GPU Cloud Pricing in India 2026: A100 and H100 Costs Compared Across Major Providers
Side-by-side pricing comparison of major GPU cloud providers in India — TensorCloud, E2E, JarvisLabs, AWS, Azure, GCP, and more. SKU-accurate, with sources.
If you're building AI products in India, GPU compute is likely your biggest infrastructure cost. Whether you're fine-tuning an LLM, running inference at scale, or training a custom model, the choice of GPU cloud provider can mean the difference between burning through your runway in months or stretching it for years.
This guide compares major GPU cloud options available to Indian startups in 2026 — from India-native providers to hyperscaler India regions to global GPU clouds. We look at real pricing, what you actually get for the money, and which provider makes sense for different workloads. We compare GPU pricing for A100 80GB, A100 40GB, and H100 80GB separately because GPU memory size often matters as much as raw GPU performance for LLM fine-tuning and inference.
How we compare prices: All prices shown are public on-demand rates unless noted otherwise. Reserved, spot, community-marketplace and government-subsidized rates are shown separately and clearly marked. GPU SKUs are normalized by memory size and form factor where possible (A100 40GB vs A100 80GB, H100 80GB SXM vs H100 NVL). Prices exclude taxes unless stated. USD conversions use an indicative exchange rate and should be verified before procurement. Sources are listed at the end of this post.
The GPU landscape in India (2026)
India's GPU cloud market is growing rapidly, driven by AI adoption, IndiaAI Mission capacity, and demand from startups, enterprises, and research institutions. The government's IndiaAI Mission has committed Rs 10,371 crore ($1.25B) toward deploying GPUs at scale, with over 38,000 GPUs already available through the programme.
With DPDP compliance, sectoral regulations (especially in BFSI and healthcare), customer security reviews, and procurement preferences becoming more important, many Indian startups prefer India-hosted GPU infrastructure. The result: Indian startups now have real choices beyond just AWS Mumbai.
A100 80GB pricing comparison
The NVIDIA A100 80GB-class GPU remains the workhorse for training and fine-tuning. Here's what a single A100 80GB costs per hour across major providers:
Important: Some hyperscaler A100 instances use the A100 40GB variant, not 80GB. We've separated these to ensure an apples-to-apples comparison. See the A100 40GB section below.
India-native providers (A100 80GB)
| Provider | On-demand (per hour) | 730-hour equivalent | Notes | Source |
|---|---|---|---|---|
| TensorCloud | $1.35 - $1.82 | $986 - $1,329 | Bundled plans with storage, IP included | Internal pricing |
| E2E Networks | ~$2.10 | ~$1,533 | NSE-listed, established provider | Public pricing page |
| JarvisLabs | $1.49 | $1,088 | GPU cloud focused, developer-friendly | Public pricing page |
| Krutrim (Ola) | ~$2.25 on-demand (Rs 189/hr) | ~$1,643 | Lower rates with commitment terms (monthly/6-month/annual). Commitment pricing is not directly comparable to on-demand. | Public pricing page |
Hyperscalers — India region (A100 80GB)
| Provider | Instance type | Per-GPU hourly | 730-hour equivalent | Notes | Source |
|---|---|---|---|---|---|
| GCP | a2-ultragpu-1g (asia-south1) | ~$5.01 | ~$3,657 | Single-GPU available | GCP pricing calculator |
Note: AWS p4d.24xlarge and Azure ND A100 v4 instances in India regions use A100 40GB, not 80GB. These are listed separately below.
