Short answer: at public list prices checked on 20 November 2026, a two-GPU private LLM setup costs roughly ₹5.7 to 6.2 lakh a month before you serve a single request, and it only beats a mid-priced API model once you pass about one million requests a month of the size modelled here (1,500 tokens in, 300 tokens out). Against a top-tier API model the break-even is about a quarter of a million requests a month. Against a small, cheap API model it is roughly twenty million, which is a volume few teams reach. So cost alone rarely justifies going private. Data control, a pinned model version and a bill that does not move with usage are the stronger reasons, and each comes with engineering work you must staff.
This guide gives you the model and the inputs so you can rerun it with your own numbers. It uses only dated public rates: published API token prices, a public GPU rental price list and a public salary range. It leaves out managed-service fees and anyone's private pricing. It is not legal advice, and the legal reading of the DPDP Act for your business goes to counsel.
How much does an API cost compared with a private LLM, in rupees?

Start with the API side. These are public list prices, checked on 20 November 2026, in US dollars per 1 million tokens, excluding tax. The rupee columns use about ₹93.37 per US dollar, the mid-market rate on 20 November 2026 from open.er-api.com. Use your own treasury rate. Prices apply to standard use for prompts below each vendor's long-context threshold. Batch use is 50% off at both OpenAI and Anthropic.
| Provider and model | Input, USD | Output, USD | Input, ₹ | Output, ₹ |
|---|---|---|---|---|
| OpenAI GPT-6.2 Astra | Input, USD: $8 | Output, USD: $40 | Input, ₹: ₹747 | Output, ₹: ₹3,735 |
| OpenAI GPT-6.2 Sol | Input, USD: $2 | Output, USD: $10 | Input, ₹: ₹187 | Output, ₹: ₹934 |
| OpenAI GPT-6 Luna | Input, USD: $0.10 | Output, USD: $0.50 | Input, ₹: ₹9 | Output, ₹: ₹47 |
| Anthropic Fable 5.2 | Input, USD: $8 | Output, USD: $40 | Input, ₹: ₹747 | Output, ₹: ₹3,735 |
| Anthropic Opus 5.5 | Input, USD: $4 | Output, USD: $20 | Input, ₹: ₹373 | Output, ₹: ₹1,867 |
| Anthropic Sonnet 5.5 | Input, USD: $2 | Output, USD: $10 | Input, ₹: ₹187 | Output, ₹: ₹934 |
| Anthropic Haiku 5.5 | Input, USD: $0.10 | Output, USD: $0.50 | Input, ₹: ₹9 | Output, ₹: ₹47 |
| Google Gemini 4 Carbon, through 30 November 2026 | Input, USD: $2 | Output, USD: $10 | Input, ₹: ₹187 | Output, ₹: ₹934 |
| Google Gemini 4 Carbon, from 1 December 2026 | Input, USD: $4 | Output, USD: $20 | Input, ₹: ₹373 | Output, ₹: ₹1,867 |
The prices fall into four rough tiers: $0.10 and $0.50, $2 and $10, $4 and $20, and $8 and $40. Now turn them into a monthly bill. The formula is simple:
Monthly API cost = requests per month × (input tokens × input price + output tokens × output price) ÷ 1 million
Take one million requests a month, each with 1,500 input tokens and 300 output tokens. That is 1.5 billion input tokens and 0.3 billion output tokens. This shape is an assumption. A chat with long history or a retrieval prompt stuffed with documents will be far larger, and LLM bill shock explains why.
| API tier (input and output, per 1M tokens) | Cost per request | Monthly cost, 1M requests | Same, if batched at 50% off |
|---|---|---|---|
| Small ($0.10 and $0.50) | Cost per request: about ₹0.03 | Monthly cost, 1M requests: about ₹0.28 lakh | Same, if batched at 50% off: about ₹0.14 lakh |
| Mid ($2 and $10) | Cost per request: about ₹0.56 | Monthly cost, 1M requests: about ₹5.6 lakh | Same, if batched at 50% off: about ₹2.8 lakh |
| Upper ($4 and $20) | Cost per request: about ₹1.12 | Monthly cost, 1M requests: about ₹11.2 lakh | Same, if batched at 50% off: about ₹5.6 lakh |
| Top ($8 and $40) | Cost per request: about ₹2.24 | Monthly cost, 1M requests: about ₹22.4 lakh | Same, if batched at 50% off: about ₹11.2 lakh |
Two things to note. First, the bill is variable: it scales with usage and costs nothing when idle. Second, it can change under you. Google's listed price for Gemini 4 Carbon doubles on 1 December 2026, which would take the mid-tier example above from about ₹5.6 lakh to about ₹11.2 lakh a month for the same traffic. Vendors can also change rates for long prompts. Prompt caching and batching can cut real bills well below these list figures.
