Compute & Cost
Cost Benchmarks From Real Indian Deployments
· 9 minute read
We will not invent a dataset of real Indian deployments and average it. We will give you a method, public anchors, and order-of-magnitude bands labelled as such.
The most requested slide in a 2026 steering pack is what did other departments spend. The most honest answer is: they have not published a dataset you can average. Tender PDFs are incomplete. Newspaper crores mix a model, a hall, a skill programme and a press note. Vendor case studies are advertisements. If we printed a table of twenty unnamed Indian deployments with rupees, we would be adding a twenty-first fiction to your file.
This data-study is a method for constructing a benchmark you can defend, plus order-of-magnitude bands drawn from public price-shaped facts — GeM-visible hardware classes, IndiaAI hour commentary, electricity stacks, and the shape of government SI quotes. Bands are not measurements of your neighbour. They are fences so a proposal that is 20× off can be seen without pretending you ran a census.
If a competitor hands you a neat national average, ask for the microdata. If they cannot produce it, you already have our article.
What a benchmark is allowed to be
A benchmark is a comparison against a named source. Acceptable sources: your own last project, a GeM price you can screenshot, an IndiaAI portal extract against your login, a SERC tariff order, a published tender BoQ with a date, a CAG or PAC figure that names the project. Unacceptable sources: industry standard in India, an anonymised consultancy table, a US dollar blog times 83, a vendor's typical PSU.
When you have only two or three acceptable points, say so. A triangle is not a distribution. Order-of-magnitude — lakhs versus crores versus tens of crores — is often the only honest national statement available in 2026. Separate the object. A helpdesk draft agent on one card is not a sovereign foundation model. Averaging them is how you get a number that helps nobody and hurts every concurrence.
Bands you may quote as bands
Standing inference, small model, one or two datacentre or professional GPUs, on-prem, one workflow, including a modest SI and a gold set: think in tens of lakhs to a couple of crores of first-year cash, not in a precision rupee. Hardware list prices move; GeM and OEM quotes on the week of sanction are the source, not this sentence.
A departmental multi-workflow node or a small SDC share — a handful of datacentre GPUs, more SI, more integration — often lives in the low-single-digit crores of first-year cash. If someone quotes ₹40 crore for this object, ask which hall they have hidden in the number. A training-capable mini-cluster (eight and up of H100-class cards), with power, facility and a serious SI: tens of crores is a shape you should be ready to see, and a shape you should refuse if your workload is a helpdesk. IndiaAI exists partly so you do not have to own this object for a one-month eval.
Metered hours: public commentary on mission bids has put unsubsidised H100-class hours in a broad low-hundreds-of-rupees band, with subsidised effective hours for some eligible users described around the ₹65–₹100 region. Use the portal. Annualise only with a cap. Recurring: AMC on hardware commonly appears in government IT as a high-single-digit to mid-teens percentage of a defined value per year after warranty — a market band, not a law. Power on a small always-on node: low-to-mid lakhs per year in many HT settings, method in the electricity article. Idle can exceed both if you bought the training object and run the helpdesk object.
| Object | First-year cash, shape | What would make it 5× |
|---|---|---|
| One-workflow small-model inference | Tens of lakhs → ~₹2 cr | A frontier API with no cap; a 8-GPU future-training add-on |
| Multi-workflow departmental node | Low-single-digit crores | Unscoped SI; full MIS rewrite billed to AI |
| Training-capable mini-cluster | High-single-digit → tens of crores | Facility build, not only the cards |
| Burst hours instead of own training | Lakhs, if capped | Uncapped on-demand through a gazetted week you missed |
How to build your own three-point set
Point one: explode your vendor quotes with the hidden-cost checklist. That is the offer, not the benchmark. Point two: pull two public BoQs — a GeM GPU host, an IndiaAI SKU, a published SDC tender if one exists this year. Date them. Point three: your own last adjacent ICT project (a MIS module, a DC expansion). It tells you what your SI market actually charges you, which a national AI blog will not.
Put the three on one page. If your offer sits inside the fence, proceed. If it sits 10× above, demand a smaller object. If it sits 10× below, look for missing rows.
