None of these appear in a headline. All of them appear in the bill.
1. The utilisation floor
Committed-spend agreements bill a floor, commonly 80% to 100% of allocated capacity, whatever you actually consume. Set above your realistic average it is a subscription to hours you will not burn.
On an 8-GPU B300 node at $4.00 per GPU-hour, a 3-year term is $840,960. A floor that overshoots your true usage by 20 points costs you roughly $168,000 on that single node, for nothing.
2. Egress and storage
They arrive on a separate page, usually after the decision. Training workloads moving checkpoints and inference workloads serving real traffic both generate volumes that turn a double-digit percentage of the compute bill into a surprise. Ask for both priced against your actual data movement, not a tier table.
3. Preemption and priority
What happens when the provider is oversubscribed. Some agreements permit reclaiming capacity with notice; some rank customers; some say nothing, which is worse. Get your position in writing, because the clause only matters on the day it is invoked and by then you have no leverage.
4. Refresh and substitution
3 years is 2 generations in this market. If a newer platform lands 14 months in, can you move onto it, at what price, and does the provider have any obligation to offer it? Without a substitution right you have bought the right to run ageing silicon at the price of current silicon.
5. The exit
A term you cannot leave is priced by what it costs to stay. Establish early termination mechanics, what happens if the provider fails to deliver contracted capacity, and whether your commitment survives a change of control at their end. This is the clause with the widest gap between what is offered first and what is available if you ask.
Read them in this order
Floor, then exit, then preemption, then refresh, then egress. That is roughly the order of how much money each one moves, and it is almost exactly the reverse of the order they appear in most agreements.
We read these line by line for buyers before signature, and the desk is paid from the supply side under disclosed arrangements, so the reading does not bend toward a vendor. Start at the GPUaaS desk, or compare against owning on the full cost breakdown.