What a State University Can Afford: Budget Math

The finance committee of a state university was shown two numbers. One was a chatbot SaaS at a price that looked like a seminar grant. The other was an on-prem stack that looked like a building. Nobody put a third number on the table: the people who would run either thing, the SI who would 'temporarily' become the exam desk, and the unofficial phone tools that would continue if the cheap SaaS was not allowed to see mark sheets. The committee picked the small number. By the third quarter they were paying the large number anyway, plus a cloud invoice, plus overtime.

This is a budget-math field piece for Indian state universities and similarly constrained campuses. It is not a survey. It does not invent average spend, GMV, or a percentage of institutions that 'already adopted AI'. Those figures, when they appear, are usually marketing. What follows are order-of-magnitude planning bands and a way to add the hidden lines. Your three quotes will beat this article. Get them.

It is not a commercial offer and not a GeM price list. Hardware, power, and SI rates move. Read current quotes. Treat every rupee figure below as a planning band for conversation with finance, not as a benchmark you can paste into a PAC.

Four bills, not one licence

People. A campus that cannot name 1–3 officers and at least one technical owner will pay an SI to be the institution. That is the most reliable line item in this entire piece.

Posture. Public SaaS, vendor cloud with a contract, on-prem in the existing data centre, or a small isolated room. Posture decides whether student files may enter. It also decides power, AMC and who is paged at 2 a.m.

Integration. ERP, identity, payment, LMS, NAD/DigiLocker if you touch them. This line regularly exceeds the model licence. Anyone who prices only the model is not pricing the work.

The model and the runtime. Sometimes the smallest of the four for a helpdesk. Sometimes the largest if someone decided you need a fine-tune and a rack you cannot cool.

Planning bands — not a survey

These bands are honest about uncertainty. They are not field measurements of a sample. They exist so a finance committee can see orders of magnitude and ask for quotes. If a vendor treats them as 'market prices from Prcept's study', they are misquoting this page.

Cheap public or campus-SaaS chatbots that must not see live student files: often in the low-lakh to mid-lakh rupees per year range for a small seat pack, sometimes less as a loss-leader, sometimes more once usage is metered. The hidden bill is DPDP and shadow use when people paste anyway.

A contracted, India-hosted or on-estate official helpdesk with integration to identity and a prospectus corpus: think in tens of lakhs over a year or two once you count SI days, not just software. Smaller if you already have identity and a clean FAQ. Larger if your ERP has no usable API.

On-prem inference that you own: a single modern inference GPU server can sit in the mid-lakhs to low-crores depending on generation, memory, vendor and whether education pricing exists. A serious multi-GPU rack, plus UPS, cooling and a year of AMC, is how you wander into crore language. Many state campuses do not need that rack for a cited helpdesk. Someone will try to sell it anyway.

People: a dedicated technical owner plus a faculty or exam-cell coordinator is a recurring cost that will dwarf a cheap SaaS if you cost it honestly. If you do not hire them, you will hire them wearing an SI badge at a higher day rate.

Order-of-magnitude conversation starters. Replace every cell with a written quote before you file a budget. Not a survey.

Shape	What you actually get	Bills people forget
Public chatbot, unofficial	Fluency, no file, no control	Leaks, inquiries, unusable as official record — 'free' is false
SaaS helpdesk, published corpus only	Official answers to public FAQs	Metering, staff paste of live files, exit of logs
On-estate agent, read-only ERP status	Authenticated completeness and status	SI days, identity, bindings, AMC, one technical owner
Owned GPU inference	Control of runtime and traces	Power, cooling, spare, patch path, idle hardware if you over-bought
Fine-tune + multi-GPU + SI programme	A project, not a helpdesk	Data rights, evals, people, a two-year calendar

SaaS is not automatically cheaper

It is cheaper when the job is public prospectus text and you will not send personal data. It is not cheaper when the real job is admissions certificates and the SaaS terms train, retain or transfer. Then the cheap line becomes an unofficial paste culture plus a later on-prem project. You pay twice.

