IND
UAE
SA
Most articles about GPU cloud pricing are written for finance teams who want a number on a spreadsheet. This one is written for the AI engineer or CTO who needs to actually choose a pricing model — and not overpay while doing it.
The difference between hourly and reserved pricing for a team running 8 × H100s continuously is roughly Rs 60 lakh per year. That’s not a rounding error. It’s a senior hire. It’s a product launch. It’s the kind of decision that deserves more than a bullet-point chart — which is exactly what this guide gives you.
The Four GPU Pricing Models at a Glance
Before getting into the nuances of each model, here is a side-by-side summary. Think of this as your quick reference — the rest of the guide fills in the details that actually matter for your workload.
| Model | How You Pay | Commitment | Discount vs On-Demand | Best For |
|---|---|---|---|---|
| On-Demand / Hourly | Per GPU per hour, start/stop any time | None | Baseline | Experiments, variable workloads, one-off jobs |
| Reserved | Fixed monthly or annual contract | 1–12 months | 30–50% cheaper | Sustained production inference, ongoing training |
| Spot / Preemptible | Per hour on surplus capacity — may be interrupted | None | Up to 70% cheaper | Fault-tolerant batch jobs, distributed training with checkpointing |
| Dedicated | Entire physical server reserved for you only | Monthly / annual | Premium over shared | Regulated industries, compliance, maximum performance |
No single pricing model is universally best. The right choice depends on one number: your average GPU utilisation as a percentage of committed capacity. Below 60%? Hourly wins. Above 70% continuously? Reserved wins. Variable mix of both? Hybrid wins. The sections below show you exactly how to calculate this for your own workload.
On-Demand / Hourly Pricing
GPU as a Service on-demand pricing is the default model for almost every cloud provider. You spin up an instance, pay for the hours it runs, and shut it down when you’re done. Simple, flexible, and — for the right workload — genuinely cost-efficient.
No Commitment, No Minimum Spend
Launch a single H100 for two hours to test a fine-tuning script, then shut it down. You pay for two hours. No monthly minimums, no setup fees, no penalties for stopping early. This is the pricing model that lets a two-person startup in Pune run the same hardware as a Fortune 500 enterprise — just for shorter windows.
Perfect for Variable and Unpredictable Workloads
Research teams running ad-hoc experiments, startups in early development, companies doing proof-of-concept work, and any team whose GPU needs spike and fall unpredictably all benefit from hourly pricing. You never pay for capacity you’re not actively using — and with 68% of on-premise GPU resources sitting idle in typical enterprise deployments, that matters.
Highest Per-Hour Rate of Any Model
The flexibility premium is real. On-demand rates are the highest per-GPU-per-hour of any pricing model. For a team running GPUs 18+ hours per day, every day, the on-demand rate quickly becomes the most expensive option available — often 2× what a reserved commitment would cost for identical compute. Workloads on the A100 GPU or L40S GPU at sustained utilisation are especially strong candidates for reserved pricing.
Bills Can Swing Dramatically Month to Month
CFOs dislike on-demand GPU billing for the same reason engineers love it: the numbers are unpredictable. A month with two large training runs can cost 5× a quieter month. If your finance team needs a stable line item, on-demand billing creates friction — unless you pair it with cost alert thresholds and auto-shutdown policies in your MLOps platform.
Your utilisation is below 60% of what you would commit to in a reserved plan, your workloads are short-duration or one-off, you are in early experimentation or PoC phase, or you need to scale dramatically for a one-time event — a product launch, a competition, a hackathon — and then scale back down immediately.
Reserved Instance Pricing
Reserved pricing is what turns GPU cloud from a flexible experiment into genuinely cost-efficient production infrastructure. By committing to a GPU instance for a fixed period — typically 1, 3, 6, or 12 months — you unlock discounts of 30–50% off the on-demand rate.
