On-premise LLM in India: a private AI assistant on your own GPU server
An on-premise LLM is a large language model that runs on GPU servers inside your own building, so prompts and documents do not go to an outside model API. BiltIQ AI Factory sets one up on your server in India, with open-weight models, serving and monitoring, and runs it as a managed service. Hardware is passed through at cost.
We'll walk you through it on sample data.
Why enterprises want a private ChatGPT-style assistant
- Staff may paste company files into public AI tools. A private assistant gives them a sanctioned place to ask.
- The DPDP Act. The DPDP Act sets duties and penalties for how organisations handle personal data.
- Metered bills. Cloud models charge per use, so every new user adds to the bill.
- Model origin. Defence and government buyers check each model's origin under Public Procurement No. 1.
What you get with BiltIQ AI Factory
| Part | What it does |
|---|---|
| Model serving | Serves open-weight models on your GPUs. A single node per model today; multi-node serving is written, not yet in use. |
| Answers from your documents | Your SI answers from your own files, with page citations. |
| Ops monitoring | Watches the health of your server and services remotely, with your permission and no document access. |
| Privacy library | Our MIT-licensed biltiq-privacy library detects Indian personal data such as Aadhaar and PAN and pseudonymises it, set per deployment. |
| Platform gate (with Agent OS) | Add Agent OS and outbound calls are blocked by default; irreversible and high-consequence actions pause for human approval. |
| Sizing | We size the server to your workload. Hardware is at cost and is never our revenue. |
Which AI models run on your private LLM server
Model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed.
Fine-tuned domain models are a roadmap item after our next GPU expansion; we promise none before then.
GPU server for AI: indicative price, at cost
| Option | Indicative, at cost, ex-GST | Power |
|---|---|---|
| RTX PRO 6000 card, 96 GB | ≈₹12.5–16 L | 600 W peak |
| 4× RTX PRO 6000 server | ₹68–80 L | ≈3.5 kW · ≈1 TR cooling · 6 kVA UPS |
Indicative, at cost; varies with the market. Site power and cooling are yours. As a guide, on our own DGX Spark a 27B model stays usable for about 5 concurrent users. Your real load decides; the GPU sizing sheet helps you estimate it.
On-premise LLM pricing in India
- Platform-only site (AI Factory serving, no agent workflows): ₹35 L one-time
- Full site licence with Agent OS workflows: ₹50–90 L one-time
- Managed service: ₹16–18 L a year (updates, monitoring, quarterly review, agent tuning; business hours IST support, 4-hour site-down response)
- Hardware: at cost
- Paid pilot: ₹5–8 L, 4–6 weeks, credited against the licence
All ex-GST. A subscription to our GPU cluster in India (API tokens at ₹8 per 1M input / ₹28 per 1M output) is planned. Subscription: not yet on sale.
Where our SI platform runs today
AI Factory is installed on an engineering college's own GPUs (4 DGX Spark); setup by BiltIQ, unpaid. There is no paid customer install yet. We run our own 4-node GPU fleet, one model per node.
What still touches the cloud
- Outbound calls. AI Factory serves the models on your server. Blocking outbound calls by default comes with Agent OS (platform gate); without it, your own network rules apply.
- Apps on top. Business OS stores files in GCS or S3 cloud storage today. AI Teacher's embeddings, vision and voice default to cloud services and can be repointed.
Key facts: On-premise LLM India
- Platform-only site licence ₹35 L one-time; full site licence with Agent OS workflows ₹50–90 L; managed service ₹16–18 L a year. Ex-GST.BiltIQ price list, Oct 2026
- Hardware at cost, indicative: ≈₹12.5–16 L per 96 GB GPU card, ₹68–80 L per 4-GPU server.Indian reseller prices, Sep 2026
- Model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed.BiltIQ code review, Oct 2026
- Serving runs a single node per model today; multi-node serving is written, not yet in use.BiltIQ code review, Oct 2026
- One install so far: an engineering college's own 4 DGX Spark, set up by BiltIQ, unpaid. No paid customer install yet.BiltIQ records
- Outbound calls are blocked by default only with Agent OS (platform gate).BiltIQ code review, Oct 2026
FAQ
Can we run a private ChatGPT-style AI on our own servers in India?
Yes. BiltIQ AI Factory installs open-weight models on your own GPU server and serves them inside your building, so staff can ask questions of your documents without prompts going to an outside model API. Answers cite the page. BiltIQ runs your SI as a managed service from India. Add Agent OS if you also want outbound calls blocked by default.
How much does an on-premise AI server with an LLM cost in India?
Two parts. Hardware is at cost and indicative: about ₹12.5–16 L per 96 GB GPU card, or ₹68–80 L for a 4-GPU server. Software: a platform-only site licence of ₹35 L, or ₹50–90 L with Agent OS workflows, plus ₹16–18 L a year for the managed service. All ex-GST. A ₹5–8 L paid pilot comes first, credited against the licence.
Which open-weight AI models can we run on premises?
BiltIQ is model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed. You can swap models as your needs change. Each model runs on a single node today. Fine-tuned domain models come after our next GPU expansion, and nothing is promised before then.
How many users can one GPU serve for on-premise AI?
It depends on document length, prompt size and model choice, so there is no single figure for a 96 GB RTX PRO 6000 card. Our free GPU sizing sheet gives a first estimate, and BiltIQ sizes your SI server against your real workload during the paid pilot, before you buy hardware.
Can a DGX Spark run on-premise AI for our team?
For a small team or a pilot, yes. On our own DGX Spark, a 27B model stays usable for about 5 concurrent users. It suits embeddings, smaller models and evaluation, not large-model production serving for many users. An engineering college runs AI Factory on four of its own DGX Spark units, set up by BiltIQ, unpaid.
Does an on-premise AI server keep all our data inside?
The models and your prompts stay on your server. Blocking outbound calls by default comes with Agent OS, through the platform gate; with AI Factory alone, your own network rules apply. Some apps on top still use cloud services today: Business OS stores files in GCS or S3, and AI Teacher's embeddings, vision and voice default to the cloud.
Size your private SI server with an engineer
We'll walk you through it on sample data. Or write to [email protected], or call or WhatsApp +91 89868 60088 (messages go straight to our CRM).