Top 10 HPC Cloud Providers in India [2026]

India went from having no real high-performance computing option to having a genuine choice in about eighteen months. L&T is building a 10,000-GPU NVIDIA B300 cluster in Chennai for Together AI on an order worth roughly $1.57 billion. Google broke ground on a gigawatt-scale AI hub in Visakhapatnam under a $15 billion commitment. Microsoft committed $17.5 billion. Yotta already runs more than 16,000 H100s, and Neysa is contracted to over 25,000 GPUs by 2027.

India is now one of the largest AI infrastructure build-outs outside the United States. These are different kinds of commitment, so read the note before comparing the bars.

So availability is no longer the constraint. Completion is. The question now is whether a multi-week parallel run survives the fabric, the checkpoint traffic, the scheduler and the node failures that come with it.

On that test the shortlist splits three ways:

• India specialists that build the cluster to your specification. Neysa, Yotta and E2E Networks have live capacity, published Slurm and parallel storage, and India-hours operations.
• Campus-scale builders whose capacity lands from 2027. L&T Vyoma and Together AI are constructing the largest single clusters in the country, but most of it is not bookable yet.
• Global clouds with strong HPC engineering. Oracle, Azure, AWS and Google are rated well on their global fleets. India-region GPU capacity is a separate conversation, and all four carry CLOUD Act jurisdiction alongside Indian residency.

Neysa leads this list because it is the only India-headquartered provider in the SemiAnalysis ClusterMAX recommended tiers, and the only one that lets you choose silicon, fabric, storage, scheduler and SLA independently rather than accepting a fixed design. If you need a small cluster this month against a published rate card instead, E2E Networks is the fastest route. 

GPUs and pricing at a glance

Provider	GPUs available in India	India locations	Pricing
Neysa	H100, H200, MI300X live. 8,192 MI350X and 9,216 B300 landing	Mumbai, NCR, Bengaluru, Kolkata	On request. On-demand or 12 to 36 month commitment
Yotta Shakti Cloud	16,000+ H100, L40S, sliced GPUs. B200 planned	Yotta-owned data centres	Full public SKU list, INR
L&T Vyoma	10,000 B300 at Chennai, building	Chennai, 250MW Phase 1	Not published
Together AI	10,000 B300 announced for Chennai	No live India region today	Public list, USD
Oracle OCI	Confirm India SKUs	Mumbai, Hyderabad	Public list, USD, per-second on many shapes
Microsoft Azure	Confirm India SKUs	Pune, Chennai, Mumbai, plus Hyderabad mid-2026	Public list, USD, per-second
AWS	Confirm India SKUs	Mumbai, Hyderabad	Public list, USD, per-second
Google Cloud	H100, L4, T4 in Mumbai. H200, RTX PRO 6000 in Delhi	Mumbai, Delhi NCR	Public list, USD, per-second
Tata Communications Vayu	H100, H200, L40S on bare metal	India	On request
E2E Networks	B200, H200, H100, A100, L40S	India	Public INR feed. B200 at INR 290/hr on commitment
IBM Cloud	Confirm India SKUs	Confirm with IBM	Public list

Storage is priced separately almost everywhere. Two published reference points: E2E lists its ultra-high-performance parallel file system at Rs 70 per GB per month, and Yotta lists SSD file storage at Rs 8 per GB per month. On a cluster that checkpoints hourly, that line is not a rounding error.

Only three providers publish an India GPU count. E2E Networks, Tata Communications and the global clouds do not, which is not the same as not having one.

How we ranked. Contracted India capacity with a date attached. Non-blocking fabric. A parallel file system that survives checkpoints. Scheduler configuration on delivery. Independent verification rather than vendor claims. India residency and India-hours support. Announced megawatts did not count: a campus that energises in 2028 does not help a run that starts in March.

If you want the reasoning behind those six criteria, it is in the buyer’s guide below the list.

1. Neysa

Best for: dedicated single-tenant clusters in India where silicon, fabric, storage and SLA are specified to the workload, with operations included.

