← All news and guides
Private AI

VPC vs On-Premise vs Indian GPU Cloud: Choosing Where Your Private AI Runs

Compare a VPC in an Indian cloud region, on-premise servers and Indian GPU clouds for private AI, with IndiaAI eligibility, dated GPU rates and a checklist.

Mono Team ·

A quiet data hall corridor between two rows of server racks, cool light on a polished floor, one cabinet door slightly open.

Short answer: run private AI in a virtual private cloud (VPC) in an Indian cloud region if your team already works in that cloud and wants to start fast. Choose on-premise servers only if you must own the hardware and can staff it. Choose an Indian GPU cloud when GPU price and availability matter more than the wider cloud toolkit. Any of the three can keep your data in India. None of them makes the system private by itself. Privacy comes from the controls you build on top: no outbound path for prompts, access limited to named people, logs you control and a tested way to delete data. This guide compares the three options, shows what each costs to start from public GPU rates, and explains who can use the IndiaAI subsidised compute programme.

It is an engineering comparison, not legal advice. Prices are public list prices checked on 20 November 2026 and change often. We give no speed or latency figures. Test your own workload. For the cost of private AI against an API, read Private LLM vs OpenAI or Claude API: a cost and risk model in rupees. For how private deployment works in India more broadly, see private AI in India.

What are the three places private AI can run?

OptionWhat it meansYou manageBest fit
VPC in an Indian cloud regionWhat it means: GPU servers inside a private network in your own AWS, Azure or Google Cloud account, in an Indian regionYou manage: The model server, network rules, keys, logs and patching of what you install. The cloud provider runs the building and the hardwareBest fit: Teams already on that cloud, with security tooling and identity set up there
On-premiseWhat it means: GPU servers you own, in your office or in rented rack space, on your own networkYou manage: Everything: hardware, power, cooling, network, firmware, the operating system, the model server and sparesBest fit: Strict data-control rules, existing data centre capacity, steady load that justifies buying hardware
Indian GPU cloudWhat it means: GPUs rented by the hour or on commitment from an India-based providerYou manage: The model server and your own security layer. Whether you get private networking, managed storage and similar services depends on the providerBest fit: Teams that want GPU supply and rupee billing, and can run the surrounding system themselves

The options can be mixed. The first question is not which is best. It is which data must go where, who may touch it and who is on call when it fails.

Does DPDP require you to host AI in India?

No. The Digital Personal Data Protection Act, 2023 does not require data to stay in India. Section 16 lets the government restrict transfers of personal data to countries it notifies. As of November 2026 none are notified, and Section 16(2) keeps stricter laws in force, such as RBI's rules on storing payment data. So hosting in India does not satisfy DPDP on its own. It does avoid any future Section 16 restriction and helps with sector localisation rules.

What DPDP does require is the same whichever option you pick. You stay responsible for personal data a processor handles for you (Section 8(1)). You need a valid contract with that processor (Section 8(2)), and your safeguards must cover it (Section 8(5)). The core duties, including Rule 6 security safeguards, start on 13 May 2027. Whether your hosting meets them is a question for your counsel. For foreign AI APIs, read DPDP and cross-border AI APIs.

When does a VPC in an Indian cloud region make sense?

A VPC suits teams that already run their product in AWS, Azure or Google Cloud. Identity, logging and network rules already exist, so the model server joins an environment you know. The provider operates the data centre. You own the configuration, and that is where most mistakes happen.

The Indian regions to know, from each provider's documentation (checked 20 November 2026):

ProviderIndian regionsNote
AWSIndian regions: Mumbai (ap-south-1) and Hyderabad (ap-south-2), three availability zones eachNote: Hyderabad needs opt-in
AzureIndian regions: Central India (Pune, three availability zones), India South Central (Hyderabad, three availability zones), South India (Chennai) and West India (Mumbai), both without availability zonesNote: Zone support differs by region
Google CloudIndian regions: asia-south1 (Mumbai) and asia-south2 (Delhi)Note: GPU types differ by zone; check before planning

Check GPU availability and quota in your chosen region before you design anything. Having a region in India does not mean it offers the GPU you need in the zone you want.

A VPC gives you three things: private networking, your own identity and access rules, and logs in your account. To make it genuinely private, add the controls below. Treat them as a checklist, not a guarantee.