The maths are straightforward. An H100 at Rs 219/hr on-demand running continuously for a year costs roughly Rs 191 lakh. The same instance on a 12-month reserved commitment at 50% off costs Rs 95 lakh. That Rs 96 lakh difference is the cost of flexibility — and for teams with predictable production workloads, it is simply wasted money. The same logic applies to the A100 80GB: at Rs 170/hr on-demand versus approximately Rs 85–100/hr reserved, the savings compound rapidly at scale.
| Commitment Length | Typical Discount | Best For | Risk Level |
|---|---|---|---|
| 1 Month | 10–15% | Teams with 4–6 weeks of confirmed roadmap | Low — short commitment |
| 3 Months | 20–25% | Quarterly sprint cycles, product launches | Low – Medium |
| 6 Months | 30–40% | Stable production inference, ongoing training projects | Medium |
| 12 Months | 40–50% | Core AI infrastructure with predictable long-term utilisation | Medium — requires utilisation forecasting |
Reserved pricing only wins if you actually use the capacity you commit to. If your utilisation drops to 40% of your reserved commitment — due to a project delay, team restructure, or model architecture change — you are paying for 60% idle capacity. Before committing, gather at least 3 months of historical usage data and build a realistic forecast for the commitment period.
You have sustained workloads running above 70% utilisation for the commitment period, you have production inference serving that runs 24/7, you have completed the experimentation phase and have a stable model architecture, or your finance team requires predictable monthly compute costs for budget planning.
Spot / Preemptible Pricing
Spot instances are the best deal in GPU cloud — for teams that know how to use them. Providers offer unused capacity at discounts of up to 70% off on-demand rates. The catch: the instance can be interrupted with short notice (typically 30 seconds to 2 minutes) when the provider needs that capacity back for on-demand or reserved customers.
This sounds scarier than it is in practice. Modern ML frameworks — PyTorch, JAX, DeepSpeed — all support checkpointing, which saves your training state to storage at regular intervals. If a spot instance is interrupted, you lose at most a few minutes of computation, not hours. Teams that build this into their training pipelines routinely run large distributed training jobs on A100 GPU spot instances, saving tens of lakhs annually without any meaningful increase in total training time.
✅ Spot Works Well For
- Large training runs using checkpointing every 10–30 minutes
- Distributed training where one node failure doesn’t kill the run
- Batch inference pipelines processing queued jobs
- Data preprocessing and feature engineering at scale
- Hyperparameter search and neural architecture search
- Any workload where a 5–10 minute restart is acceptable
🚫 Spot Does Not Work For
- Real-time inference APIs serving live production traffic
- Interactive notebooks or active development environments
- Long training runs without any checkpointing in place
- Stateful applications that cannot handle graceful termination
- Regulated workloads requiring guaranteed uptime SLAs
Dedicated GPU Pricing
Dedicated GPU instances give you an entire physical server — not a virtualised slice of one. No other customer’s workloads run on the same hardware. This matters in two specific situations: when you need maximum, consistent performance without the “noisy neighbour” effect of shared infrastructure, and when your compliance requirements mandate physical isolation of your data. Dedicated configurations are available across all GPU tiers — from the L40S for graphics and inference workloads to the A100 for regulated production deployments.
DPDP & Regulatory Compliance
BFSI and healthcare organisations processing sensitive personal data under the DPDP Act 2023 often require dedicated instances with full audit trails, Data Processing Agreements, and physical data isolation. Shared multi-tenant instances cannot meet these requirements.
Maximum Consistent Performance
Shared GPU instances can experience throughput variability depending on other tenants’ workloads. Dedicated instances deliver predictable, peak-level performance — critical for latency-sensitive production inference and time-bound training jobs with hard deadlines.
Data Sovereignty & IP Protection
Your model weights, training data, and inference outputs never touch hardware that other organisations have access to. For IP-sensitive work — proprietary model architectures, confidential business data — dedicated instances provide the strongest data isolation available in cloud.
Cyfuture AI India GPU Rates (2026)
Here are the on-demand rates for Cyfuture AI’s GPU cloud — India’s leading GPUaaS platform with data centres in Mumbai, Noida, and Chennai. All pricing is in Indian Rupees and includes full NVLink interconnect within nodes and InfiniBand HDR networking for multi-node clusters.
Multiply any on-demand rate by 0.5–0.6 to estimate your effective hourly cost on a 12-month reserved commitment. An H100 at Rs 219/hr on-demand becomes approximately Rs 110–130/hr on a 12-month reserve. An A100 80GB drops from Rs 170/hr to approximately Rs 85–100/hr. For a team running 8 × H100s continuously, that is a saving of over Rs 50 lakh per year — before accounting for reserved pricing being 51% cheaper than AWS to begin with.