At a glance: H100, H200, MI300X live · 8,192 MI350X and 9,216 B300 landing · Mumbai, NCR, Bengaluru, Kolkata · Pricing on request, on-demand or 12 to 36 month · ClusterMAX recommended tier

Neysa is the only India-headquartered provider currently in the recommended tiers of the SemiAnalysis ClusterMAX rating.

Nothing about the cluster is pre-decided. Most providers make the architectural choices for you. Neysa makes them with you.

• Silicon. NVIDIA and AMD, current generation. Background: NVIDIA B300 and AMD MI300X.
• Fabric. NVIDIA Quantum-X800 InfiniBand at 800G where you want the validated reference architecture. Non-blocking open-standards RoCE v2 Ethernet at 400G and 800G where you want vendor independence. Both offered non-blocking.
• Storage. WEKA, VAST or DDN. Open-source software-defined storage. Or your own licences on Neysa hardware, which matters if you have already bought them.
• Scheduler. Pure bare metal, bare metal with Slurm pre-configured for batch and checkpoint recovery, or bare metal with Kubernetes.
• Contract. Control planes, non-GPU node types, SLA terms and hot-spare counts specified per deployment. See AI cloud architecture and design.

Operations, which is what decides long runs.

• 99.9% uptime SLA, held 100% for six months running.
• Thirty-minute first response on Severity-1, against a four-hour resolution target.
• GPU-replacement SLA in the contract. Affected instances restored in 30 to 60 minutes.
• 24×7 India-based NOC and SOC. Named technical account manager. Four-level escalation.
• Roughly one incident in three caught by monitoring before the customer reports it.
• ISO 27001:2022, ISO 27017, ISO 27018, SOC 2 Type II, CSA STAR Level 2, DPDP Act readiness. See AI infrastructure security.

Proof from production. ITQ Technologies: about 141 billion tokens a month, 40% lower cost, P99 latency from 14 seconds to under 2. TIFIN: 65% reduction in GPU cost. Innoviti: 60% lower TCO across 50,000+ merchants.

2. Yotta Shakti Cloud

Best for: the largest deployed NVIDIA fleet in the country, from an operator that owns the buildings as well as the racks.

At a glance: 16,000+ H100, L40S, sliced GPUs · B200 planned · Yotta-owned data centres · Full public SKU price list · TOP500 listed

• Over 16,000 H100 GPUs deployed, with B200 in plan.
• A Shakti system appears on the TOP500 list on H100 over InfiniBand NDR400. That is worth more than it looks: TOP500 entry requires running LINPACK across the whole machine, so the fabric has been proven under a real collective workload rather than described in a datasheet.
• Slurm and Kubernetes both supported, with published guidance on when to use each.
• ·       File storage published at Rs 8 per GB per month on SSD. Bare metal as 8x HGX H100 with 640GB of GPU memory.
• Powers national-scale work including the PARAM-1 foundation model.
• Publishes a full SKU price list, which is rare in this market.
• Owns its own data centres, removing a counterparty between you and the rack.

3. L&T Vyoma and LTN Compute

Best for: campus-scale capacity from 2027, from a builder that has delivered national infrastructure for decades.

At a glance: 10,000 B300 at Chennai, building · Chennai gigawatt campus, 250MW Phase 1 · Pricing not published · Not rated

• L&T Vyoma is L&T’s sovereign AI cloud and data centre business, launched February 2026.
• The Chennai campus is gigawatt-scale, Phase 1 designed for 250MW with 150MVA of power readiness. For scale, that is roughly five times the total commissioned power most Indian GPU providers operate today.
• LTN Compute sells AI Factory services, GPU-as-a-Service and colocation.
• Partnered with Fortanix on confidential computing with NVIDIA technology, so data stays encrypted in GPU memory during processing and even the operator cannot read it. For some regulated workloads that is the only acceptable answer.
• In August 2026 it secured an order worth roughly $1.57 billion to build a 10,000-GPU B300 cluster for Together AI.
• Worth confirming: the platform is new. Ask for published cluster service levels, fabric oversubscription ratio and scheduler configuration. The largest Chennai block already has an anchor tenant, so confirm what is available to you and when.

4. Together AI

Best for: a mature training and inference software stack, once its India capacity is live.

