I think the biggest argument I've seen between sales people is "Do you sell something that will increase the customer's revenue or reduce their costs?"
For a well established business saving $1 in costs adds $1 to the bottom line whereas adding $1 in revenue also adds some cost of revenue so maybe it just ends 20 cents to the bottom line or something.
I think every business using AI is different but right now many users are in a "gold rush" phase where they don't have a rational economic model for the whole thing: they might be happy to have GPUs at all and not be worried that they have 30% utilization. Internal politics plays a big role, say my unit has a GPU cluster, I might be really possessive of it and resist sharing it with other units. For that matter I might get a big Mac Studio for my desk because then I can run any model I want that fits on it and never have to have a talk with anyone about model licenses and data privacy and token costs, none of that. From that point of view I could care less about utilization fraction and the one thing I care about is having my head.
I guess there is somebody who can say GPU compute is part of their cost of revenue and increasing the utilization fraction has a clear impact on the bottom line. That somebody probably knows a lot about cost optimization already (or thinks they know!) and might think you are bringing a tic-tac-toe master to a chess tournament! Might still be the kind of person you want to talk to just to talk to.
I think the biggest argument I've seen between sales people is "Do you sell something that will increase the customer's revenue or reduce their costs?"
For a well established business saving $1 in costs adds $1 to the bottom line whereas adding $1 in revenue also adds some cost of revenue so maybe it just ends 20 cents to the bottom line or something.
I think every business using AI is different but right now many users are in a "gold rush" phase where they don't have a rational economic model for the whole thing: they might be happy to have GPUs at all and not be worried that they have 30% utilization. Internal politics plays a big role, say my unit has a GPU cluster, I might be really possessive of it and resist sharing it with other units. For that matter I might get a big Mac Studio for my desk because then I can run any model I want that fits on it and never have to have a talk with anyone about model licenses and data privacy and token costs, none of that. From that point of view I could care less about utilization fraction and the one thing I care about is having my head.
I guess there is somebody who can say GPU compute is part of their cost of revenue and increasing the utilization fraction has a clear impact on the bottom line. That somebody probably knows a lot about cost optimization already (or thinks they know!) and might think you are bringing a tic-tac-toe master to a chess tournament! Might still be the kind of person you want to talk to just to talk to.