Table of Contents
RunPod is the dominant per-second GPU rental marketplace. GigaGPU rents the same RTX 4090 hardware by the month. Same card, different billing model. Which is cheaper depends entirely on how much you use it.
If you use a 4090 more than ~3.7 hours per day on average, GigaGPU dedicated is cheaper. Below that, RunPod per-second is cheaper. For 24/7 production inference, dedicated is dramatically cheaper. For occasional fine-tuning, RunPod is the right choice.
Current rates
- RunPod RTX 4090 (community cloud): ~$0.34/hour = ~£0.27/hour
- RunPod RTX 4090 (secure cloud): ~$0.69/hour = ~£0.54/hour
- GigaGPU RTX 4090 dedicated: £289/month flat
RunPod community cloud uses partner-operator hardware (Vast-style); secure cloud uses RunPod’s own datacenters.
Break-even hours-per-day
Per month: 30 days × 24 hours = 720 GPU-hours available.
Against RunPod community cloud (£0.27/hour):
- £289 / £0.27 = ~1,033 hours = more than a month. Dedicated wins at any utilisation.
Against RunPod secure cloud (£0.54/hour):
- £289 / £0.54 = ~517 hours/month = ~17 hours/day. Dedicated wins above 17h/day usage.
- If you use it 24/7, dedicated saves ~£110/mo.
Against AWS g6e.xlarge (similar 24 GB GPU at ~£0.95/hour):
- £289 / £0.95 = ~294 hours/month = ~9.7 hours/day. Dedicated wins above 10h/day.
By workload type
| Workload pattern | Daily GPU usage | Cheaper option |
|---|---|---|
| 24/7 production inference | 24h | Dedicated (saves 50%+) |
| Business-hours chatbot | ~12h | Dedicated (saves 30%) |
| Nightly batch jobs | ~6h | Tied — depends on RunPod tier |
| Occasional fine-tuning | ~2h | RunPod per-second |
| Weekly experiments | ~1h | RunPod per-second |
| Bursty image-gen API | Variable | RunPod Serverless |
Verdict
- Steady production workload: GigaGPU dedicated. Cheaper, no cold start.
- Long fine-tuning runs: GigaGPU dedicated. Per-hour billing for 12-hour training jobs is brutal.
- Occasional / experimental: RunPod per-second. Don’t pay for idle.
- Spiky inference: RunPod Serverless or hosted API.
- Need data residency: GigaGPU (UK datacenter) regardless.
Bottom line
The honest break-even is 3-4 hours of GPU-time per day. Above that, dedicated is cheaper. For most production teams the answer is dedicated; for ML researchers and one-off jobs, RunPod is the right tool. See serverless vs dedicated for the broader analysis.
