An on-premise AI server in India costs about ₹12.5–16 L per 96 GB GPU card or ₹68–80 L for a 4-card server, at cost, ex-GST, moving with the market. The software sits on top: a BiltIQ AI Factory platform-only site licence of ₹35 L, or ₹50–90 L with Agent OS workflows, plus a managed service of ₹16–18 L a year.
At a glance
| Item | Price (ex-GST) | Note |
|---|---|---|
| RTX PRO 6000 card, 96 GB | ≈₹12.5–16 L | 600 W peak; Indian reseller prices, Sep 2026 |
| 4× RTX PRO 6000 server | ₹68–80 L | ≈3.5 kW, ≈1 TR cooling, 6 kVA UPS |
| Platform-only site licence (AI Factory serving, no agent workflows) | ₹35 L one-time | |
| Full site licence with Agent OS workflows | ₹50–90 L one-time | By segment; Voice channel included |
| Managed service | ₹16–18 L a year | Updates, monitoring, quarterly review, tuning |
| Paid pilot | ₹5–8 L, 4–6 weeks | Credited against the licence |
| Hardware | At cost | Passed through; never BiltIQ revenue |
Why is hardware "at cost"?
BiltIQ sizes the server to your workload, buys at reseller prices and passes the hardware through. It is never counted as revenue, which is why the site shows indicative ranges rather than a price: the ranges move with the market. Site power and cooling are yours.
What does a full site add up to?
Illustrative, from the published ranges, ex-GST.
| Configuration | Hardware | Licence | Managed service, 3 years | Three-year total |
|---|---|---|---|---|
| 4-card server, platform only | ₹68–80 L | ₹35 L | ₹48–54 L | ₹1.51–1.69 Cr |
| 4-card server, with Agent OS workflows | ₹68–80 L | ₹50–90 L | ₹48–54 L | ₹1.66–2.24 Cr |
| One 96 GB card in your own chassis | ≈₹12.5–16 L + chassis | ₹35 L or ₹50–90 L | ₹48–54 L | Scoped in the pilot |
The paid pilot (₹5–8 L) comes off the licence if you proceed, so it is not an extra line.
When does owning beat renting?
Cloud models are metered: every new user adds to the bill. Owning costs a one-time licence plus a yearly service, with hardware at cost, so the cost per query falls as use grows. At low volume cloud is cheaper; at high, steady volume ownership is. BiltIQ works out your case on the scoping call. The live post on on-prem vs cloud AI cost sets out when owning beats renting and when it does not.
Subscribe, for those who do not want hardware: API tokens from BiltIQ's GPU cluster in India at ₹8 per 1M input and ₹28 per 1M output, coming in 2027.
What does the managed service cover for ₹16–18 L a year?
- Model and platform updates, including security updates, installed by BiltIQ
- Remote health monitoring of the server and services, with your permission and no document access
- A quarterly health and usage review
- Agent and workflow tuning (the number of changes is set per contract)
- Support in business hours IST, Mon–Sat 9:00–19:00; site-down first response within 4 hours, remote first
- GPU faults through the OEM warranty, with the claim managed by BiltIQ; no spares held, your site runs on the remaining GPUs meanwhile
What does a reference install look like?
An engineering college runs AI Factory on its own GPUs, 4 DGX Spark, as a reference install set up by BiltIQ. It is not a paying customer, and BiltIQ publishes no throughput figures for it. Each model runs on its own GPU node; multi-node serving is on the roadmap.
What else do you need in the room?
The published figures for the 4-card server: ≈3.5 kW draw, ≈1 TR of cooling, a 6 kVA UPS. A single card peaks at 600 W. Site power and cooling are the buyer's. The post "Sizing an on-premise LLM box: one 96 GB GPU or four?" covers whether you need one card or four.
How do you budget for year one?
Illustrative year-one budget for a 4-card server with Agent OS workflows, from the published ranges, ex-GST:
| Line | Amount | Note |
|---|---|---|
| Paid pilot | ₹5–8 L | Credited against the licence if you proceed |
| Hardware | ₹68–80 L | At cost, passed through, never BiltIQ revenue |
| Site licence | ₹50–90 L | By segment; the exact figure is confirmed in writing after the scoping call |
| Managed service, year one | ₹16–18 L | Updates, monitoring, quarterly review, tuning |
| Site power and cooling | Yours | ≈3.5 kW, ≈1 TR, 6 kVA UPS |
| GST | 18% on each line |
Years two and three carry the managed service and power only, which is why the three-year totals above are dominated by year one.
What moves the price inside the ranges?
- Segment. Government and large-enterprise sites sit at the top of ₹50–90 L; medium firms lower. Platform-only (no agent workflows) is ₹35 L.
- Workload. The pilot sizes the server; a department's document Q&A may need one card, several teams and a vision model need four.
- Market. Hardware is at cost, so the GPU line moves with reseller prices.
- What you already have. A firm with a chassis and power in place buys cards, not a server.
Nothing in the ranges is negotiated per client in public; BiltIQ publishes the full range and confirms the number inside it in writing.
When should you NOT buy an on-premise AI server?
- Low, bursty volume. Metered cloud is cheaper until use is steady.
- No room for ≈3.5 kW and a tonne of cooling. A 4-card server needs both; one card is easier.
- Your data may leave the building. Then the Subscribe tier from BiltIQ's India cluster (2027) is the cheaper route.
- You want a price before a pilot. The ranges are public; the exact licence inside the range is confirmed in writing after the scoping call, and the server is sized during the pilot.
The figure buyers ask for and BiltIQ does not publish
"How many users per card?" is the first question on most calls, and the honest answer is that it depends on documents, prompt length and model choice. A number printed on a web page would be wrong for most buyers in one direction or the other. The pilot measures it on your files. Treat any vendor's users-per-GPU figure as a claim to test on your own documents, not a specification.
Frequently asked questions
How much does an on-premise AI server cost in India?
Hardware at cost, moving with the market: about ₹12.5–16 L per 96 GB GPU card or ₹68–80 L for a 4-card server. Software on top: ₹35 L platform-only or ₹50–90 L with Agent OS workflows, plus ₹16–18 L a year managed service. All ex-GST.
Does the ₹68–80 L include the software?
No. It is hardware at cost. The site licence and the managed service are separate and listed above.
How many users can one card serve?
It depends on your documents, prompt length and model choice. BiltIQ sizes the server during the paid pilot, before you buy hardware.
Related: Sizing an on-premise LLM box: one 96 GB GPU or four? · Open-weight LLMs for enterprise, 2026: which model for which job · Managed on-premise AI: what the yearly service actually covers · Air-gapped LLM deployment: what changes with no internet · Custom AI on your own servers: when a product is not enough · On-Prem vs Cloud AI Cost Comparison: Priced in Electricity, Not Tokens
Next step: pricing for every figure with its label; AI Factory; scoping call on contact.
Facts checked against the BiltIQ product pages, 5 November 2026.