If you're running ML training from India, you've hit the pricing wall. AWS Mumbai GPU instances cost 15-20% more than us-east 1. And availability is inconsistent.
Here's how the cost breakdown works for a typical team doing LLM fine-tuning:
Cloud GPU (8x H100, AWS ap-south-1): ~$98/hr on-demand
Cloud GPU (8x H100, AWS us-east-1): ~$85/hr on-demand
Dedicated colo (8x H100, Memphis): ~$32/hr effective (3-year TCO)
The 3x cost difference isn't magic. It's facility economics. Memphis: $0.07/kWh power, $4.98M/MW facility cost, zero state income tax. Mumbai: $0.12/kWh, $8-10M/MW facility cost, 18% GST on services.
For teams running sustained workloads (not burst), the break-even on dedicated infrastructure is 8-12 months. After that, every GPU-hour is pure savings.
Not saying everyone should move to Memphis. But the math is worth running if your monthly cloud bill crossed $50K.
Top comments (0)