GPU Rental Prices: AWS vs GCP vs Lumino Comparison
Paying $8/hour for H100 on AWS? Yeah, we were too. Then we did the math across 50+ training runs. Here is what we found — and why most teams are overpaying b
Comparison | 10 min read | 2026-03-17
Paying $8/hour for H100 on AWS? Yeah, we were too. Then we did the math across 50+ training runs. Here is what we found — and why most teams are overpaying by 30-50% without realizing it.
The Real Cost of GPU Rental (April 2026)
We compared three providers running the same workloads: Llama 3.1 70B fine-tuning, Stable Diffusion XL batch generation, and vLLM inference serving. All tests ran from Mumbai to minimize latency differences. Here are the raw numbers.
H100 80GB - Flagship GPU
A100 80GB - The Workhorse Comparison
This is where the gap gets painful. The A100 is the most commonly rented GPU for fine-tuning and production inference. Here is how the three providers stack up:
Monthly Cost Breakdown: The Numbers That Matter
Hourly rates are useful for comparison. But what you actually care about is the monthly bill. Here is what a typical ML workload costs across providers.
Scenario: You run an A100 for 100 hours/month (typical for a team fine-tuning 2-3 models per week). Here is the total cost including all fees:
You save ₹20,900-25,200/month. That is ₹2.5-3 lakhs/year on a single GPU.
Where AWS and GCP Hide Their Real Costs
The headline price is never the final price. Here are the charges that show up on your bill but never appear on the pricing page.
Data Transfer (Egress) Charges
AWS charges ₹7.50/GB for data egress from Mumbai region. GCP charges ₹6.80/GB. If your training job generates 50GB of checkpoints, model weights, and logs that you need to download, that is ₹340-375 in egress fees alone.
And this is per download. If you pull checkpoints daily for a week-long training run, egress can add ₹2,000-5,000 to your bill.
Lumino: Free data transfer. Upload datasets, download checkpoints, move files around — no per-GB charges. This alone can save ₹5,000-15,000/month for active teams.
Storage Billing When Instances Are Stopped
AWS EBS and GCP Persistent Disks continue billing even when your GPU instance is stopped. A 500GB boot volume costs ₹2,500-3,500/month just sitting there. If you forget to delete it after a project ends, it keeps billing indefinitely.
The trap: Teams stop instances thinking billing is zero. Two weeks later, they find ₹4,000 in storage charges for disks they are not using.
Fix: Always check storage policies. Download your data before terminating. Use object storage (S3, GCS) for long-term checkpoint storage instead of attached disks.
Minimum Billing Increments
AWS bills in 1-hour minimums for most GPU instance types. GCP bills in 1-minute increments. If your fine-tuning job takes 23 minutes, AWS charges you for a full hour. GCP charges for 23 minutes.
Over a month of 50 short test runs, AWS's hourly minimum can add ₹3,000-5,000 in charges for time you never used.
Lumino: Per-second billing. A 23-minute run costs exactly 23 minutes. No rounding up, no minimums.
Performance Comparison: Same GPU, Different Speed
We ran the same Llama 3.1 8B fine-tuning job (10K examples, batch size 8, LoRA) on A100 across all three providers. Here are the results:
| Metric | AWS | GCP | Lumino |
|---|---|---|---|
| Training time | 4h 12m | 4h 08m | 3h 55m |
| Provisioning time | 8-12 min | 5-8 min | 30-60 sec |
| GPU utilization (avg) | 62% | 65% | 71% |
| Total cost (job) | ₹1,428 | ₹1,290 | ₹678 |
The Lumino A100 was 5-7% faster in actual training time. This is likely due to less noisy neighbors (shared infrastructure on AWS/GCP) and faster NVMe storage for dataset loading. The bigger difference is provisioning time — 30 seconds vs 8-12 minutes. Over 20 training runs per month, that is 2.5-4 hours of wasted wait time on cloud providers.
RTX 4090: The GPU AWS Does Not Even Offer
Here is something most comparisons miss: AWS and GCP do not offer RTX 4090 instances. This is a problem because the RTX 4090 is the sweet spot for 70% of AI workloads in India.
RTX 4090 (24GB) at ₹73/hr:
- Fine-tuning 7B-13B models with LoRA
- Stable Diffusion XL batch generation
- vLLM inference serving for 7B models
- Development and testing workloads
Cost comparison: An RTX 4090 at ₹73/hr is 58% cheaper than an A100 at ₹173/hr. For many workloads, the performance difference is negligible. You are paying for VRAM you do not use.
The Verdict: Which Provider Should You Choose?
It depends on your workload size and budget. Here is the honest breakdown:
- Choose AWS if: You need enterprise compliance, multi-region deployment, or already have AWS infrastructure. You will pay 30-50% more, but get integrated services.
- Choose GCP if: You use TPU workloads, need Vertex AI integration, or have GCP credits. Slightly cheaper than AWS but still 40-45% more than Indian providers.
- Choose Lumino (Indian provider) if: You want the lowest cost, per-second billing, free egress, and fast provisioning. Best for startups, individual developers, and teams running frequent training jobs.
Bottom Line
The A100 comparison is the most striking. At ₹173/hr vs ₹340/hr on AWS, you save 49% on the same hardware. Over a 100-hour training run, that is ₹16,700 saved. Over a month of continuous usage, that is ₹1,21,200 saved on a single GPU.
But the real savings come from billing granularity, zero egress fees, and faster provisioning. A 23-minute job should cost 23 minutes, not a full hour. Your checkpoints should download for free. Your instance should be ready in seconds, not minutes.
Try It Yourself
Start with ₹100. Rent an A100 for 30 minutes. Run your workload. Compare the actual bill against what AWS would have charged you. Then decide.
Compare Prices →