  • No default internet path. Block outbound traffic from the model subnet. Allow only named destinations, such as your package mirror, and review the list.
  • Private model files. Download model weights once into your own storage, check their hashes, and serve from there. Do not let production servers pull from a public hub at start-up.
  • Narrow access. Give people and services only the roles they need, with no shared admin accounts.
  • Logs you own. Decide what prompts and outputs you log, who can read them and how long they live. Raw prompts often hold personal data.
  • A way to delete. Know where prompts, embeddings, caches and backups live, so an erasure request or a retention limit can reach them.

When does on-premise make sense?

A small, tidy server room with one open equipment rack and neatly bundled cables, lit by cool overhead light, the far wall softly out of focus.

On-premise means you own the machines. That gives you the most direct control over where data sits, and makes you the operator of everything below the model. It fits when a rule or contract requires data to stay on equipment you own, when you already have space, power and staff, or when steady load keeps the hardware busy. It fits poorly when demand is spiky, the model is undecided or nobody has run GPU servers before.

Things teams forget to cost for on-premise:

  • Lead time. Buying and installing GPU servers takes time. Rented GPUs can usually be running much sooner.
  • Power and cooling. GPU servers draw far more power per rack than ordinary servers. Check your site's capacity before you order.
  • Spares and support. A failed GPU or network card must be replaced. Ask what warranty and replacement time you get.
  • Patching, backups and recovery. Drivers, firmware and serving software need regular updates, and you build and test recovery yourself.
  • People. Someone must run all of the above. Read do you need an MLOps team to self-host an LLM for the staffing side before you commit.

We give no hardware prices here, because they depend on the supplier and the date. Get written quotes for the full system, including support, and compare them with rental over the period you expect to keep the hardware.

What do Indian GPU clouds offer?

We have not tested or ranked any Indian GPU provider. This section records what each says publicly, checked on 20 November 2026.

  • E2E Networks. Its site shows "MeitY Empaneled" and "NSE: E2E", and lists B200, H200, H100, A100 and RTX PRO 6000 GPUs. It publishes prices in rupees, which we use below.
  • Yotta (Shakti Cloud). Its site lists "NVIDIA H100 and L40S GPUs", says B200s are coming soon, and names data centres in Navi Mumbai and Greater Noida. It shows no public price. The MeitY-empanelled label on its site belongs to Yntraa Cloud, not to Shakti Cloud, so check which legal entity you would contract with.
  • Other providers. The IndiaAI compute portal lists 14 empanelled providers: CMS, Cyfuture, E2E, Ishan, JPL, Locuz, NxtGen, NTT, Neysa, Orient, Sify, Tata Communications, Vensysco and Yotta.

Empanelment is a listing, not a security certificate. Ask for the evidence directly:

  • Which data centre and city will hold your servers and storage, and who operates it?
  • Which legal entity signs the contract, and will it sign a data processing agreement?
  • Is private networking available, and can you block all outbound traffic from your GPU servers?
  • Which GPU type and how many can you get, under what commitment, and can you scale at short notice?
  • What are the support hours, and what does the provider keep about your usage?
  • Which independent security audit reports can it share? Read the scope, not just the title.

How much does a month of GPU rental cost in India?

Here is one provider's public list price, so you can see the order of magnitude. E2E Networks publishes these rates at e2enetworks.com/pricing.json (checked 20 November 2026, rupees per hour, excluding GST). The monthly column assumes 730 hours, which is a GPU left on around the clock.

GPURate per hourPer month (730 hours)
NVIDIA H100 80GB, on-demandRate per hour: ₹379.05Per month (730 hours): ₹2.77 lakh
NVIDIA H100 80GB, committed (3 months or more)Rate per hour: ₹347.70Per month (730 hours): ₹2.54 lakh
NVIDIA A100 80GB, on-demandRate per hour: ₹189Per month (730 hours): ₹1.38 lakh
NVIDIA H100 80GB, spotRate per hour: ₹70Per month (730 hours): ₹0.51 lakh (can be taken back; not for a live service)

A production service rarely runs on one GPU, because you need spare capacity and room for peak traffic. Two on-demand H100s cost about ₹5.53 lakh a month at these rates, before GST, storage, networking, monitoring and the people who run it. Hyperscaler GPU prices vary by instance, region and commitment, so get a quote and compare like with like. A rate card is not a total cost: add egress, storage, support tiers and staff time. The cost model in rupees turns these rates into a break-even against an API.

