Buy vs Rent GPUs in India: A Cost Analysis

One of the most important infrastructure decisions for Indian AI companies is whether to buy GPU hardware outright or rent it from cloud providers. The answer depends on your utilisation rate, timeline, and operational maturity. Here is a real-numbers analysis for 2026.

Cloud GPU Pricing in India

The major cloud providers offer NVIDIA A100 and H100 instances in their Mumbai (and increasingly, Hyderabad) regions. Typical on-demand pricing as of early 2026:

A100 80GB (single GPU): INR 200-300/hour on-demand
H100 80GB (single GPU): INR 400-600/hour on-demand
8x H100 instance: INR 3,000-4,500/hour on-demand

Reserved instances (1-year commitment) reduce these by 30-40%, and spot/preemptible instances offer 60-70% discounts but with no availability guarantees.

Owned Hardware Costs

For a single 8x H100 SXM5 server (Supermicro or equivalent), the total cost of ownership over 3 years looks like this:

Cost Component	Amount (INR)
Server hardware (8x H100 HGX, CPUs, RAM, NVMe, networking)	2.5 crore
Colocation (full rack, 15kW power, Tier 3 Mumbai)	30 lakh/year (90 lakh over 3 years)
Electricity (10kW average, INR 10/kWh)	26 lakh/year (78 lakh over 3 years)
Network bandwidth (1 Gbps commit)	5 lakh/year (15 lakh over 3 years)
Maintenance and spares	10 lakh/year (30 lakh over 3 years)
3-year TCO	~4.6 crore

This works out to approximately INR 1,750/hour for 8x H100 GPUs assuming 24/7 availability (8,760 hours/year x 3 years = 26,280 hours).

The Break-Even Calculation

Comparing owned hardware at INR 1,750/hour against cloud on-demand pricing at INR 3,500/hour for 8x H100:

Break-even point: approximately 15-18 months at continuous utilisation
At 50% utilisation (12 hours/day), break-even extends to roughly 30-36 months
At 25% utilisation (6 hours/day), buying rarely makes financial sense within 3 years

The key variable is GPU utilisation rate. If your GPUs will run 18+ hours per day consistently, buying is almost always cheaper. If utilisation is sporadic, cloud is more cost-effective.

Beyond Raw Cost: Hidden Factors

Several factors tilt the decision beyond simple hourly cost:

Advantages of buying:

No per-hour anxiety: teams experiment more freely when compute is not metered
Data sovereignty: your training data stays on your own hardware in an Indian data centre
Customisation: you control the exact OS, driver, networking, and storage configuration
Residual value: GPU servers retain 30-50% of their value after 3 years

Advantages of renting:

Zero capex: no upfront investment of INR 2-3 crore
Elastic scaling: spin up 64 GPUs for a training run, release them when done
Managed infrastructure: no need for data-centre relationships, spare parts inventory, or hardware expertise
Global availability: train in any cloud region with lowest latency to your data

The Hybrid Approach

Many successful Indian AI companies use a hybrid strategy:

Own a base cluster (e.g., 8-16 GPUs) for day-to-day development, fine-tuning, and inference serving
Burst to cloud for large training runs that need 64-256 GPUs for a few days to weeks
Use spot instances for hyperparameter sweeps and experiments where preemption is acceptable

This approach captures the cost savings of ownership for sustained workloads while leveraging cloud elasticity for peak demand.

Our Recommendation

If you expect to run GPUs at 50%+ utilisation for 2+ years, buying hardware and colocating it in India is significantly more cost-effective. Rawcompute.in makes the buying path easier by handling hardware sourcing, import logistics, GST, configuration, delivery to your colocation facility, and ongoing support. Contact us for a TCO comparison customised to your specific workload profile.