I didn’t read the article, but I really hope the headline is true
It makes the opposite point as the headline.
You don’t really believe that Germany only has 500 datacenters, do you?
Copyright owners are not going to be paid and electricity costs aren’t that high. One query is about equal to 10 web searches. Which isn’t nothing but it’s not crazy either. This may improve with time as well. By the time investors stop the subsidies, I expect the price will be moderate.
So far it’s been getting more expensive, not less
Training is the problem. Inference is pretty manageable, training isn’t.
Well training has been done. So I don’t see prices increasing for existing models. If anything, future models may be more expensive if they need a lot of computation, although there will likely be cost declines in that over time, especially now that there’s a huge data center construction boom.
Which where built with loans that need to be paid off. Which also have a lot of GPUs that need replacement every 5 years, sometimes sooner. There’s more cost centers than just electricity.
Training needs to continue though and I don’t think continous learning is solved.
I’m assuming efficiency gains are being delivered - but that’s combating growing scale.
I wouldn’t mind an annual training run performed by some global academic body, with revenues flowing from private/public customers through to data sources.