Cyfuture AI vs AWS & GCP: Price Comparison
For Indian teams, the location of your GPU cloud matters beyond compliance. India-hosted GPU cloud eliminates data egress fees — which can add Rs 5–15 lakh per year on large training datasets transferred to and from foreign data centres. Here is how Cyfuture AI’s on-demand rates compare to AWS ap-south-1 and GCP Mumbai:
| GPU | Cyfuture AI (India) | AWS (ap-south-1) | GCP (Mumbai) | Savings vs AWS |
|---|---|---|---|---|
| A100 80GB | Rs 170/hr (∼$2.03) | ∼$3.20/hr | ∼$2.93/hr | ∼37% cheaper |
| H100 SXM | Rs 219/hr (∼$2.62) | ∼$5.40/hr | ∼$4.80/hr | ∼51% cheaper |
| L40S | Rs 61/hr (∼$0.73) | ∼$1.60/hr | ∼$1.40/hr | ∼54% cheaper |
| V100 | Rs 39/hr (∼$0.47) | ∼$1.00/hr | ∼$0.90/hr | ∼53% cheaper |
Training datasets for large language models routinely run into terabytes. At AWS’s standard data egress pricing, transferring 10 TB of training data from a foreign region costs approximately $1,200 — every single time you run the job. India-hosted GPU cloud eliminates this entirely. Add DPDP Act compliance obligations for regulated sectors, and the total cost advantage of Cyfuture AI vs hyperscalers is often 50–70% when all costs are factored in.
Calculate Your Exact GPU Cost — Then Launch in Under 60 Seconds
H100 from Rs 219/hr. A100 from Rs 170/hr. L40S from Rs 61/hr. No minimums, no procurement delays, no data centre headaches. India-hosted and DPDP-compliant from day one.
Real Cost Scenarios: Hourly vs Reserved
Abstract percentages only go so far. Here are three real-world scenarios with actual Indian Rupee numbers to help you see which model applies to your situation — and how much the right choice actually saves.
Scenario 1 — Production LLM Inference Running 24/7
An Indian fintech company is serving a customer-facing AI assistant via a single A100 80GB running around the clock. They need guaranteed uptime and consistent latency.
| Pricing Model | Effective Rate | Monthly Cost | Annual Cost |
|---|---|---|---|
| On-Demand | Rs 170/hr | Rs 1,22,400 | Rs 14,68,800 |
| Reserved — 12 months (∼50% off) | Rs 85/hr | Rs 61,200 | Rs 7,34,400 |
| Annual Saving with Reserved | — | — | Rs 7,34,400 |
Verdict: Reserved wins decisively. At 100% utilisation, committing saves Rs 7.3 lakh per A100 GPU per year. An 8-GPU A100 cluster saves over Rs 58 lakh annually — enough to hire two senior AI engineers.
Scenario 2 — Research Lab with Intermittent GPU Use
A pharmaceutical research team runs GPU jobs 40 hours/week, 45 weeks/year on 4 × L40S GPUs. The rest of the time, the GPUs would sit idle under a reserved commitment.
| Pricing Model | Rate | Annual Hours Paid | Annual Cost |
|---|---|---|---|
| On-Demand (pay for actual use) | Rs 61/hr | 7,200 hrs (actual) | Rs 4,39,200 |
| Reserved 12-month (pays for full year) | Rs 30/hr committed | 35,040 hrs (committed) | Rs 10,51,200 |
| Annual Saving with On-Demand | — | — | Rs 6,12,000 |
Verdict: On-demand wins by a wide margin. The team only uses about 20% of what a reserved commitment would require them to pay for. Reserving L40S capacity here means paying for 80% idle time month after month.
Scenario 3 — Startup Monthly Training Sprint
A generative AI startup runs a major training job each month using 8 × H100 GPUs for 10 days (240 hours), then switches to a single H100 for inference the rest of the month (504 hours).
| Workload | Model | Effective Rate | Monthly Cost |
|---|---|---|---|
| Training: 8 H100 × 240 hrs | Spot (70% off) | Rs 66/hr | Rs 1,26,720 |
| Inference: 1 H100 × 504 hrs | Reserved | Rs 110/hr | Rs 55,440 |
| Hybrid Total | — | — | Rs 1,82,160 |
| Same workload, pure on-demand | On-Demand | Rs 219/hr all GPUs | Rs 5,32,872 |
| Monthly Saving with Hybrid | — | — | Rs 3,50,712 |
Verdict: The hybrid approach — spot for training, reserved for inference — saves Rs 3.5 lakh per month, or over Rs 42 lakh per year, compared to running everything on-demand. This is a real strategy used by AI-native companies in India today.