Who can use IndiaAI subsidised GPU compute?

The Government of India runs a compute portal at compute.indiaai.gov.in (last updated 14 November 2026 when we checked). IndiaAI subsidises approved projects, and users pay the rest. Eligibility follows the portal's criteria, which we summarise here. Read the portal before you plan on it.

ApplicantEligibility on the portal
Startups and MSMEsEligibility on the portal: DPIIT recognition, AI/ML experience, and either ₹50 lakh annual revenue or ₹100 lakh funding
Early-stage startupsEligibility on the portal: 4 years of AI/ML experience, plus a letter from the IndiaAI CEO or a JS-rank officer
OthersEligibility on the portal: Researchers, academia, students, IndiaAI fellows and government

The process, as the portal describes it:

  • Requests are filed between the 1st and 25th of each month.
  • Requests under 5,000 GPU-hours are auto-approved. Larger requests go to a committee (PMEC).
  • Approvals are listed on the 10th of the following month and are valid for 30 days.
  • The compute comes from empanelled providers. Check where your data would sit and who you contract with.

A PIB release of 25 March 2026 said more than 38 thousand GPUs had been onboarded against a target of 100,000. We could not confirm a newer count, so treat "38,000+ (PIB, March 2026)" as the latest figure we can cite.

What this means for planning: it is a monthly, per-request subsidy. Do not plan production capacity on it. It suits experiments, fine-tuning runs and proofs of concept. A live service needs capacity you can count on past the next approval cycle, so it needs a standing contract with a provider. We have not verified the subsidy rates, so we give no figure. If you are foreign-owned or not DPIIT-recognised, do not assume you qualify. The portal's criteria decide.

How big a GPU do you need to run an open-weight model?

A steel tape measure laid across an empty metal shelf in a server room, soft overhead light, the background out of focus.

Start with memory. The weights must fit in GPU memory, with room left for the key-value cache that grows with context length and the number of users. A rough estimate for the weights is parameters multiplied by bytes per parameter: about 2 bytes in 16-bit (BF16), 1 in 8-bit (FP8 or INT8) and about 0.5 in 4-bit. This is a rough estimate, not a benchmark. Real use needs more.

Model size16-bit (BF16)8-bit4-bit
8 billion parameters16-bit (BF16): about 16 GB8-bit: about 8 GB4-bit: about 4 GB
31 billion parameters16-bit (BF16): about 62 GB8-bit: about 31 GB4-bit: about 16 GB
70 billion parameters16-bit (BF16): about 140 GB8-bit: about 70 GB4-bit: about 35 GB

These are weights only. A 70 billion model at 8-bit leaves only about 10 GB of an 80 GB card for the cache and overhead, so plan for more than one GPU or a smaller model. Quantisation "trades off model precision for smaller memory footprint", in the words of the vLLM documentation, so test quality on your own tasks before you adopt it. The 8-bit FP8 W8A8 format needs Ada, Hopper or AMD GPUs.

vLLM is a widely used open-source server for this job. Its latest release when we checked was v0.34.0 (16 November 2026). `vllm serve ` starts an OpenAI-compatible server, so applications written for a hosted API can point at your own endpoint. Set the port with `--port 8000`. The flags that matter most for fit are:

  • `--gpu-memory-utilization`, the share of GPU memory vLLM may use, from 0 to 1.
  • `--max-model-len`, the context length, prompt plus output. A lower value frees memory for the cache.
  • `--tensor-parallel-size`, which splits the model across GPUs. It defaults to 1.
  • `--kv-cache-dtype`, which can store the cache in a smaller format such as fp8.

How many users one GPU can serve depends on the model, prompt length, settings and your latency target. We publish no throughput or latency numbers, and you should distrust any that do not come with your workload. Run a load test with your own prompts at your expected peak, measure from where your users sit, and decide from that.

Check the model's licence too, because it is set per model, not per vendor. Llama 4 is open-weight under Meta's community licence, which requires a licence request above 700 million monthly active users. Gemma 4 is under Apache 2.0. Qwen and Mistral mix permissive and restricted licences across models. Read the licence on each model card, and see open-weight model licences.

Which option should you choose?

The first question you answer "yes" to usually settles it.

