GPUaaS contracts, read properly before you sign.

1. What a node-hour actually means. Whether the clock runs on allocation or on availability, whether a degraded node bills at full rate, and whether a failed node inside a cluster stops the meter for the whole cluster. On large training runs this single definition can dominate everything else in the contract.
2. Goodput, not raw hours. A cluster delivering nominal availability while dropping jobs to node failures is not delivering what you bought. Where possible we push toward a measure tied to completed work rather than hours allocated.
3. Commitment burn-down and rollover. If you underconsume in a period, does the commitment roll forward, expire, or convert? Expiring commitments on a workload whose shape is still moving is how buyers end up paying for capacity they never used.
4. Preemption and reclaim rights. On anything below firm reserved capacity, the provider's right to reclaim your nodes — and the notice you get — is the difference between a usable tier and a cheap one that cannot carry production.
5. Storage, egress and interconnect. Rarely in the headline pricing, frequently a material share of the bill, and the primary mechanism by which a cheap compute rate becomes an expensive relationship. Caps and included allowances are negotiable.
6. Term length against the hardware cycle. Offtakers commit in three-to-five-year windows for a reason: beyond five years you are underwriting a GPU generation you cannot forecast, and below three the provider cannot finance the capacity. Anything outside that band needs a structural reason, and if a provider is pushing for more, the price should reflect what you are absorbing.
7. Exit, portability and step-downs. Whether the rate steps down as the market moves, what happens on a change of control, and how you get your data and checkpoints out. A multi-year term without a step-down mechanism is a bet that prices will not fall.

The rate is a marketing number. The definitions are the contract.