it at the same time, or even automated processes running concurrently with users. Most medium sized businesses could likely generate more than $50 in benefit from it per running hour, especially since you can just shut it down at night and not even pay for that time.
You can just look at it from a much smaller perspective too. A small business could buy access to consumer GPU based systems and use them profitably with 30B or 120B parameter open source models for dollars per hour. I know this is possible, because I’m actively doing it.