Conversation

Notices

1. Embed this notice
   GreenSkyOverMe (Monika) (greenskyoverme@ohai.social)'s status on Tuesday, 30-Jun-2026 21:13:34 JST GreenSkyOverMe (Monika)

   What do you think will be a realistic price for „LLM please improve the style of my e-mail“ once the AI companies charge a price per token that generates profits for them, maybe later this year or next year?

   Asking for my manager.

   • Embed this notice
     GreenSkyOverMe (Monika) (greenskyoverme@ohai.social)'s status on Tuesday, 30-Jun-2026 23:58:22 JST GreenSkyOverMe (Monika)

     in reply to

     • Wladimir Palant

     @WPalant Thank you!

   • Embed this notice
     Wladimir Palant (wpalant@infosec.exchange)'s status on Tuesday, 30-Jun-2026 23:58:23 JST Wladimir Palant

     in reply to

     @GreenSkyOverMe Let’s see… You are essentially renting a GPU in the cloud or comparable hardware. Not a consumer-grade GPU, those aren’t capable of running the more advanced models – it has to be something with more memory. Let’s say we are talking about Nvidia A100 with 80 GB of memory. Google currently rents those for approximately $5 per hour: https://cloud.google.com/products/compute/pricing/accelerator-optimized

     The big unknown is how long it takes this GPU to process your email. Depends on the length of the email of course and on how well the respective LLM performs on the GPU. Maybe someone can contribute reliable numbers but we are probably talking about roughly a minute of GPU time. Which gives you something like 10¢.

     However, this assumes that the model itself is free which it is not. Millions of GPU hours went into training this model, not to mention the human work required (much of it underpaid African workers). It should be possible to get financial info on the investment here but the next hurdle is: how many customers are going to pay for that? It’s a very competitive market, and models have to be continuously improved – that isn’t a one-time investment. At the same time the number of customers is bound to fall drastically once the companies start charging realistic fees. So the bulk of these fees will be offsetting the cost of developing the models. If I have to make a guess, this is going to be at least ten times the costs of the hardware to run your prompt. Meaning that we are talking about total prices beyond $1.