GPU cloud pricing in India 2026: H100, H200 and B200 rates compared

Summary. Renting an NVIDIA H100 in India starts near ₹217/hour on JarvisLabs and about ₹329/hour on-demand on Cyfuture AI as of July 2026, while the government-backed IndiaAI Mission puts H100-class compute at roughly ₹92/hour after subsidy and ₹67/GPU-hour for standard GPUs. Buying the same card outright costs ₹30-50 lakh before power, cooling and networking. This guide compares H100, H200 and B200 cloud rates across Indian providers (E2E Networks, Yotta, Cyfuture, AceCloud), global neoclouds (Spheron, RunPod, Lambda Labs, Vast.ai) and the Mumbai regions of AWS, Azure and Google Cloud, then works through the rent-versus-reserve-versus-buy maths for an Indian engineering team.

The number that matters has changed. Two years ago the question was whether you could get an H100 in India at all. In 2026 the supply exists, the billing is in rupees, and the spread between the cheapest and most expensive way to run the same GPU is close to 100x. A single H100-hour costs about ₹67 through the IndiaAI subsidy, ₹217-400 on a domestic cloud, and roughly ₹1,000 per GPU once you split an eight-way hyperscaler instance in the Mumbai region. Picking the wrong lane is now the biggest controllable line item in an AI budget.

This is a pricing and decision guide, not a vendor pitch. Every rate below is dated and sourced, and rates move with GPU supply, so treat the figures as a July 2026 snapshot and re-check the provider portal before you commit spend.

What Indian teams actually pay for a GPU-hour in 2026

Three markets sit side by side, and they price the same silicon very differently.

The first is the domestic cloud market: E2E Networks, JarvisLabs, Cyfuture AI, AceCloud and Yotta, all billing in INR with GST invoices. The second is the global neocloud market: Spheron, RunPod, Lambda Labs, Vast.ai and CoreWeave, billing in US dollars but accepting Indian cards. The third is the hyperscaler market: AWS, Azure and Google Cloud, which list H100 instances in their India regions only in eight-GPU blocks at a heavy premium. Layered under all of it is the IndiaAI Mission, a subsidised compute pool that undercuts every commercial option for teams that qualify.

Currency conversions in this guide use the sources' own rates, roughly ₹84-95 per US dollar during 2026; the exact figure shifts daily, so price in the provider's native currency where you can.

The pattern is consistent. Domestic clouds beat the hyperscalers on single-GPU work by a wide margin, global neoclouds win on raw spot price, and the IndiaAI Mission beats everything if your project is eligible. The rest of this guide breaks down each lane with real SKUs.

On-demand rental: Indian providers compared

For most teams that need one to eight H100s without a procurement cycle, a domestic cloud is the practical default. You pay in rupees, you get a GST invoice you can claim as input credit, and the data stays inside India, which matters for regulated workloads under the DPDP Act.

E2E Networks is the anchor. It is India's only publicly listed pure-play GPU cloud, with data centres in Delhi NCR, Mumbai and Bengaluru, and it is empanelled on MeitY's GI Cloud. Its published H100 80GB SXM5 rate is ₹350-400/hour on-demand, with spot instances at ₹120-180/hour and a monthly commitment near ₹2,50,000-2,80,000 for continuous use. A four-way HGX H100 with NVLink runs ₹1,400-1,600/hour.

JarvisLabs lists a single H100 at ₹217.89/hour with per-minute billing and no minimum commitment, which is among the lowest transparent on-demand rates in the India region. Cyfuture AI quotes ₹329/hour on-demand, dropping to ₹219/hour on a 12-month reserved plan, with per-second billing and DPDP, SOC 2 and ISO-focused positioning. AceCloud prices differently, selling the H100 HGX as a monthly block from ₹1,80,000/month, falling to an effective ₹1,62,000/month on a 12-month term, which suits teams that run a card continuously rather than in bursts.