QuestionIf yes, lean towards
Must the data stay on hardware you own, by rule or contract?If yes, lean towards: On-premise
Does your product already run in AWS, Azure or Google Cloud, with security tooling in place?If yes, lean towards: VPC in an Indian region of that cloud
Do you need GPUs that your cloud region cannot supply, or want rupee billing and a provider you can call?If yes, lean towards: Indian GPU cloud
Is the work an experiment or a short project that could use a 30-day allocation?If yes, lean towards: IndiaAI compute, if you qualify, then a standing option for production
Is your load spiky, or are you still choosing a model?If yes, lean towards: Rent first (VPC or GPU cloud). Do not buy hardware yet
Do you lack anyone to run GPU servers?If yes, lean towards: Fix that first. Without an operator, none of the three options works

Whichever you choose, keep the application separate from the model host. Put an OpenAI-compatible interface in front, keep prompts and settings in version control, and write down how to rebuild the host. A later move then becomes a project, not a rewrite.

If you want help placing your AI workloads, Mono's Own Your AI service sets up open-weight models on a client's own servers or cloud account. Mono does not rent GPUs and is not a certification body or law firm. Your counsel confirms the legal position.

Sources: Digital Personal Data Protection Act, 2023, Sections 8 and 16, and the Digital Personal Data Protection Rules, 2025, Rule 6, Rule 15. AWS Regions documentation (docs.aws.amazon.com/global-infrastructure/latest/regions/aws-regions.html), Azure regions list (learn.microsoft.com/en-us/azure/reliability/regions-list) and Google Cloud GPU regions and zones (cloud.google.com/compute/docs/gpus/gpu-regions-zones), checked 20 November 2026. E2E Networks (e2enetworks.com and e2enetworks.com/pricing.json) and Yotta Shakti Cloud (shakticloud.ai), checked 20 November 2026. IndiaAI compute portal (compute.indiaai.gov.in), last updated 14 November 2026. PIB, 25 March 2026 (pib.gov.in, NoteId 156786). vLLM documentation and releases (docs.vllm.ai, github.com/vllm-project/vllm). Llama 4 Community License (dev.meta.ai/llama/llama4/license) and Gemma 4 licence (ai.google.dev/gemma/docs/gemma_4_license). Prices are provider list prices and change often. Last reviewed 20 November 2026.

FAQ

Is a VPC the same as private AI?

No. A VPC is a private network inside your cloud account. It gives you a place to build private AI, but the system is only private if you also block outbound traffic from the model host, limit who can access it, control the logs and keep model files in your own storage. A model in a VPC that can call the open internet or ship prompts to a third-party tool is not private.

Does DPDP require AI to be hosted in India?

No. The DPDP Act does not require data to stay in India. Section 16 allows the government to restrict transfers to notified countries, and as of November 2026 none are notified. Sector rules can still require local storage, such as RBI's payment-data rules. You remain responsible for any processor you use, whichever way you host.

Which Indian cloud regions can run private AI?

AWS has Mumbai and Hyderabad, Azure has Central India (Pune), India South Central (Hyderabad), South India (Chennai) and West India (Mumbai), and Google Cloud has Mumbai and Delhi. A region existing does not mean it has the GPU you need. GPU types differ by zone, so check availability and quota before you plan.

Who can use IndiaAI subsidised GPU compute?

According to the IndiaAI compute portal, last updated 14 November 2026, eligible users include DPIIT-recognised startups and MSMEs with AI/ML experience and either ₹50 lakh annual revenue or ₹100 lakh funding, plus researchers, academia, students, IndiaAI fellows and government. Requests are filed monthly and approvals last 30 days, so do not plan production capacity on it. Eligibility follows the portal's criteria.

How much does it cost to rent an H100 in India?

On E2E Networks' public price list, checked 20 November 2026, an NVIDIA H100 80GB costs ₹379.05 an hour on-demand and ₹347.70 an hour on a commitment of three months or more, excluding GST. Left on for a 730-hour month, that is about ₹2.77 lakh and ₹2.54 lakh. Other providers list different rates, and some publish none.

Can I move from rented GPUs to on-premise later?

Yes, if you plan for it. Serve the model behind an OpenAI-compatible interface, keep your prompts, settings and deployment scripts in version control, and keep model files in your own storage. Then moving the model host is an infrastructure project and the application does not change.

Decide where your private AI should run.