What does it cost to run a private LLM in India?
A private LLM has three cost lines: GPUs, the people who run them, and everything around them. Here is each from public data.
GPU rental
E2E Networks, an Indian cloud provider, publishes GPU prices in rupees at e2enetworks.com/pricing (checked 20 November 2026, per hour, excluding GST). The table converts them to a month of 730 hours, which is a GPU left on around the clock.
| GPU on E2E Networks | Rate per hour | Per month (730 hours) | Per year |
|---|---|---|---|
| NVIDIA H100 80GB, on-demand | Rate per hour: ₹379.05 | Per month (730 hours): about ₹2.77 lakh | Per year: about ₹33.2 lakh |
| NVIDIA H100 80GB, committed (3+ months) | Rate per hour: ₹347.70 | Per month (730 hours): about ₹2.54 lakh | Per year: about ₹30.5 lakh |
| NVIDIA A100 80GB, on-demand | Rate per hour: ₹189 | Per month (730 hours): about ₹1.38 lakh | Per year: about ₹16.6 lakh |
| NVIDIA H100 80GB, spot | Rate per hour: ₹70 | Per month (730 hours): about ₹0.51 lakh | Per year: about ₹6.1 lakh |
Spot capacity can be taken back, so it suits batch jobs and experiments, not a live service. Other providers list different prices, and the Indian government's IndiaAI compute portal offers subsidised rates to eligible entities. We have not verified its current rates, so check its eligibility rules and prices directly.
Sizing is where teams get surprised. A rough rule: model weights need about the number of parameters times the bytes per parameter. A 70 billion parameter model at 16-bit precision needs roughly 140 GB for weights alone, which is more than one 80 GB card, and you still need room for the context cache. An 8 billion parameter model at 16-bit needs roughly 16 GB. How many requests a GPU can serve depends on the model, the prompt length, the serving software and your latency target. Measure it with a load test on your own traffic. Do not trust a vendor's headline tokens-per-second figure.
People
Someone has to run the serving stack, patch it, watch it, test new model versions and answer the pager. AmbitionBox's self-reported data for the title "AI Engineer" in India shows a typical salary of ₹14.8 to 16.4 lakh a year for 1 to 6 years of experience, from about 5,600 salaries, updated 19 November 2026. That is salary only, before employer costs. It is also a broad title, and a person who can run GPU inference in production may cost more. The model below assumes half of one such person's time at the midpoint, ₹15.6 lakh a year, which is about ₹0.65 lakh a month. Change that if your setup needs more.
Everything else
Storage, network egress, monitoring, evaluation runs, security review, backups, a staging environment and GST are extra, and we have not priced them because they depend on your setup. A realistic budget adds a margin for them.
When does a private LLM become cheaper than an API?
Compare a fixed monthly cost with a variable one. We model two private setups, both from the rates above:
- Lean: one A100 80GB on-demand, no failover, plus half an engineer. ₹1.38 lakh + ₹0.65 lakh = about ₹2.03 lakh a month. Suits a pilot or an internal tool where an outage is tolerable.
- Production: two H100 80GB on committed rates, so one can fail while the other serves, plus half an engineer. ₹5.08 lakh + ₹0.65 lakh = about ₹5.73 lakh a month, or about ₹68.7 lakh a year. On on-demand rates it is about ₹6.18 lakh.