Yotta Data Services and Tata Communications sit at the enterprise end. Yotta Shakti runs on a Tier IV+ campus in Navi Mumbai and sells H100 and A100 capacity through enterprise agreements rather than a self-serve portal, so pricing is quote-based and typically at or above E2E's rates for the SLA. Tata Communications offers H100 and A100 through its Vayu AI Cloud, bundled with private MPLS connectivity for hybrid deployments. ESDS in Nashik serves the SME segment but largely on older V100 and T4 hardware, with limited H100 supply.

The takeaway for a startup: JarvisLabs or E2E spot for experiments, Cyfuture reserved or AceCloud monthly for a card you keep busy, and Yotta or Tata only when an India-resident enterprise SLA is contractual. Teams weighing this against renting inference capacity by the token should also read our breakdown of H100 inference cost per million tokens.

H200 and B200: what Blackwell costs and where to get it

H100 supply is now healthy in India. H200 and B200 (Blackwell) supply is not, and where you find it changes the price sharply.

Domestic self-serve B200 capacity is thin. E2E Networks lists H200 and B200 alongside H100 on its portal, with catalog "from" rates that Spheron's May 2026 comparison put near $2.20/GPU-hour for H200 and $4.90/GPU-hour for B200, but availability fluctuates with demand. Yotta does not yet sell H200 or B200 on self-serve, even though it is building one of Asia's largest Blackwell clusters (more on that below). For guaranteed Blackwell access today, global neoclouds are usually faster.

Spheron, an aggregator that pools bare-metal capacity, lists H100 SXM5 at $3.84/hour on-demand and $1.63 spot, H200 at $4.56/$1.89, and B200 SXM6 at $7.16 on-demand but only $1.71 spot, which is below its own on-demand H100 rate when supply is loose. RunPod's Community tier lists H100 PCIe at $1.99, H200 near $3.59 and B200 at $5.98. Lambda Labs, aimed at multi-node training with InfiniBand, runs H100 at $3.99-4.29, H200 at $4.99+ and B200 SXM6 at $6.69-6.99. Vast.ai's marketplace produces the lowest floor, with H100 at roughly $1.55-2.50 and A100 as low as $0.67, at the cost of variable host reliability and no SLA.

Two practical notes. First, spot B200 can be cheaper than on-demand H100 on aggregators when Blackwell supply outruns demand, so a checkpointed training job may finish faster and cheaper on spot B200 than on-demand H100. Second, latency from Mumbai to US and EU nodes runs 120-200ms, which is fine for training throughput but poor for latency-sensitive inference, where an India-resident node wins. If you are still choosing between generations on a cost-per-token basis, our B200 versus H100 inference cost analysis has the per-token maths.

Hyperscalers in India: why AWS, Azure and GCP cost 3-5x

The large clouds do offer H100 in their India regions, but the packaging makes them expensive for anything under a full node. As compiled by E2E Networks, AWS P5 (p5.48xlarge, eight H100s) in Mumbai lists at about $98/hour, Azure's Standard_ND96isr_H100_v5 (eight H100s) at about $83/hour, and Google Cloud's a3-highgpu-8g (eight H100s) at about $85/hour. All three sell eight GPUs at a time, with no single-GPU H100 option.

Split those rates per GPU and you land around $10-12/GPU-hour, roughly ₹850-1,050, which is three to five times a domestic on-demand H100 and more than ten times an IndiaAI-subsidised hour. Committed-use discounts on the hyperscalers require one-to-three-year terms, and licensing is more involved.

That does not make them the wrong choice for everyone. If your data, identity and pipeline already live on AWS, Azure or GCP, keeping GPU training on the same cloud avoids cross-cloud egress and integration work, and egress is a real cost that we break down in our note on cloud egress fees on AI inference bills. The rule of thumb: use a hyperscaler when the workload is glued to that cloud's other services, and a domestic cloud or neocloud when the GPU job is standalone.

The IndiaAI Mission: subsidised compute at ₹67-92/GPU-hour

The cheapest legitimate GPU in India is not on any commercial price list. It is the IndiaAI Mission, a ₹10,300 crore (about $1.25 billion) government programme to build sovereign AI infrastructure, and its subsidised compute portal is the single biggest lever on cost for an eligible team.

