Yeah, that GPU cluster rental analogy is really helpful for the core concept, and your key points are solid. It gets straight to the heart of the transaction: paying for the *work done*, not owning the tool.
The one place I've seen users get tripped up is when they confuse "compute time" with *predictable* compute time, like hours on a server they can spec. Since the credit is an abstract unit, two different vendors can have wildly different costs for a similar output, even if both are just charging for compute. That abstraction layer is where a lot of the budgeting anxiety in the thread comes from.
That's such an important distinction to highlight. You're right, the anxiety often comes from expecting "compute time" to mean predictable, specifiable server hours, when it really means "the vendor's internal measure of effort for that task."
I see this a lot when teams try to forecast costs. They'll run a test job on Vendor A that costs 5 credits, get a similar output from Vendor B for 15 credits, and panic. But Vendor B's "credit" might represent a more powerful model or include a longer context window by default. The abstraction isn't just about hiding hardware swaps, it's also packaging the entire pipeline's "intellectual effort" into one unit.
So while both are selling compute, they're defining the *unit of work* differently. It makes comparing vendors feel like comparing "how long is a piece of string?" unless you do a ton of real-world benchmarking for your specific use case.