At a glance: 10,000 B300 announced for Chennai · No live India region today · Public list pricing, USD · ClusterMAX Silver

• Holds a Silver rating in ClusterMAX, among the better-run GPU clouds globally.
• Slurm is a first-class managed capability with published configuration controls, not a toolkit you assemble.
• Instant Clusters support long-lived, resizable in-datacentre shared storage, which is the right shape for checkpoint traffic.
• Its India position is the L&T Chennai deployment: 10,000 NVIDIA B300 GPUs, described by both parties as the largest single GPU cluster in the country.

Worth confirming: Together’s documentation names no live India region today, and data residency is currently an enterprise option. For a 2026 start, plan on their existing regions and treat India as forthcoming.

5. Oracle Cloud Infrastructure

Best for: regulated enterprises already standardised on Oracle that want cluster networking without adding a vendor.

At a glance: Confirm India GPU SKUs · Mumbai, Hyderabad · Public list pricing, USD, per-second on many shapes · ClusterMAX Gold

• Holds a Gold rating in ClusterMAX, the highest of any general-purpose cloud.
• Bare metal is a first-class product, not an exception. Oracle was the first major cloud to offer bare metal instances, and publishes plans to scale to 65,536 NVIDIA GPUs in an OCI supercluster.
• RDMA cluster networking with low-latency GPU interconnect, so the collectives problem is designed for rather than worked around.
• OCI File Storage and Lustre cover parallel workloads.
• Two live India regions, India West (Mumbai) and India South (Hyderabad).
• OCI Dedicated Region and dedicated bare-metal GPU clusters cover single-tenant requirements.

Worth confirming: the Gold rating reflects the global fleet, so confirm contiguous multi-node GPU capacity inside the India regions in writing. Slurm is orchestrated rather than a first-party managed service, so budget configuration time. Oracle is US-headquartered, so CLOUD Act jurisdiction applies alongside India residency.

6. Microsoft Azure

Best for: enterprises with an existing Microsoft agreement, where procurement speed outweighs unit cost.

At a glance: Confirm India GPU SKUs · Pune, Chennai, Mumbai, plus Hyderabad from mid-2026 · Public list pricing, USD · ClusterMAX Gold

• The other Gold-rated general-purpose cloud in ClusterMAX.
• InfiniBand-connected ND-series clusters, with Azure CycleCloud orchestrating Slurm.
• Azure Managed Lustre provides a managed parallel file system, so checkpoint throughput is a solved problem here.
• India regions in Central India (Pune), South India (Chennai) and West India (Mumbai).
• A new India South Central region in Hyderabad from mid-2026, three availability zones, with published customer data residency options, under a $17.5 billion India commitment.

Worth confirming: no bare-metal GPU size appears on the VM sizes page and GPU availability on Dedicated Hosts is limited, so ask how single-tenant isolation is delivered. India GPU quota competes with global demand. Microsoft is US-headquartered, so CLOUD Act jurisdiction applies. Comparison: Microsoft Azure alternatives in India.

7. Amazon Web Services

Best for: clusters that need to sit inside a much larger managed estate.

At a glance: Confirm India GPU SKUs · Mumbai, Hyderabad · Public list pricing, USD · ClusterMAX Silver

• Holds a Silver rating in ClusterMAX.
• Every HPC problem has a named AWS product: ParallelCluster for Slurm, FSx for Lustre for parallel storage, Elastic Fabric Adapter for the RDMA path, UltraClusters for physical placement.
• EC2 Dedicated Hosts and .metal instance types give real single-tenant options.
• Two India regions: Asia Pacific (Mumbai) since 2016 and Asia Pacific (Hyderabad) since 2022.
• Backed by a $12.7 billion India infrastructure commitment through 2030, since raised by a further $13 billion to a total of $48 billion between 2026 and 2030.

Worth confirming: newest accelerators typically reach India regions later than US regions, so check your SKU is available locally before planning around it. Egress is chargeable, which matters if your training data lives elsewhere. AWS is US-headquartered, so CLOUD Act jurisdiction applies. Comparison: AWS alternatives in India.

8. Google Cloud

Best for: TPU and JAX workloads, and organisations planning for 2028 as much as this year.

