On-demand vs reserved vs spot pricing
Compare contract choices.
← Back to Compute College
Compute College
What is a Compute Reservation?
A compute reservation secures defined GPU or accelerator capacity for a buyer over an agreed period.
Plain-English definition
A compute reservation is an agreement that gives a buyer access to a defined amount of compute capacity for a specified period. For AI workloads, a reservation may secure GPUs or a connected cluster before the buyer runs training, serves users, or needs overflow capacity.
Memory trick: A compute reservation is a seat saved on the GPU train: it helps you board on time, but you may pay even if plans change.
Why it matters
Reservations show how much buyers value certainty. A company that cannot risk waiting for GPUs may commit earlier, accept minimum payments, or choose a provider based on reliable availability rather than the lowest visible hourly rate. That choice affects cost and removes capacity from the open market.
- A training deadline can make guaranteed simultaneous cluster access more valuable than flexible on-demand use.
- A production service may need predictable capacity to protect latency and uptime even if utilization varies.
- Reserved blocks can tighten available supply for other buyers while giving providers more predictable demand.
- Unused commitments can become stranded cost when forecasts, models, or product demand change.
Simple example
Consider an illustrative startup expecting to use 64 H100 GPUs for 720 hours in one month. At an on-demand rate of $8 per GPU-hour, full-month usage would equal 64 x 720 x $8 = $368,640. A reservation at an illustrative $6.50 rate would equal $299,520 if the full commitment is required and used.
- At full use, the illustrative reservation is $69,120 lower for that month before other overhead.
- If the buyer needs only half of the committed hours but still pays for the full block, the effective useful-hour cost rises sharply.
- Availability, GPU generation, networking, support, and cancellation terms must be comparable before treating the difference as savings.
- These numbers illustrate commitment math; they are not current observed quotes.
Example figures are illustrative calculations, not current quoted market prices.
Common mistake
Do not assume a reserved rate is automatically cheaper than on-demand capacity. A reservation buys access and possibly a discount, but it also creates obligation. If demand arrives later than expected, a model architecture changes, or the reserved system does not fit the workload, the buyer can pay for unused or ineffective capacity.
Practical takeaway
What you can do with this
Compare reservation choices using expected utilization, critical deadlines, workload reliability needs, and the cost of being unable to run. A buyer should model base, low-use, and high-demand cases before committing to a capacity block.
- Procurement teams: capture GPU type, connected quantity, term, minimum payment, region, support, SLA, renewal, and exit rights.
- Founders: reserve capacity when predictable demand or launch risk justifies paying for certainty.
- Product managers: distinguish recurring serving baseload from occasional experiments or overflow capacity.
- Analysts: watch whether long commitments absorb availability that would otherwise appear in open rental markets.
Decision check: approve a reservation only after showing the utilization level at which it beats flexible capacity and the consequence if capacity is unavailable without it.
Compute College track
Compute Market Structure
Step 4 of 8: What is a compute reservation