The reason this works is that you get lesser inference time compute used for queries on these cheaper plans which makes it sustainable and this is enough for the tasks these guys do.
And some local telcos are bundling this subscription as well, so most people just get it for "free". For example, I get Google AI Pro for free with my 350 INR/mo telco plan.
Unless they can massively scale down training and inference cost or implement AGI I don't know what their plan is. Just provide a subsidized plan for the next 10 or 20 years? Their costs are directly proportional to the amount of tokens the LLM produces. How is a monthly subscription plan supposed to account for such costs?
Or else?
Their investors will throw the CEO under the bus and hire a new one that will enshittify it enough to make them money.
Even Deepseek at its absurd prices is a very healthy business. Regarding their return on capex multiple, their CEO said they make a six-fold profit on their compute capex with 10 month recuperation. Because of this, all of them are spending aggressively on compute. Apart from that, user acquisition and data labelling are the major costs that are preventing net profitability right now. High quality data labelling is said to not have a cost advantage in china etc as well and they pay global market prices for this. I can confirm this is true in india too the model companies I know pay global market rates for high quality data.
> Their costs are directly proportional to the amount of tokens the LLM produces. How is a monthly subscription plan supposed to account for such costs?
By limiting the number of tokens you use per month? per week, per hour? And by limiting the inference time compute dedicated to each turn in each session.
> need to recoup
the world economy has shown itself capable of handling decade-scale recouping easily
The main obstacle today in the inference business is the high variability in usefulness/token. This does not need to be solved, but rather only quantified. Innovation is needed to be able to reasonably bound this variance for a reasonable subset of tasks. And we are making progress on this. Naturally though, tasks on the frontier of current capabilities have very high variance. The last couple of years has followed the pattern where tasks no longer on the frontier have reduced variance, but I am not claiming this will continue to be the case generally as the frontier improves.
That's "theoretical profit" - in some imaginary world where the free subscribers would pay the top tier cost.
https://techcrunch.com/2025/03/01/deepseek-claims-theoretica...
I am not familiar with chineese model companies as much as I am with US based ones so I don't have much to say beyond that the CEO is incentivced to pump up those numbers.
> By limiting the number of tokens you use per month? per week, per hour? And by limiting the inference time compute dedicated to each turn in each session.
If this was so simple I don't know why GitHub copilot went to token based billing at my company.
> This does not need to be solved, but rather only quantified. Innovation is needed to be able to reasonably bound this variance for a reasonable subset of tasks
It's much better to make a business case for them after finding this bound right? Currently I can't use copilot for anything serious since I cannot predict how many credits one request is going to consume.
Your point about non frontier tasks using less tokens makes sense. As you said, let's see if it holds up
> If this was so simple...copilot...
Github copilot still has subscriptions. They moved away from request based accounting to token based accounting for the usage limits, as did cursor, and everybody else.