It is also not cheaper when metering explodes in admissions week, or when exit of logs is a change request. Read the unit. Per seat, per conversation, per GPU-hour, per SI-day. Ask what admissions week costs at last year's ticket volume times a safety factor. Do not use the quiet-week demo as the annual number.

What a constrained campus should buy first

An official, cited, multilingual helpdesk on the prospectus and live notices, with logs you hold, no training on student files, and no write to ERP. That is the job that reduces shadow AI.

A technical owner on the payroll or a tightly scoped AMC with named hours. Not a visiting enthusiast.

Only then authenticated completeness and status. Only then, much later, any isolated exam-cell logistics. Almost never a fine-tune in year one.

If the state has a data centre or a higher-education network you can honestly use, prefer it to a heroic departmental rack you cannot cool. If that data centre cannot isolate or log, do not pretend. The companion pieces on NIC readiness still apply.

How to present this to finance without theatre

Three options, each with the four bills. A do-nothing option that prices unofficial tools as residual risk, not as zero. A note that no option uses invented satisfaction percentages.

A sentence on DPDP: cheap processors that you cannot erase from are not cheap. A sentence on UGC: there is no mandatory AI SKU. A sentence on NEP: policy is not a budget head.

Ask for a year of people and AMC, not only for a one-time GPU. Capital without revenue is how racks become heaters.

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.

Give us a single rupee figure for 'campus AI'.

There is no honest single figure. There are four bills and a posture. Anyone with one figure is selling one SKU.

State universities cannot afford on-prem.

Many cannot afford a multi-GPU programme. Many can afford an official helpdesk on existing compute or a modest server, plus a person. Do not let the crore-rack story erase the tens-of-lakhs story. Get quotes.

We will use students as cheap labour instead of an SI.

Student projects are fine on public data. They are not a processor for mark sheets and not an AMC. Do not build production custody on a final-year timetable.

DPIIT vendors are cheaper by policy.

DPIIT recognition can affect eligibility and relaxations. It does not set a price. Score the four bills.

A 21-day budget pack

The pack is three options and three quotes. The article is not a quote.

1. Days 1–5: write the job in one page (public FAQ vs authenticated files vs exam isolation). Name the posture you can actually operate.
2. Days 6–12: request quotes that split licence, SI days, hardware, AMC, and what is not included. Ask the metering question at peak volume.
3. Days 13–18: add people and power. Add the do-nothing residual-risk paragraph.
4. Days 19–21: finance note with three options. No GMV. No fake benchmark. No NEP stamp.

How this shows up in the file

Keep the quotes, the four-bill table filled with those quotes, and this article's disclaimer that our bands are not a survey. If a later PAC asks why you did not 'use the industry average', you can show that you refused to invent one.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, academic-regulation or engineering advice. Confirm against the current Gazette, DPDP text and Rules, CERT-In direction, India AI Governance Guidelines, UGC/AICTE/NAAC notices, NEP documents, GFR, departmental manual and your counsel before you file it. Guidelines are not statute. Circulars move.

How this works on an Indian campus

A P3 University should be able to run “What a State University Can Afford: Budget Math” without importing a US playbook. “university AI budget India” hits UGC/AICTE/NAAC clocks, exam secrecy, reservation rules, and students who may be minors.

A state university can afford a cited helpdesk. It cannot afford a hidden SI, a silent GPU, and a SaaS that eats mark sheets. Budget the people and the posture, not the demo. DPDP applies to student personal data. Chatbots are not a strategy. Exam and admissions writes stay with officers. Affiliated colleges need isolation, not one shared index.

• No production student data in a vendor SaaS sandbox.
• Write the academic integrity policy before the tool.
• Consent and purpose tags on student-facing agents.
• Budget for staff training, not only licences.

Close this loop before the next CAB

Put “What a State University Can Afford: Budget Math” 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 P3 University, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “university AI budget India” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.