At a glance: H100, L4, T4 in Mumbai · H200, RTX PRO 6000 in Delhi NCR · Public list pricing, USD · ClusterMAX Silver

• Holds a Silver rating in ClusterMAX.
• Two India regions with real, published GPU capacity. asia-south1 Mumbai carries H100, L4 and T4 across A3 Ultra, A3 Edge, G4, G2 and N1 families. asia-south2 Delhi carries H200 and RTX PRO 6000. Publishing accelerator availability per region is unusual and useful.
• Google Cloud Managed Lustre and Filestore cover parallel and shared storage. Cluster Toolkit deploys Slurm.
• Building a gigawatt-scale AI hub in Visakhapatnam on a $15 billion investment across 2026 to 2030, including a new international subsea gateway.

Worth confirming: sole-tenant nodes give physical isolation but there is no bare-metal GPU product, and Slurm is deployed via Cluster Toolkit rather than offered as a managed service. Vizag is a 2026 to 2030 build. Google is US-headquartered, so CLOUD Act jurisdiction applies. Comparison: GCP alternatives in India.

9. Tata Communications Vayu AI Cloud

Best for: clusters that move large volumes of data between India sites and an existing estate.

At a glance: H100, H200, L40S on bare metal · India · Pricing on request · ISO/IEC 42001 certified AIMS

• Vayu AI Cloud offers NVIDIA H100, H200 and L40S on bare metal, from what Tata announced as one of India’s largest NVIDIA Hopper deployments.
• The distinctive strength here is the network. Training data usually starts somewhere other than the cluster, and moving tens of terabytes into position is its own project. IZO Multi Cloud Connect makes that a problem Tata already owns end to end.
• Runs under an ISO/IEC 42001:2023 certified AI management system, the first AI management standard. Very few providers anywhere hold it.

Worth confirming: the published fleet is Hopper-generation, and fabric design, parallel file system options and maximum contiguous cluster size are not detailed publicly. Ask for all three before a large multi-week run.

10. E2E Networks

Best for: small and mid-sized clusters where a published rate card and a fast start carry the decision.

At a glance: B200, H200, H100, A100, L40S · India · Public INR price feed, B200 at INR 290/hr on commitment · ClusterMAX tested

• NSE-listed, and publishes a genuine machine-readable price feed: NVIDIA B200 at INR 290 per hour on commitment, plus H200, H100, A100 and L40S.
• Slurm-native scheduling. All workloads are scheduled through Slurm via Slinky and Pyxis, described as purpose-built for HPC and large-scale distributed training. That is a stronger published position on the scheduler question than most of this list.
• A shared high-throughput parallel file system for multi-node GPU training, published at Rs 70 per GB per month for the ultra-high-performance tier. Published storage tiering lets you model the checkpoint arithmetic before you sign.
• Private clusters with bare-metal compute nodes. Managed Kubernetes with node pools and autoscaling.
• IAM single sign-on with enforced MFA through Microsoft Entra ID, Terraform support, and a published 99.9% SLA on compute VMs.

Worth knowing: L&T acquired a 21% stake in E2E in late 2024 as a strategic partnership. L&T has since launched its own AI cloud under Vyoma and won the order to host Together AI’s India capacity. E2E operates independently. Useful context when planning a multi-year commitment.

Worth confirming: ask for the fabric oversubscription ratio, and for benchmark access on a short commitment so you can validate collective performance on your own model before scaling.

Side by side comparison

Statuses come from a first-party documentation audit. Not published means we could not find it in the vendor’s own material. That is not the same as absent.

