Formula
8 GPUs × 10 hours = 80 GPU-hours
GPU-hours measure time-based access to accelerators, much like kilowatt-hours measure energy use over time.
← Back to Compute College
Compute College
What is a GPU-hour? Meaning, formula, and cost
One GPU running for one hour: the formula behind AI rental cost and cloud price comparisons.
Plain-English definition
A GPU-hour is one GPU running for one hour. GPU-hours = GPU count x runtime, so 8 GPUs used for 10 hours equals 80 GPU-hours. Multiply that total by the quoted per-GPU hourly rate to estimate raw accelerator cost before storage, networking, utilization loss, or platform fees.
Memory trick: GPU-hour = one GPU available for one hour, like a kilowatt-hour measures electricity used through time.
Why it matters
- Time-based access to an accelerator.
- A common way to discuss rental pricing across providers.
- A starting point for utilization, revenue, and capacity analysis.
Simple example
If a workload uses 8 GPUs for 10 hours, it consumes 80 GPU-hours. Multiply that result by the quoted rate to estimate raw accelerator cost before storage, networking, utilization loss, or platform overhead.
Any figures shown are illustrative calculations, not current quoted market prices.
Common mistake
A GPU-hour tells you how long capacity is available, not how powerful that capacity is. Chip generation, memory, networking, region, commitment length, and bundled services can all change the value of one GPU-hour versus another.
- Chip generation and memory configuration matter.
- Networking, storage, support, and software stack can change effective value.
- Commitment length, region, and availability can move pricing even within the same chip family.
Practical takeaway
What you can do with this
Normalize AI compute quotes into GPU-hours before comparing them, then record the GPU type, region, network, reliability terms, utilization, and additional fees that change useful cost.
- Buyers: multiply GPUs by runtime and hourly rate for a common starting comparison.
- Founders and analysts: distinguish paid GPU-hours from useful output delivered by those hours.
- For example, an illustrative reservation of 8 GPUs for 10 hours represents 80 GPU-hours; multiplying by an illustrative $5 rate produces $400 before storage, networking, or platform overhead.
- Next compare useful output: a slower run, interruption, weak interconnect, or unused reservation time can increase the effective cost even when the posted GPU-hour rate is lower.
- Keep provider observations separated from calculated workload estimates. A calculated bill helps make a decision, but it is not a new observed market price.
- A procurement comparison should also state whether capacity is immediately available, reserved for later delivery, or interruptible, since identical GPU-hour arithmetic can represent different operational risk and different useful output.
Decision check: a cheaper GPU-hour is meaningful only when it can complete the relevant workload on acceptable terms.
Compute College track
AI Compute 101
Step 3 of 7: What is a GPU hour