Hyperscalers — India region (A100 40GB)
These instances use the A100 40GB variant. Pricing is not directly comparable to A100 80GB — the 40GB variant has half the HBM2e memory, which limits batch sizes for training and restricts which models can be loaded for inference.
| Provider | Instance type | Per-GPU hourly | 730-hour equivalent | Notes | Source |
|---|---|---|---|---|---|
| AWS | p4d.24xlarge (ap-south-1) | ~$4.10 per GPU | ~$2,993 | 8x A100 40GB only, price per GPU | AWS on-demand pricing |
| Azure | ND96asr A100 v4 (Central India) | ~$3.68 per GPU | ~$2,686 | 8x A100 40GB only | Azure pricing calculator |
Global GPU clouds (A100 80GB, no India region)
| Provider | On-demand (per hour) | 730-hour equivalent | Notes | Source |
|---|---|---|---|---|
| Lambda Labs | $1.10 | $803 | US/EU only | Public pricing page |
| RunPod | $1.19 | $869 | Community cloud, spot pricing lower | Public pricing page |
| Vast.ai | $0.70 - $1.20 | $511 - $876 | Marketplace model, variable reliability | Marketplace listing |
Key takeaway: India-native providers offer A100 80GB at significantly lower rates than hyperscaler India regions. TensorCloud's A100-1 plan at $1.82/hr includes 16 vCPUs, 64 GB RAM, 500 GB NVMe storage, and a public IP — specs that would cost significantly more on AWS or GCP when you add storage, networking, and the fact that hyperscaler A100 instances in India are often the 40GB variant.
H100 pricing comparison
The H100 80GB SXM5 delivers roughly 2-3x the inference throughput of the A100, making it the GPU of choice for production inference and large-scale training.
| Provider | H100 80GB SXM (per hour) | 730-hour equivalent | Data Residency | Source |
|---|---|---|---|---|
| TensorCloud | $3.22 | $2,351 | India | Internal pricing |
| E2E Networks | ~$2.90 | ~$2,117 | India | Public pricing page |
| JarvisLabs | $2.69 | $1,964 | India | Public pricing page |
| Krutrim (Ola) | ~$2.54 on-demand (Rs 213/hr) | ~$1,855 | India | Commitment rates available; verify availability |
| Yotta / Shakti Cloud | Enterprise pricing | Contact sales | India | — |
| AWS (p5.48xlarge, Mumbai) | Varies by purchase model | Verify in AWS calculator | India | 8x H100 only; normalized per-GPU pricing varies by offer type |
| Azure (ND H100 v5, Central India) | Varies by offer | Verify in Azure calculator | India | 8x H100 only; pricing depends on region, offer, and reservation terms |
| Lambda Labs | $2.49 | $1,818 | US/EU | Public pricing page |
| CoreWeave | $2.06 | $1,504 | US/EU | Public pricing page |
| RunPod | $2.39 | $1,745 | US/EU | Public pricing page |
Key takeaway: On raw H100 hourly pricing, E2E and JarvisLabs offer lower per-GPU rates than TensorCloud. Where TensorCloud differentiates is in bundled value — the H100-1 plan at $3.22/hr includes 24 vCPUs, 192 GB RAM, 1 TB storage, and a public IP. Hyperscaler H100 pricing in India regions tends to be significantly higher, though exact rates vary by purchase model and should be verified directly.
Subsidized government compute: IndiaAI Mission
IndiaAI compute is available at subsidized rates for eligible users — approved startups, MSMEs, academic institutions, research organizations, and government entities — through the IndiaAI portal. Public government communications have referenced rates as low as Rs 65/GPU-hour, but exact pricing, GPU class, quota, and availability may vary by allocation.
If you qualify, IndiaAI compute is worth exploring for research and experimentation workloads. However, it is not a direct substitute for production cloud infrastructure — availability is limited, there's an approval process, and the platform depth (managed services, networking, storage, databases) is minimal compared to commercial providers.
The hidden costs hyperscalers don't show upfront
GPU-per-hour pricing tells only part of the story. Here's what adds up on hyperscaler bills:
Data egress
AWS charges $0.09/GB for the first 10 TB of data leaving ap-south-1 (after a small free allowance). If you're serving a model that transfers 500 GB/month of inference results, that's $45/month just in egress — on top of your GPU costs. India-native providers typically include egress or charge dramatically less.