Provider	Bare metal	Slurm	Parallel file system	Single-tenant	Live India capacity	Independent signal
Neysa	Yes, first-class	Yes, pre-configured	WEKA, VAST, DDN or BYO	Yes, by default	Mumbai, NCR, Bengaluru, Kolkata	Only India-headquartered provider in ClusterMAX recommended tiers
Yotta Shakti Cloud	Yes	Yes	Shared file storage published	Yes	Yes	TOP500 listed
L&T Vyoma	Yes	Not published	Not published	Yes	Chennai campus building	Not rated
Together AI	Yes	Yes, managed	Yes	Yes	Announced, not live	ClusterMAX Silver
Oracle OCI	Yes	Orchestrated	Yes, Lustre	Yes	Mumbai, Hyderabad	ClusterMAX Gold
Microsoft Azure	Partial	Via CycleCloud	Yes, Managed Lustre	Partial	Pune, Chennai, Mumbai, +1 in 2026	ClusterMAX Gold
AWS	Yes	Yes, ParallelCluster	Yes, FSx for Lustre	Yes	Mumbai, Hyderabad	ClusterMAX Silver
Google Cloud	Partial, sole-tenant	Via Cluster Toolkit	Yes, Managed Lustre	Partial	Mumbai, Delhi NCR	ClusterMAX Silver
Tata Communications Vayu	Yes	Not published	Not published	Yes	Yes	ISO/IEC 42001 certified AIMS
E2E Networks	Yes	Yes, Slurm-native	Yes, published parallel FS	Yes, private clusters	Yes	ClusterMAX tested

The buyer’s guide: what actually separates these providers

Everything below is the reasoning behind the ranking. Read it if you are running the evaluation rather than skimming the shortlist.

Three scales of GPU work, and why only one of them is HPC

• One GPU. You are renting an accelerator. Performance is a property of the chip: how much memory it has, and how fast that memory is.
• One node, up to eight GPUs. The GPUs inside a server talk over NVLink, coordinated by NVSwitch for all-to-all traffic. Bandwidth here is enormous and effectively free.
• Many nodes. Now the GPUs talk across a network. This is where HPC begins, and where the differences between these providers live.

A GPU cloud sells you the first two. An HPC cloud sells you the third. Once you cross that line, four things decide your throughput.

1. The fabric, because leaving the node costs an order of magnitude

Every training step ends the same way. Each node computes gradients on its slice of the batch, then all nodes agree on a single averaged set before the next step starts. That operation is an all-reduce, and it is a collective: every participant must finish before any participant moves on.

A collective runs at the speed of its slowest path.

Figure 2: Inside a Blackwell node, GPUs talk at 1,800 GB/s. The moment traffic leaves the node, the best available link drops to roughly 100 GB/s.

Work it through. A 70-billion-parameter model in bf16 holds 140 GB of gradients. A ring all-reduce moves roughly twice that across each node’s uplink, about 280 GB per node per step.

• On an 800G InfiniBand fabric, roughly 100 GB/s, that exchange takes under 3 seconds.
• On standard 100 gigabit Ethernet, roughly 12.5 GB/s, it takes over 20 seconds.

Frameworks overlap communication with computation, so you never pay the full number. But you cannot overlap bandwidth you do not have. On the wrong fabric the GPUs sit waiting, and your monitoring reports it as low GPU utilisation rather than as a network problem. That is a common misdiagnosis in multi-node training.

Two terms you will meet in a proposal:

• Non-blocking, or 1:1. Every node can send at full rate simultaneously without contending for a shared uplink.
• Oversubscribed, often 3:1. The uplink leaving a switch carries a third of the bandwidth of the ports beneath it. Sensible for web servers, which rarely all talk at once. In an all-reduce every node talks at once, so you receive a third of the fabric you thought you bought.

Ask for the oversubscription ratio in writing. It is a number, and a provider running a non-blocking fabric will be glad to give it to you.

InfiniBand or Ethernet? Both work. The trade is real.

	NVIDIA Quantum InfiniBand	RoCE v2 over open Ethernet
Congestion handling	Managed in the fabric, mature, predictable at scale	Requires careful DCQCN and PFC tuning to match
Ecosystem	NVIDIA reference architectures assume it	Multi-vendor, standard switches, no single supplier
Typical speed in India today	800G	400G and 800G
Best when	You want the validated NVIDIA design	You want vendor choice, or you are running AMD Instinct

A provider that offers only one of the two has made that decision for you. More in HPC architecture explained.

2. The file system, because checkpoints stall the whole cluster

A parallel file system lets thousands of processes read and write the same namespace at once, striping data across many storage servers. WEKA, VAST, DDN and Lustre are the common implementations.

A 70-billion-parameter model does not checkpoint 140 GB. It checkpoints the full training state: bf16 weights, an fp32 master copy, and two Adam optimizer moments. That is roughly 12 to 14 bytes per parameter, so about 1 TB per checkpoint.