Industry-Specific GPU Pricing Strategies
The optimal pricing model is not just about utilisation. It is also shaped by the compliance requirements, workload rhythms, and business constraints specific to your industry.
Reserved Dedicated Instances — Non-Negotiable for Compliance
Banks and NBFCs processing customer financial data under the DPDP Act 2023 must use India-hosted infrastructure with physical data isolation. This means dedicated GPU instances — typically A100 on 12-month reserved contracts — with full audit trails and Data Processing Agreements. Most BFSI teams also run separate on-demand instances for model development and testing to avoid mixing production and development workloads on the same dedicated hardware.
Hybrid Model: Spot for Training, Reserved for Inference
AI-native companies and ML platform teams typically have two distinct GPU needs: compute-intensive training runs (high burst, intermittent) and always-on inference serving (steady, predictable). The optimal strategy is spot or on-demand for training jobs — with robust checkpointing — and reserved instances for the production inference layer. Teams commonly use A100 reserved instances for production inference and spot H100s for training sprints. This hybrid approach consistently delivers 40–60% lower total compute costs versus running everything on-demand.
Reserved Dedicated for Diagnostics, On-Demand for Research
Hospital AI systems running diagnostic models (radiology, pathology) need consistent, guaranteed latency — reserved dedicated instances are the right choice. Research teams doing drug discovery, genomics, or clinical trial modelling have burst needs better served by on-demand or short-term reserved instances. Healthcare teams must factor HIPAA compliance into their instance selection — Cyfuture AI provides all required compliance documentation.
Project-Based On-Demand or Short Reserved
Animation and VFX studios have extremely variable GPU needs — quiet during pre-production, then massive crunch before delivery deadlines. On-demand or 1–3 month reserved instances align with project timelines. Studios that have moved to generative AI for asset creation are finding the L40S on-demand at Rs 61/hr to be the sweet spot for Stable Diffusion and Flux throughput at a manageable cost-per-render, thanks to its Ada Lovelace architecture optimised for graphics and AI workloads simultaneously.
Spot for Simulations, On-Demand for Analysis
Academic and government research institutions typically have budget constraints but generous time windows for computation. Spot instances are ideal for climate models, molecular dynamics, and HPC simulations — these workloads are long-running and naturally fault-tolerant when built with restart checkpointing. On-demand instances — particularly the L40S at Rs 61/hr — suit shorter analytical runs and interactive data exploration during working hours.
The Hybrid 70-20-10 Model
The most cost-efficient GPU cloud strategies used by mature AI organisations are not pure plays on any single pricing model. They use a structured hybrid approach that matches pricing to workload characteristics. The most battle-tested version is the 70-20-10 rule:
A fintech company applying this model on a 10-GPU baseline saw monthly GPU costs fall from Rs 12.6 lakh (pure on-demand) to Rs 7.4 lakh with the hybrid approach — a 41% saving with no reduction in capability or production reliability. The implementation requires one engineering investment: checkpointing for spot jobs. Most frameworks support it natively, and it pays for itself within the first month of spot savings.
The hybrid model requires that you map each workload type clearly to the right instance category and review the split every quarter as your workload patterns evolve. Teams that set this up once and forget it tend to drift back toward over-provisioning on reserved capacity as projects change. A quarterly 30-minute utilisation review is all it takes to keep the model working.
How to Choose the Right GPU Pricing Model
Use this decision table based on your specific situation. Answer honestly — the “wrong” answer in each case is not a failure, it is useful signal for which model fits your workload today.
| Your Situation | Recommended Model | Why |
|---|---|---|
| Less than 3 months of GPU usage history | On-Demand | Gather real data before committing — your estimates will be wrong |
| GPU utilisation consistently above 70%, ongoing | Reserved (12-month) | The maths are clear — 40–50% savings on your largest cost line |
| Large training jobs with checkpointing already in place | Spot | Up to 70% cheaper; interruptions are recoverable with checkpointing |
| Production inference API serving live traffic | Reserved + On-Demand Fallback | Reserve core capacity, maintain on-demand headroom for traffic spikes |
| Regulated industry (BFSI, Healthcare, HR) | Dedicated Reserved | DPDP/HIPAA requires physical isolation — dedicated is the only compliant option |
| Seasonal or project-based GPU needs | Short Reserved + On-Demand | Reserve for the known duration; use on-demand for overflow during crunch |
| Early startup, PoC, or hackathon | On-Demand | Maximum flexibility, no commitment risk, lowest barrier to getting started |
| Predictable production + active R&D running in parallel | Hybrid 70-20-10 | Match pricing to each workload type; consistently delivers 35–45% total savings |
Need Help Choosing the Right GPU Pricing Model for Your Workload?