Newspapers are not microdata
A headline that a state will spend ₹200 crore on AI is not a benchmark for your grievance cell. That number usually mixes a skill mission, a park, a challenge prize and a wish. If you cannot explode it into objects, you cannot use it. Attach the tender BoQ or do not attach the headline.
The same is true of vendor case studies that say a PSU saved 40 percent. Saved against what baseline, on which object, with which utilisation? If those answers are missing, the case study is an advertisement. Advertisements can live in an annex marked colour. They cannot live in Ledger A.
Objections you will hear — and what to do with them
These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.
Just give us the numbers from your customers
Our customers' contracts are not a public dataset. Where we can show a method on your file, we will. We will not anonymise three quotes into a fake national mean.
NASSCOM / a consultancy has a figure
Read whether it is a survey of marketing budgets, a global number, or a press note. If the microdata is not public, it is colour, not a benchmark.
Without a benchmark we cannot sanction
You can sanction against a complete Ledger A and two dated public anchors. Committees did this for servers before AI had a brand.
Order-of-magnitude is too vague for finance
Finance prefers a labelled band to a fake precise mean. Offer both the band and your exploded quote.
A ten-day benchmark page
- Days 1–3: explode your quotes. Classify the object in one sentence.
- Days 4–7: collect two dated public anchors and one internal adjacent project.
- Days 8–10: draw the fence. Write why you are inside, above or below. Attach the screenshots. Refuse any typical Indian deployment sentence that has no microdata.
How this shows up in the file
Subject: Cost anchors for [object] — method, not a census.
No anonymised national dataset has been used. Anchors: [GeM extract date], [IndiaAI extract date], [internal project]. Offer exploded in Annex A. Position versus order-of-magnitude fence: [inside / above / below] because [reason]. Recommendation: [proceed / shrink object / demand missing rows].
This study is not a price list and not legal advice. There is no honest public census of Indian government AI deployment costs in this file.
This article is informational field guidance for Indian public institutions, not legal, procurement, tax, accounting, tariff or engineering advice. Confirm against the current Gazette, GFR, GeM term, SERC tariff order, IndiaAI portal rule, CAG mandate, DPDP text, departmental finance manual and your counsel before you file it. Figures are methods and order-of-magnitude illustrations, not a dataset of real deployments and not a substitute for a live quote.
How to put this in the finance note
A P1 CIO/CTO searching “AI deployment cost benchmark India” needs a number a CFO can defend, not a GPU brand. “Cost Benchmarks From Real Indian Deployments” belongs in a cost model with people, power, idle time, AMC and the cost of a failed pilot.
We will not invent a dataset of real Indian deployments and average it. We will give you a method, public anchors, and order-of-magnitude bands labelled as such. IndiaAI subsidy, if you use it, is a live notice — not a permanent discount. On-prem TCO includes ops headcount. Do not invent Rs/hour. Cite the source of every rupee.
- Separate capex, opex, and one-time cleanup.
- Show utilisation, not just peak GPUs.
- Price the human fallback, not only inference.
- Date every tariff and subsidy assumption.
Close this loop before the next CAB
Put “Cost Benchmarks From Real Indian Deployments” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”
Revisit the item when the model, the GeM term, the region, or the SI changes. “AI deployment cost benchmark India” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.
Questions this usually raises
- What does an Indian government AI deployment typically cost?
- There is no typical we will stand behind. A one-workflow small-model agent and a training hall are different objects. Use order-of-magnitude fences and dated anchors.
- Why will Prcept not publish a customer cost table?
- Because it would be a selected, contractual, non-random set that readers would treat as a census. That would be a pretty lie.
- Are IndiaAI hours a benchmark for on-prem?
- They are a benchmark for burst hours. They are not a TCO for a standing air-gapped duty. Convert with utilisation and power, or do not convert.
- Can we use US cloud list prices times a rupee rate?
- As a sanity fence for tokens, maybe. As a sanction figure, no. You will buy on GeM, IndiaAI or a rate contract, in rupees, with GST and support.
- What will you put in our steering pack?
- The method, the fence, your exploded quote, and a shrink option. Not a fake peer average.
- Is this a price list?
- No. It is a method. Dated public anchors beat an anonymised national mean we do not have.