Checkpoint hourly, normal for a run this size, and over seven days you write it 168 times.

• At 10 GB/s aggregate write throughput, each checkpoint stalls the cluster for about 100 seconds. Over the run, 4.7 hours of paid GPU time doing nothing.
• At 100 GB/s, each costs about 10 seconds, and the run loses 28 minutes in total.

Figure 4: The same 168 checkpoints, four different storage tiers. Storage throughput is a GPU cost line, not an IT line.

The GPUs are identical in both cases. The bill is not.

Object storage over a standard network is the wrong tool here. It is excellent for datasets and archives, and it is not built for thousands of ranks writing a coordinated checkpoint. See object vs block vs file storage.

3. The scheduler, because placement follows topology

Slurm is the standard batch scheduler for HPC. It decides which physical nodes your job runs on.

By default it does not know how those nodes are wired. You tell it through a file called topology.conf, which describes which nodes sit under which switch.

Without it, Slurm can scatter 64 ranks across the fabric in a way that pushes every collective through the busiest uplinks in the building. The job runs. It is simply slower than it should be, on every step, forever, and nothing in your logs says why.

Three ways providers hand you a cluster:

• Pure bare metal. Root access, your own kernel. Take this if you have a platform team and strong opinions.
• Bare metal with Slurm. Batch scheduling, queueing and checkpoint recovery, configured before handover. Take this for classical training runs.
• Bare metal with Kubernetes. Container orchestration for workloads that mix training with serving, or scale dynamically.

For the Slurm option, ask specifically whether topology awareness is configured on delivery or left to you.

4. Failure detection, because long runs fail and short runs do not

GPUs rarely fail cleanly. They throttle under thermal pressure, log correctable ECC errors, or slow just enough to lag their rank. In a synchronous training job, one slow rank sets the pace for every other rank in the cluster.

Then there is outright node loss, and the arithmetic is unforgiving.

Figure 3: Interruption risk compounds with cluster size. A seven-day run on 512 nodes is a coin flip even at a modest per-node failure rate.

A single node might be reliable. Five hundred of them, running for a week, are not. Two contractual numbers matter more than the hourly rate:

• How long a failed node stays in your allocation. A four-hour replacement instead of thirty minutes loses three and a half hours of the entire cluster, not of one node.
• What share of incidents the provider catches first. The difference between a swap that happens while you sleep and one that starts when you file a ticket.

The term for what you are protecting is goodput: the fraction of allocated GPU time that produces work you keep. Everything above is a goodput argument. See GPU monitoring and ML observability.

What ClusterMAX is, since it appears throughout this list

SemiAnalysis rents clusters from GPU providers, runs real workloads on them, and publishes what happened. Providers are placed in tiers from Platinum down through Gold, Silver, Bronze, Underperforming and Unavailable. Bronze is the lowest tier SemiAnalysis still recommends.

It is the only independent, hands-on evaluation this market has, which is why it carries weight here.

One structural difference worth naming

AWS, Azure, Google and Oracle all keep data in the India region you select. All four are US-headquartered, so CLOUD Act jurisdiction applies on top of Indian residency.

Residency answers where the bytes sit. Jurisdiction answers whose courts can compel access to them. They are different questions, and for most workloads only the first matters. Where the second matters, as it often does in banking and financial services or insurance, an India-headquartered provider answers something the others structurally cannot.

What to put in the contract

Four clauses decide how an HPC engagement goes. Each maps to one section above.

A dated delivery schedule with a slip consequence. Cluster size, silicon, fabric, storage, in service by a date. Capacity that arrives two quarters late is capacity you did not buy.

The fabric oversubscription ratio, in writing. The difference between the bandwidth you were quoted and the bandwidth you receive during an all-reduce.

A GPU replacement SLA in hours. Everyone quotes uptime. On a 500-node run, this single number caps your exposure.

Proactive monitoring with a defined escalation path. Ask what share of incidents the provider catches before the customer does, and who you reach at 3am.

Getting a cluster in India is now solvable. Getting a run to finish on it is the purchase that matters.