You're Probably Comparing GPU Rentals the Wrong Way
If the first number you compare is hourly price, there is a good chance you are optimizing for the wrong thing. We analyzed 200+ GPU rental decisions and fou
Pricing | 10 min read | 2026-03-21
If the first number you compare is hourly price, there is a good chance you are optimizing for the wrong thing. We analyzed 200+ GPU rental decisions and found that teams who compare only hourly rates end up paying 30-60% more than teams who compare total job cost. Here is the right way to do it.
The Default Comparison Everyone Uses
Open two pricing pages. Find the hourly number. Pick the smaller one. It feels rational, but it hides the variables that actually decide what the job will cost and how painful the workflow will be.
This is like comparing cars by their sticker price without considering fuel efficiency, maintenance costs, insurance, or how long they last. The cheapest car to buy can be the most expensive to own. The same is true for GPU rentals.
What Most People Miss
1. Time-to-result matters more than hourly rate
A faster GPU with a higher rate can still finish cheaper than a slower GPU with a lower rate. The total cost is rate × time, not just rate.
Example: Fine-tuning a 13B model:
- • RTX 4090 at ₹73/hr: takes 3 hours → ₹219 total
- • A100 at ₹173/hr: takes 2 hours → ₹346 total
- • H100 at ₹583/hr: takes 1 hour → ₹583 total
The RTX 4090 has the lowest hourly rate AND the lowest total cost. The H100 has the highest hourly rate AND the highest total cost. But the A100 is 58% more expensive per hour than the 4090 and only 33% faster — making it the worst value of the three for this workload.
2. VRAM fit matters more than brand name
If the workload fits on a 4090, jumping to an H100 too early is usually a budget mistake, not a smart optimization. The H100 costs 8x more per hour. If the 4090 can do the job, the H100's extra speed rarely justifies 8x the cost.
Rule: Always start with the smallest GPU that can fit your model. Only upgrade if the smaller GPU cannot complete the job or takes unacceptably long.
3. Storage rules can completely change the bill
Many users stop a pod and think billing has stopped. That assumption gets expensive fast. AWS charges ₹7.50/GB/month for EBS storage. A 500GB disk costs ₹3,750/month just sitting there. If you forget to delete it after a project ends, it keeps billing indefinitely.
Check: Does the provider charge for storage after stopping? What is the per-GB rate? Can you delete storage independently of the GPU instance?
4. Billing granularity changes experiment cost
Short runs, failed runs, and iterative tuning all get punished if billing rounds aggressively. A provider that bills per hour will charge you ₹173 for a 5-minute test. A provider that bills per second will charge you ₹1.44. Over 50 tests per month, the difference is ₹8,650 vs ₹72.
Rule: If you run frequent short tests (under 30 minutes), per-second billing saves you 90-95% on test costs. This is non-negotiable for experimentation-heavy workflows.
5. Egress fees are the silent budget killer
AWS charges ₹7.50/GB for data egress from Mumbai. GCP charges ₹6.80/GB. If your training job generates 100GB of checkpoints, model weights, and logs that you need to download, that is ₹680-750 in egress fees alone. Over a month of daily downloads, that is ₹20,000-22,500.
Check: Does the provider charge for data transfer out? If yes, calculate your expected monthly egress and add it to the hourly cost. The "cheaper" provider may become the more expensive one after egress fees.
The Better Comparison Framework
Here is the step-by-step process that actually works. Do these in order — do not skip steps.
- Pick the smallest GPU that comfortably fits the workload.
Calculate: model weights + optimizer states + gradients + activations + batch size + context length = total VRAM needed. Pick the smallest GPU that fits this with 20% headroom.
- Estimate total runtime, not just hourly rate.
Run a small test (100 examples) on the GPU. Measure time per step. Extrapolate to full dataset. Total cost = runtime × hourly rate.
- Check storage behavior after stop.
Does storage keep billing? At what rate? Can you delete it independently? Factor this into your total cost.
- Look at setup time and workflow friction.
How long from payment to SSH access? Is the environment pre-configured or do you need to install everything? Setup time is billed time.
- Check egress fees.
How much does it cost to download your checkpoints and data? Calculate expected monthly egress and add to total cost.
- Only then compare effective total cost.
Effective cost = (runtime × rate) + storage + egress + provisioning waste. Compare this number across providers, not the hourly rate.
Real Comparison: Three Providers, One Workload
Here is what happens when you compare three providers using the full framework, not just hourly rates.
| Cost component | Provider A (AWS) | Provider B (GCP) | Provider C (Lumino) |
|---|---|---|---|
| GPU hourly rate (A100) | ₹340/hr | ₹310/hr | ₹173/hr |
| Job runtime | 4.5 hrs | 4.2 hrs | 3.9 hrs |
| Compute cost | ₹1,530 | ₹1,302 | ₹675 |
| Provisioning time (billed) | 10 min (₹57) | 6 min (₹31) | 30 sec (₹1.44) |
| Storage (500GB, 1 week) | ₹875 | ₹800 | ₹0 (included) |
| Egress (50GB downloads) | ₹375 | ₹340 | ₹0 (free) |
| Total effective cost | ₹2,837 | ₹2,473 | ₹676 |
Provider A (AWS) looks competitive at ₹340/hr. But after adding provisioning, storage, and egress, the total effective cost is ₹2,837 — 4.2x more than Provider C (Lumino) at ₹676. The hourly rate told one story. The total cost told a very different one.
The One Question to Ask
Not "Which provider is cheapest per hour?" The better question is "Which option gets this workload done with the lowest total cost and least friction?" That usually leads to a very different answer.
Quick decision checklist
- ☐ Does the GPU have enough VRAM for my model + batch size?
- ☐ What is the total job cost (runtime × rate), not just the hourly rate?
- ☐ Does the provider charge for storage after stopping?
- ☐ What is the egress fee per GB?
- ☐ How long does provisioning take? Is it billed?
- ☐ Is billing per-second, per-minute, or per-hour?
- ☐ What is the effective total cost including all fees?
The Bottom Line
Hourly rate is the easiest number to compare. It is also the least useful. The real cost of a GPU rental is the sum of compute time, provisioning time, storage, egress, and idle time. Teams that compare only hourly rates consistently overpay by 30-60%. Teams that compare total effective cost consistently find the cheapest option — and it is rarely the one with the lowest headline price.
Before your next GPU rental, run the full comparison framework. Calculate the effective total cost. Then decide. Your budget will thank you.
Compare Smarter
Look at the workload, the billing model, and the stop-state behavior before you trust the hourly rate. Start with ₹100 and measure the real cost.
Browse GPUs