Storage
AWS EBS gp3 in Mumbai costs $0.08/GB/month. A 500 GB volume for your training data is $40/month. TensorCloud's bundled plans include NVMe storage at $0.058/GB/month, and many plans include 500 GB - 2 TB.
Public IP addresses
AWS charges $0.005/hr for public IPv4 addresses ($3.65/month each). TensorCloud bundles one public IP with every instance.
Networking and load balancers
AWS ALB minimum cost is ~$22/month plus per-request charges. TensorCloud's load balancers are $0.01/hr ($7.30/month) flat.
Total cost of ownership: a real example
Let's price a realistic AI workload: one A100 80GB GPU instance with 500 GB storage, a public IP, and moderate egress (200 GB/month).
For the hyperscaler comparison, we use the GCP a2-ultragpu-1g as a publicly documented single-GPU A100 80GB hyperscaler option in an India region.
| Cost component | GCP India (A100 80GB) | TensorCloud |
|---|---|---|
| GPU compute (730 hrs) | $3,657 | $1,329 (A100-1 plan) |
| Storage (500 GB SSD) | ~$85 | Included |
| Public IP | ~$3 | Included |
| Data egress (200 GB) | $24 | Included |
| Monthly total | ~$3,769 | $1,329 |
| Annual total | ~$45,228 | $15,948 |
| Annual savings | — | ~$29,280 (65%) |
Even comparing against India-native competitors who offer lower raw GPU rates, TensorCloud's bundled plans often come out ahead on total cost because storage, IP, and platform services are included.
Bundled plans vs. custom pricing
Most GPU cloud providers offer two pricing models:
Pay-per-resource (custom)
You pick vCPUs, RAM, storage, and GPU independently. Good for unusual configurations, but typically 20-25% more expensive than bundled plans.
At TensorCloud, custom rates are:
- vCPU: $0.028/hr
- RAM: $0.0045/GB/hr
- Storage: $0.00008/GB/hr
- A100 GPU: $1.49/hr
- H100 GPU: $2.39/hr
Bundled plans (recommended)
Pre-configured instances with GPU, compute, storage, and networking at a discount. These are optimized for common workloads:
| Plan | GPU | vCPUs | RAM | Storage | Price/hr | Best for |
|---|---|---|---|---|---|---|
| A100-Dev | 1x A100 80GB | 8 | 32 GB | 200 GB | $1.35 | Notebooks, experimentation |
| A100-1 | 1x A100 80GB | 16 | 64 GB | 500 GB | $1.82 | Training, inference |
| A100-2 | 2x A100 80GB | 24 | 128 GB | 1 TB | $3.55 | Multi-GPU fine-tuning |
| A100-4 | 4x A100 80GB | 48 | 256 GB | 2 TB | $6.78 | Distributed training |
| H100-1 | 1x H100 80GB | 24 | 192 GB | 1 TB | $3.22 | Production inference |
| H100-2 | 2x H100 80GB | 48 | 384 GB | 2 TB | $6.24 | High-throughput serving |
Which plan for your workload?