Break-even is the fixed cost divided by the API cost per request. The table shows how many requests a month you need before the private setup costs less, using the 1,500 and 300 token shape.
| Compared with this API tier | Lean setup breaks even at | Production setup breaks even at |
|---|---|---|
| Small ($0.10 and $0.50) | Lean setup breaks even at: about 7.3 million requests a month | Production setup breaks even at: about 20.5 million requests a month |
| Mid ($2 and $10) | Lean setup breaks even at: about 0.36 million | Production setup breaks even at: about 1.0 million |
| Upper ($4 and $20) | Lean setup breaks even at: about 0.18 million | Production setup breaks even at: about 0.5 million |
| Top ($8 and $40) | Lean setup breaks even at: about 0.09 million | Production setup breaks even at: about 0.26 million |
One million requests a month is about 23 requests a minute on average, and peaks run higher, so check that your GPUs can carry your peak at an acceptable speed. If they cannot, the true fixed cost is higher and the break-even moves out.
Read the table with three cautions.
- Compare like with like. The private setup runs an open-weight model. It is only a fair comparison against the API tier whose quality your own tests show is equivalent for your task, which is often a lower tier than you assume, and sometimes a higher one.
- Batch changes the answer. If your work is not live, such as overnight document processing, API batch pricing halves the variable bill and doubles the break-even volume.
- Idle time is yours. A rented GPU bills at night and on weekends. An API does not. If your traffic is spiky, the average GPU utilisation can be low, and the fixed cost is spread over fewer requests.
If your numbers sit near the break-even, the decision is not about cost. Look at the risks.
What are the risks of each option?

| Risk | Public API | Private LLM |
|---|---|---|
| Price change | Public API: The vendor can change rates. Gemini 4 Carbon doubles on 1 December 2026. | Private LLM: Rental prices change too. A committed rate may hold the GPU price for the term; check the provider's terms. |
| Model changes | Public API: Vendors retire older versions on their own schedule. Check each vendor's deprecation page. | Private LLM: You pin a version and upgrade when you choose. You also own the testing. |
| Availability | Public API: You depend on the vendor's uptime and your rate limits. | Private LLM: You depend on your own uptime. You need failover and monitoring. |
| Data exposure | Public API: Prompts leave your network. Check the vendor's retention and training terms and sign a contract that fits your obligations. | Private LLM: Prompts stay in your account or servers. You are responsible for securing them. |
| Quality | Public API: Top models are strong at hard reasoning and long, messy tasks. | Private LLM: Open-weight models can match on narrow, well-defined tasks. They may fall behind on hard ones. |
| Operations | Public API: Little to run. You manage keys, cost and logging. | Private LLM: You run serving, scaling, patches, evaluation and on-call. |
| Staff dependency | Public API: Low. | Private LLM: High. If the one person who knows the stack leaves, you are exposed. |
| Idle cost | Public API: None. | Private LLM: GPUs bill whether used or not. |
The pattern is clear. The API moves risk to the vendor and gives you a variable bill. A private setup moves control to you, with a fixed bill and an operations job.
Does the DPDP Act force you to run a private LLM?
No. The DPDP Act does not ban foreign AI APIs and does not require data to stay in India. Section 16 lets the government restrict transfers to notified countries, and as of November 2026 none are notified. Section 16(2) keeps stricter laws in force, such as RBI's rules on storing payment data, so a regulated business may have a sector reason to keep data in India even though the Act itself does not.
For the cost model, what matters is that the duties apply either way. You stay responsible as the Data Fiduciary for processing a provider does on your behalf (Section 8(1)), you need a valid contract with it (Section 8(2)), and your security safeguards must cover it (Section 8(5)). Rule 6 sets minimum safeguards, and the core duties start on 13 May 2027. Hosting in India keeps you clear of any future Section 16 restriction and helps with sector localisation rules, but it does not replace those controls, so budget for access control, logging and an erasure path in either option.
For a data-flow map of an AI feature, read DPDP compliant AI architecture. For how private deployment works in India, see private AI in India.
Open-source or proprietary LLM: what changes for an enterprise?
This choice is separate from API against private, though the two are linked. You can call an open-weight model through a hosted API, and you can run some proprietary models in a vendor's private environment. Think of it as three questions.
- Weights or no weights. "Open-weight" means you can download and run the model yourself. It does not always mean open source. Licences differ in what they allow for commercial use, large user counts, fine-tuning and use of the outputs. Read the licence for the exact model and version, and have counsel review it.
