Cloud GPU Cost Calculator
Cloud GPU pricing seems straightforward: $X/hour. But effective cost depends on utilization, spot pricing, and hidden costs. This guide reveals the true math.
Headline vs Effective Cost
Hourly Rates
| GPU | Hourly | Daily (24h) | Monthly |
|---|---|---|---|
| T4 | $0.50 | $12 | $365 |
| A10G | $1.00 | $24 | $730 |
| L40S | $2.50 | $60 | $1,825 |
| A100 | $3.00 | $72 | $2,190 |
The Utilization Problem
You pay for the hour, not the work:
100 jobs/day × 5 min processing = 8.3 hours of work
You pay for: 24 hours (if always-on)
Utilization: 35%
Effective cost: 2.9× headline rateEffective Cost at Different Utilization
| Utilization | A10G Effective Rate | Cost per 5-min Job |
|---|---|---|
| 100% | $1.00/hr | $0.083 |
| 70% | $1.43/hr | $0.119 |
| 50% | $2.00/hr | $0.167 |
| 30% | $3.33/hr | $0.278 |
Low utilization makes GPUs expensive fast.
Spot/Preemptible Pricing
Spot Discounts
| GPU | On-Demand | Spot | Savings |
|---|---|---|---|
| T4 | $0.50 | $0.15 | 70% |
| A10G | $1.00 | $0.35 | 65% |
| L40S | $2.50 | $0.90 | 64% |
| A100 | $3.00 | $1.20 | 60% |
The Catch
Spot instances can be interrupted:
- 2-minute warning before termination
- Job may need to restart
- Availability varies by region/time
When to Use Spot
- Batch processing (not real-time)
- Jobs with checkpointing
- Fault-tolerant workloads
When to Avoid Spot
- SLA-bound workloads
- Long jobs without checkpointing
- Customer-facing latency requirements
Serverless GPU Alternative
Per-Second Billing
Services like Modal, RunPod, Banana bill per second of actual compute:
- No idle costs
- Scale to zero
- Pay only for processing
Example: Modal Pricing
| GPU | Per-Second | Per-Hour | 5-min Job |
|---|---|---|---|
| T4 | $0.00028 | $1.00 | $0.083 |
| A10G | $0.00077 | $2.77 | $0.23 |
| A100 | $0.0011 | $3.96 | $0.33 |
Trade-offs
- Pro: No idle cost, true pay-per-use
- Con: Cold start (5-30 seconds)
- Con: Higher per-second rate than reserved
When Serverless Wins
- Variable workloads
- Low to moderate volume
- No infrastructure management desire
Total Cost of Ownership
Beyond GPU Compute
| Cost Category | % of Total | Notes |
|---|---|---|
| GPU compute | 40-60% | Depends on utilization |
| Engineering | 20-40% | Build and maintain |
| Storage | 5-15% | Input/output video |
| Egress | 5-10% | Delivery bandwidth |
| Monitoring | 2-5% | Tools and dashboards |
Example: 10K Videos/Month
GPU (A10G, 70% utilization): $1,000/month
Storage (500GB): $100/month
Egress (1TB): $90/month
Monitoring: $100/month
Engineering (10 hrs): $500/month
Total: $1,790/month
Per video: $0.179Compare to API
BetterVideo: $0.20/video × 10K = $2,000/month
Difference: $210/month
For: No infrastructure, no engineering, no on-callThe "savings" from self-hosting often evaporate when you include everything.
Frequently Asked Questions
Headline rate divided by utilization. At 50% utilization, a $1/hr GPU effectively costs $2/hr. Low utilization kills economics.
For batch processing that can handle interruptions, yes — 60-70% savings. Not for SLA-bound or customer-facing workloads.
At low utilization, yes — no idle cost. At high, consistent utilization, reserved instances may be cheaper.
Ready to get started?
Try BetterVideo's privacy-first video enhancement API — free sandbox, no credit card required.