Current market-vs-spot spread
See the latest daily CT-MKT vs CT-SPOT gap on the Market page.
← Back to Compute College
Compute College
On-Demand vs Reserved vs Spot GPU Pricing
On-demand, reserved, and spot GPU pricing are three ways buyers obtain and pay for AI compute capacity.
Plain-English definition
On-demand vs reserved vs spot GPU pricing describes three ways to pay for AI compute capacity. On-demand capacity is flexible, reserved capacity exchanges commitment for access or discount, and spot capacity is typically discounted because it may be reclaimed or interrupted.
Memory trick: On-demand is a hotel room, reserved is a lease, and spot is a standby ticket: lower certainty can mean a lower price.
Why it matters
Contract type changes both buyer risk and what a displayed price means. The mix of buyers choosing flexible capacity, longer reservations, or interruptible supply reveals how much certainty is worth in a market shaped by GPU availability, power, and workload deadlines.
- On-demand rates are visible and flexible, but may be expensive for steady usage.
- Reserved terms can show buyers locking in supply because future availability matters.
- Spot pricing can expose slack or scarcity in capacity suited to interruptible work.
Simple example
Assume a job needs 8 GPUs for 100 hours, or 800 GPU-hours. At an illustrative on-demand rate of $8 per hour it costs $6,400. Reserved at $6.50 costs $5,200 if the commitment fits. Spot at $3 costs $2,400 only if interruptions do not force costly delay or repeated work.
- On-demand: pay for flexibility when timing or scale is uncertain.
- Reserved: pay or commit for dependable access to known recurring demand.
- Spot: accept interruption risk for checkpointable or deferrable workloads.
Example figures are illustrative calculations, not current quoted market prices.
Common mistake
A beginner mistake is assuming the lowest quoted rate wins. Spot capacity can be expensive when interruptions waste engineering time, miss a deadline, or cause a run to restart. Reserved capacity can be expensive when the buyer overcommits and leaves GPUs idle.
Practical takeaway
What you can do with this
Match contract type to workload criticality and predictability. Flexible experiments may suit on-demand access; dependable production serving or scheduled training may justify reservation; fault-tolerant batch jobs may use spot when savings exceed interruption cost.
- Procurement teams: request interruption rules, term length, availability guarantees, and minimum use alongside rate.
- Founders: decide which workloads can wait or restart without damaging customer commitments.
- Analysts: follow pricing spreads and reservation behavior as different views of capacity tightness.
- ML teams: checkpoint interruptible jobs and price the storage, recovery, and delay needed to make spot usable.
- Finance teams: treat unused reserved capacity as cost when comparing a committed deal with flexible buying.
Decision check: assign each workload a tolerance for delay, interruption, and unused commitment before selecting the price type that appears cheapest.
Compute College
Turn the lesson into a number
Use the GPU-Hour Cost Calculator, AI Training Cost Calculator, or Model Serving Cost Calculator.
Compute College track
GPU Pricing & Capacity
Step 5 of 8: On demand vs reserved vs spot GPU pricing