- Quality on your task. Public benchmarks do not predict how a model does on your tickets, contracts or code. Build a test set from your own real examples, remove personal data, and score each candidate on it. Many narrow tasks, such as classification, extraction and redaction, work well on smaller open models. Hard open-ended reasoning is where the gap is most likely to show.
- Control. With open weights you can pin a version, fine-tune on your data inside your own account, and keep the model as long as you like. With a proprietary API you gain the vendor's updates and lose control of the timing.
When should a data-sensitive team go private?
Use this checklist. The more of the first list that is true, and the more of the second that is false, the stronger the case.
- For private: the data cannot leave your network by contract, by sector rule or by your own policy; your volume is steady and high on a narrow task; you need to reproduce an answer from a pinned model version; you can staff on-call; the open model passes your test set.
- For an API: volume is low or spiky; you need frontier-level reasoning; no one can run GPUs; you are still piloting; the data is low risk or already covered by a contract you are comfortable with.
Many teams end up with a mix. Route a high-volume, narrow, data-heavy step, such as redacting personal data or classifying documents, to a small private model. Send the hard, low-volume steps to an API, with personal data removed first. This keeps the fixed GPU cost small and the data exposure smaller. If you are not sure where your product calls models, a technical audit can map it.
If you decide to go private, Mono's Own Your AI service sets up open models in your own cloud account or on your own servers. We do not give prices here, and the rates in this article come from public sources, not from our own cost data.
How do you rerun this model with your own numbers?
- Measure your traffic. Pull a month of logs: requests, input tokens and output tokens per request, and peak requests per minute.
- Price the API. Apply the formula to each tier you would accept, with and without batching and caching.
- Test the open model. Run your own examples through the candidates. Record quality, speed and the GPU memory needed.
- Price the private setup. GPU hours at your committed rate, times the number of cards including failover, plus people and the extras above.
- Compare, then stress it. Double the traffic, then halve it. Raise the API price by 100%, as Gemini 4 Carbon's list price does on 1 December. Raise GPU price by 20%. See whether the answer holds.
- Add the non-cost factors. Data control, version pinning, staffing and the cost of an outage.
Sources: OpenAI API pricing (openai.com/api/pricing), Anthropic pricing (anthropic.com/pricing) and Google Gemini API pricing (ai.google.dev/gemini-api/docs/pricing), all checked 20 November 2026. E2E Networks GPU price list (e2enetworks.com/pricing.json), checked 20 November 2026. AmbitionBox, AI Engineer salary in India (ambitionbox.com/profile/ai-engineer-salary), self-reported, updated 19 November 2026. Exchange rate from open.er-api.com, 20 November 2026. Digital Personal Data Protection Act, 2023, Sections 8 and 16. Digital Personal Data Protection Rules, 2025, Rule 6. Prices are vendor list prices and change often. Last reviewed 20 November 2026.
FAQ
Is a private LLM cheaper than the OpenAI or Claude API?
Only at enough volume. At public prices checked on 20 November 2026, a two-GPU production setup on committed H100 rates plus half an engineer costs about ₹5.7 lakh a month. It beats a mid-priced API model at roughly one million requests a month of 1,500 input and 300 output tokens, and a top-tier model at roughly a quarter of a million. Below that, the API is cheaper.
How much does a GPU cost to rent 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 3-month or longer commitment, excluding GST. An A100 80GB costs ₹189 an hour. Left running all month, an on-demand H100 costs about ₹2.77 lakh.
Does the DPDP Act require me to host an LLM 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. Stricter sector laws still apply, for example RBI's payment-data storage rules. You remain responsible for your processors, so you need a contract and safeguards whichever way you host.
What does a private LLM cost besides GPUs?
People, mainly. Someone must run the serving stack, test new models and handle outages. AmbitionBox's self-reported data puts typical AI engineer pay in India at ₹14.8 to 16.4 lakh a year, before employer costs. You also pay for storage, networking, monitoring, evaluation, security review and GST, which depend on your setup.
Can I use both a private LLM and an API?
Yes, and many teams should. Run a small open model privately for high-volume, narrow, data-heavy steps such as classification or redaction, and call an API for hard, low-volume reasoning with personal data removed first. This keeps the fixed GPU bill small and limits how much data leaves your network.