By mid-2026 the mission had empanelled more than 38,000 GPUs, well past its original 10,000-chip target, with roughly 15,000 of them H100 or H200-class. The GPUs are owned by private partners (Jio, Tata Communications, Yotta and E2E Networks among 14 providers) and the government subsidises the per-hour rate to end users. The lowest winning bids came in at ₹115.85/GPU-hour for standard GPUs and ₹150/GPU-hour for high-end H100-class GPUs; eligible users then receive up to a 40% subsidy, which brings effective rates to about ₹67/GPU-hour for standard cards and ₹92/hour for an H100.

Union IT minister Ashwini Vaishnaw has framed this as deliberate industrial policy, saying India would "provide the cheapest compute facilities in the world which is ₹67/GPU hour." For a narrow set of teams the price falls to zero: the government has offered a 100% compute subsidy for approved foundational-model builders, and Sarvam AI received 4,096 H100 SXM GPUs backed by a ₹98.68 crore subsidy to build 70B-parameter Indic models, with Gnani.ai, Soket AI Labs and GAN.ai also cleared.

The catch is access, not price. You apply through the IndiaAI Compute Portal, approval is discretionary and prioritises projects of national importance, and H100 availability can tighten during peak demand. For a recognised Indian AI startup or research group, though, the maths is decisive: ₹67-92/GPU-hour undercuts every commercial route below. We cover the policy context and its data-residency implications in our piece on India's sovereign AI push and the IndiaAI Mission, and the company that anchored the 100% subsidy in our profile of Sarvam AI's rise to sovereign-AI unicorn.

Rent vs reserve vs buy: the real maths

The instinct to buy hardware rarely survives contact with the total cost of ownership. NVIDIA does not sell the H100 at retail in India; you go through channel partners such as Dell, HP, Lenovo or Supermicro, and import duties of 18-28% plus GST push the landed price of an H100 PCIe to ₹30-40 lakh and an SXM to ₹40-50 lakh. An eight-GPU HGX server lands at ₹3.5-5 crore.

Then the hidden costs start. Each H100 draws 700W, so an eight-GPU node needs dedicated power that can cost ₹50 lakh to provision, liquid or immersion cooling from ₹15 lakh to ₹1 crore, and InfiniBand networking from ₹50 lakh. A functional owned cluster in India is a ₹2-5 crore commitment before the first training run, plus scarce GPU-infrastructure engineers to keep it alive.

Run the comparison for a ten-person team that uses GPUs 20-30% of the time, roughly five to seven hours a day.

At that utilisation the break-even for buying versus renting is roughly 42 months, and that ignores the crore-scale infrastructure bill. The plain version: buy only if you will run the GPUs close to 24/7 for 18 months or more and already have the team to operate on-premise hardware. Everyone else rents, and picks the lane by workload. The same lesson shows up in enterprise GPU budgeting for the next generation in our note on the NVIDIA Rubin buildout and cloud budgets, and GPU spend now tops most FinOps teams' worry list, as we cover in why GPU spend is the top FinOps concern of 2026.

India-specific considerations: GST, RBI, DPDP and domestic supply

Four India-specific factors change the decision beyond the sticker rate.

Tax and input credit. When you pay a foreign GPU cloud, 18% IGST applies under the reverse-charge mechanism, and a GST-registered business can reclaim it as input tax credit. On ₹1.5 lakh a month of foreign GPU spend, that is roughly ₹27,000 recoverable, so the real gap between a domestic INR invoice and a USD one is smaller than the headline once credits are filed.

Cross-border payments. Payments to foreign providers fall under the RBI's Liberalised Remittance Scheme. International cards issued by Indian banks process them automatically, and no special RBI approval is needed below $250,000 per year, so a startup can pay Spheron or RunPod by card without friction.