From single on-demand L40S and A100 instances to 64-GPU InfiniBand clusters on reserved pricing — Cyfuture AI builds and manages GPU infrastructure for India’s fastest-growing AI teams. DPDP-compliant, India-hosted, and backed by GPU engineers available around the clock.
Frequently Asked Questions
Straight answers to the pricing questions AI teams in India ask most often.
The most affordable entry point on Cyfuture AI is the V100 GPU at Rs 39/hr on-demand. Reserved pricing brings the effective hourly rate down 30–50% further. Spot instances go up to 70% cheaper than on-demand but may be interrupted. For teams doing regular inference or training, an L40S on a reserved commitment at approximately Rs 30–35/hr effective rate offers the best balance of cost, performance, and reliability.
On Cyfuture AI, an H100 SXM5 GPU costs Rs 219/hr on-demand — approximately $2.62/hr at current exchange rates. This is around 51% cheaper than AWS ap-south-1, which charges approximately $5.40/hr for equivalent H100 capacity. On a 12-month reserved commitment, the effective H100 rate on Cyfuture AI drops to approximately Rs 110–130/hr, making it among the most cost-competitive H100 pricing available for Indian enterprises. For teams where the H100 is more than needed, the A100 80GB at Rs 170/hr is a popular cost-performance alternative.
Choose hourly (on-demand) pricing when your GPU utilisation is variable or unpredictable. The practical rule: if your average monthly utilisation would be below 60–65% of the capacity you would commit to in a reserved contract, on-demand is cheaper. Below that threshold, you end up paying for idle capacity with a reserved plan. Above it, the reserved discount more than compensates for the commitment.
Reserved pricing means committing to a GPU instance for 1–12 months in exchange for a 30–50% discount on the standard on-demand rate. On Cyfuture AI, a 12-month reserved A100 80GB drops from Rs 170/hr to approximately Rs 85–100/hr. For a single A100 running continuously, that saving is roughly Rs 7.3 lakh per year. For an 8-GPU A100 cluster, the annual saving exceeds Rs 58 lakh — enough to meaningfully change your AI infrastructure budget.
Spot instances are unused GPU capacity offered at up to 70% off on-demand pricing. They can be interrupted when the provider needs that capacity back. They are absolutely safe for training runs — provided you implement checkpointing, which saves your model’s training state to storage at regular intervals. If interrupted, you resume from the last checkpoint rather than restarting from zero. Most modern training frameworks (PyTorch, JAX, DeepSpeed) support checkpointing natively. Teams that build this in routinely train large models on A100 and H100 spot instances at a fraction of on-demand cost.
On-demand: pay per hour, no commitment, start and stop any time — highest per-hour rate, maximum flexibility. Reserved: commit for 1–12 months, 30–50% cheaper, guaranteed capacity availability. Spot: up to 70% cheaper than on-demand but may be interrupted with short notice — best for batch jobs with checkpointing. Dedicated: entire physical server reserved exclusively for you, no shared tenancy, maximum performance and compliance isolation, premium pricing. Most mature AI teams use a combination of all four, matching each to the right workload type.
Yes. Cyfuture AI’s GPU cloud is 100% hosted in Indian data centres — Mumbai, Noida, and Chennai — and provides the Data Processing Agreements required for DPDP Act 2023 compliance. For regulated industries including BFSI, healthcare, and HR that must process Indian user data within Indian borders, this is a legal requirement rather than a preference. Foreign GPU cloud providers do not automatically satisfy this requirement and would require additional configuration and compliance documentation.
Joita specialises in AI infrastructure, GPU cloud economics, and MLOps for Cyfuture AI. She writes for engineering teams and CTOs who need to make practical decisions about compute costs, cloud architecture, and infrastructure strategy — translating complex pricing mechanics into clear, actionable guidance for teams at every stage.