| Workload | Recommended plan | Why |
|---|---|---|
| Learning, notebooks, prototyping | A100-Dev | Low cost, enough for most experiments |
| Fine-tuning 7B-14B parameter models | A100-1 | 80GB VRAM fits most 7B-14B models with LoRA/QLoRA |
| Fine-tuning 32B-70B with QLoRA | A100-2 or A100-4 | Multi-GPU for large models that exceed single-GPU VRAM |
| Production inference (single model) | H100-1 | 2-3x inference throughput vs A100 |
| High-throughput serving (multiple models or high concurrency) | H100-2 | Dual H100 for demanding production workloads |
| Enterprise regulated workloads (BFSI, healthcare) | Custom deployment | Dedicated resources, compliance controls, SLA |
When to use which provider
Choose an India-native provider (TensorCloud, E2E, JarvisLabs) when:
- Data residency is important — DPDP compliance, BFSI regulations, healthcare data, government procurement
- You need INR billing — no FX risk, GST invoices for input tax credit
- You want a full platform — compute + storage + databases + networking + email in one place
- Latency to Indian users is critical — 5-15 ms from India vs 150+ ms from US regions
Choose a hyperscaler (AWS, Azure, GCP) when:
- You need global multi-region — serving users across 20+ regions
- You require specific managed services — SageMaker, Vertex AI, or Azure ML
- Enterprise procurement requires it — some large organizations mandate hyperscaler contracts
- You need 8+ GPUs in a single instance — hyperscalers offer p5.48xlarge (8x H100) as a single machine
Choose a global GPU cloud (Lambda, RunPod) when:
- Lowest possible GPU-hour price is the only priority — and data residency doesn't matter
- You need spot/interruptible instances — for fault-tolerant batch training
- You're experimenting — short-term, no commitment needed
Beyond GPUs: the full-stack advantage
The cheapest GPU per hour isn't always the cheapest total cost. If you use Lambda for GPU, AWS S3 for storage, Supabase for your database, and Resend for transactional email, you're managing four vendors, four billing cycles, and cross-provider data transfer costs.
TensorCloud bundles everything an AI startup needs on one India-hosted platform:
- GPU Instances — A100 and H100 with near bare-metal GPU passthrough
- Jupyter Notebooks — GPU-powered, one-click setup
- LLM Deployments — 254+ open-source models, one-click deploy with OpenAI-compatible API
- Managed PostgreSQL — with pgvector for embeddings
- Object Storage — S3-compatible, 5 GB free
- Block Volumes — NVMe SSD, attach to any instance
- Load Balancers — L4/L7 with SSL termination
- DNS Management — managed zones
- Transactional Email — DKIM-signed, 10K messages free
- Security Groups — per-instance firewalls
No cross-provider data transfer. One invoice. One dashboard. We may not always be the lowest raw GPU-hour provider, but for production workloads, bundled infrastructure and reduced operational complexity make TensorCloud a compelling choice on total cost.
Get started
Not sure which GPU you need? Tell us your model name, expected traffic, context length, and monthly budget. We'll recommend the right TensorCloud configuration and estimate your cost per 1M tokens.
- Sign up at tensordata.com — no credit card required
- Choose a plan — A100-Dev at $1.35/hr is ideal for getting started
- Launch an instance — SSH-ready in under 5 minutes
- Deploy a model — one-click LLM deployment from our 254+ model registry
Or reach out to us directly for a custom quote.
Pricing sources and methodology
| Provider | Source |
|---|---|
| TensorCloud | Internal published pricing |
| E2E Networks | Public pricing page |
| JarvisLabs | Public pricing page |
| Krutrim (Ola) | Public pricing page |
| AWS | EC2 on-demand pricing, instance type docs |
| Azure | Azure pricing calculator, ND A100 v4 docs, ND H100 v5 docs |
| GCP | GCP pricing calculator |
| Lambda Labs | Public pricing page |
| RunPod | Public pricing page |
| CoreWeave | Public pricing page |
| Vast.ai | Marketplace listings |
| IndiaAI | PIB press release, PIB press note |
Pricing changes frequently and may vary by region, availability, commitment term, quota, taxes, exchange rate, and negotiated enterprise discounts. Always verify final pricing with the provider before procurement.
Pricing data last verified: June 2026. All prices are on-demand rates in USD unless otherwise noted. TensorCloud also supports INR billing with GST invoices. This post will be updated quarterly. Subscribe to our RSS feed for updates.
AWS, Amazon Web Services, Azure, Google Cloud Platform, NVIDIA, Lambda, RunPod, CoreWeave, E2E Networks, JarvisLabs, Krutrim, Yotta, and Vast.ai are trademarks of their respective owners. TensorCloud is not affiliated with or endorsed by any of these companies.
Ready to build with GPU cloud?
Get started with TensorCloud — A100 & H100 GPUs, one-click LLM deployment, and full-stack AI infrastructure.
Get Started Free