Post #2449179
2026-02-21 08:46 UTC
Replies (1)
-
@fancysandwiches@neuromatch.social 2026-02-21 16:08
@elexia@catcatnya.com correct. No one is doing a single prompt, especially the developers who are enthusiastic, and definitely not the OpenClaw enthusiasts. The argument I've seen for the cost per prompt pricing being low has supposedly included the training costs. Essentially folks are saying you can amortize the cost of training across all prompts later, so cost per prompt is: amortized training cost + inference cost. However, no one actually knows the amortized cost per prompt, AND no one seems to be accounting for the fact that these models aren't around for very long. It would be one thing if we could amortize the cost of the newest frontier model over 10 years, but we're simply not doing that. No one knows the true cost, and as a result the boosters want to argue that it's cheaper than you think, but all evidence shows that it's incredibly expensive when you consider that we're now trying to increase how many terawatts of power we can produce by what seems to be an order of magnitude, just so we can support these models. I don't recall seeing a ton of news about how we need to significantly increase our energy output because Amazon was scaling up AWS, or Microsoft scaling up Azure, or Google scaling up GCloud. I'm sure they had an impact, but not like what we're seeing.