Data residency. The Digital Personal Data Protection Act 2023 does not currently force India-located compute for training on public or synthetic data, or for inference that touches no personal data. It becomes decisive for regulated fintech and health pipelines processing Indian citizens' personal data at scale, and for government contracts with residency clauses. The clean pattern is to keep the regulated inference layer on E2E Networks or Yotta and run non-regulated training on the cheapest neocloud. Our DPDP engineering playbook for Indian startups covers how to draw that line.

Domestic supply is arriving fast. Yotta raised $150 million in July 2026 at a ₹37,000 crore (about $4.4 billion) valuation and is deploying 20,736 liquid-cooled NVIDIA Blackwell Ultra GPUs in a supercluster worth more than $2 billion, expected live by August 2026, on the way to about 85,000 GPUs by the end of the financial year. Reliance is running Blackwell GPUs in a 1GW data centre in Gujarat. Sunil Gupta, co-founder, MD and CEO of Yotta Data Services, put the ambition plainly: "India's AI ambition requires sustained, high-performance compute at scale. By combining Blackwell Ultra infrastructure with open models like NVIDIA Nemotron and the full NVIDIA AI stack, we are enabling developers to build sovereign, globally competitive AI applications from India." For buyers, more domestic Blackwell supply through 2026 should ease H200 and B200 availability and put mild downward pressure on rates.

How to choose: a short decision guide

Match the lane to the job. For a recognised Indian AI startup or research group building something of national relevance, apply to the IndiaAI Mission first, because ₹67-92/GPU-hour is unbeatable and the approval wait is worth it. For fine-tuning, inference and anything that must keep Indian personal data in India, use a domestic on-demand cloud such as JarvisLabs, Cyfuture AI or E2E Networks, and move steady workloads onto a reserved or monthly plan. For multi-GPU training that can tolerate interruption, use global neocloud spot capacity on Spheron, Vast.ai or RunPod and checkpoint aggressively. For Blackwell today, go to a global neocloud on-demand or E2E's limited catalog. And stay on a hyperscaler only when the GPU job is welded to services you already run on AWS, Azure or GCP. When training is bursty and inference is steady, the cheapest answer is usually two providers, not one, and teams weighing self-hosting against buying tokens should read our comparison of self-hosting a large model versus API cost.

FAQ

How eCorpIT can help

eCorpIT helps Indian and global teams cost, provision and run GPU workloads without overpaying for compute or over-committing to hardware. We benchmark your training and inference profile against domestic clouds, neoclouds, hyperscalers and IndiaAI eligibility, then design a placement that keeps regulated data India-resident while pushing bursty jobs to the cheapest spot capacity. If you want a private, cost-controlled deployment, our senior engineering team can build and operate it end to end, as described in our private LLM deployment service. To model your own GPU spend or plan an IndiaAI application, contact us for a working session.

References

1. E2E Networks, H100 GPU pricing in India (2026).

1. Spheron, GPU cloud providers in India 2026: H100, H200 and B200 availability with INR billing (23 May 2026).

1. Daya Shankar, NVIDIA H100 price in India 2026: buy, rent, or get it for 70% less through the government?, Hugging Face (28 April 2026).

1. IndiaAI Mission, IndiaAI Compute Capacity portal (2026).

1. Analytics India Magazine, India will provide the cheapest compute facilities in the world at ₹67 per GPU hour: IT minister Ashwini Vaishnaw (2026).

1. DD News, Transforming India with AI: ₹10,300 crore mission, 38,000 GPUs (2026).

1. Business Standard, Yotta raises $150 mn at ₹37,000 crore valuation to fund AI expansion (5 July 2026).

1. NVIDIA Blog, Yotta CEO Sunil Gupta on supercharging India's fast-growing AI market (2026).

1. Outlook Business, Govt offers 100% compute subsidy for foundational AI model development (2026).

1. HPCwire, Yotta plans $2B NVIDIA Blackwell Ultra supercluster deployment in India (2026).

1. AceCloud, NVIDIA H100 HGX pricing (2026).

1. Inc42, Govt to offer 100% compute subsidy to Sarvam, others for foundational models (2026).

_Last updated: 23 July 2026._