The Cost of Sovereignty, Quantified

A finance controller in a central PSU asked for the sovereign premium as a single number so the board note would fit on one slide. The CIO, who had just priced two used accelerators, a year of a site reliability engineer, and an evaluation contractor, refused to invent one. The controller invented 30 percent anyway, attributed it to industry, and the note went up. Nobody could later find the industry study.

This article will not give you that study. It does not exist in a form honest enough to put in a government file. What exists are order-of-magnitude ranges for the three bills you will actually pay: people, accelerators, and the operating tax of not being a SaaS customer.

Use these ranges to structure a costing sheet, not to copy a rupee figure into a sanction. Prices move. Utilisation moves. Your union, your SDC power rate and your ability to share a cluster with another bureau will move the total more than a vendor discount on a card.

Three bills, not one premium

Ranges, not quotes. Re-price against current invoices before a sanction.

Bill	Order-of-magnitude, departmental agent	What moves it
People	A small dedicated pod is typically the largest line: think one technical owner, fractions of SRE, security, DPO time, and an evaluator. In many Indian institutions this is mid-to-high seven figures INR per year fully loaded, more if you hire on the open market in a metro.	Whether you already have an SDC team; whether you try to fine-tune; shift coverage
Accelerators and hosts	A 7B–8B class Q4 serve can live on a single modern card with roughly 8–12 GB usable VRAM, often a low-to-mid seven figure INR capital line including a host, or a rental. A 70B-class serve with concurrency wants multiple heavier cards and jumps an order of magnitude.	Model size, context length, concurrency, whether you already have a box
Operations tax	Power, cooling, internal registry, SIEM retention, staging, and — if air-gapped — media control. Often a noticeable fraction of the hardware bill, sometimes larger than the hardware bill in a weak SDC.	Air-gap vs on-prem; CERT-In retention; how many environments you keep

SaaS looks cheaper because the people are hidden in the subscription and because you are not paying for the failure mode. The failure mode is an undeclared transfer. Price that as residual risk, not as zero.

A worked costing shape, not a fake study

Take a ministry helpdesk that needs a few dozen concurrent officers, retrieval over a scheme corpus, and a 7B-class instruct model. The hardware story is one or two modest accelerators plus a CPU box for the app and the index. The people story is the real line: someone who can keep serving up, someone who can evaluate answers before a change, someone who owns the egress proxy, and counsel time on the DPA.

If the same ministry insists on a 70B-class model because a demo was fluent, the hardware line can jump from a box you could hide in an existing cage to a cluster that needs power, cooling and a waiting list for parts. The quality jump is real on some tasks and invisible on a retrieval-heavy scheme FAQ. Measure the task before you buy the cluster.

Shared capacity changes the shape. Renting time on an Indian cluster can make the 70B experiment honest as a spike. It becomes dishonest if the rental includes a hosted notebook that logs prompts. Cheap tokens with expensive legal residue is not cheap.

Hidden lines that wreck the comparison

Shadow API keys on officer phones never appear in the SaaS column. Put a row for them. Even a rough count of unofficial seats is better than pretending the department currently spends zero.

Evaluation is another hidden line. Someone must write the fifty frozen questions, score the answers after each promotion, and refuse a bad artefact. If that someone is a consultant, price the days. If it is a joint secretary’s weekend, you are stealing from another file and calling it free.

Power and cooling look small until the SDC says the cage cannot take a fourth heavy card without a board for electrical work. Ask facilities before you fall in love with a 70B quote. The electrical sanction can dwarf the card invoice and will not arrive in the same quarter.

Incident labour is the line finance hates and then pays. A single undeclared transfer of a beneficiary file consumes counsel, the DPO, the CISO, and a month of officer time. You cannot put a precise rupee on it. You can put a qualitative high on the hosted-path sheet and make the secretary look at it.

Objections about money

Give us the percentage or we cannot compare. Compare two filled sheets for the same workflow. A percentage without a denominator is how the 30 percent industry figure got into that board note.

Open weights make TCO zero. They make the licence line zero if the licence fits. They do not pay the SRE.

We will use existing SDC staff at zero marginal cost. Existing staff are already late on something else. Write the hours. Zero is a political number.

Air-gap is only for defence, so ignore that tax. If you are not air-gapped, ignore it. If you are, budget media control and a second staging environment or you will pay the tax as outages.

How to present ranges without looking evasive

Finance officers have been trained to distrust ranges. Give them a point estimate built from named assumptions, then a low and a high that change one assumption each. Low: 7B-class, existing SDC staff hours, no air-gap. High: 70B-class, hired SRE, media-controlled estate. The point estimate is the one you recommend. The high is what happens if someone fell in love with the demo.

Cite invoices, not blogs, for the hardware line. This article’s bands will age. A distributor quote dated this month will not. If you cannot get a quote in time, say so, and do not replace it with a confident rupee figure copied from a keynote.

Separate one-time stand-up from annual run. Committees collapse them and then think sovereignty is a capital spike. The annual run — people, power, eval, licences — is what decides whether the cluster is still honest in year three.

A 90-day costing sheet

1. Days 1–15: pick one workflow. Write current SaaS or unofficial API spend, including shadow use.
2. Days 16–40: price a small-model on-prem path with people hours, not only hardware quotes.
3. Days 41–70: price the larger-model path only if the frozen eval set says you need it.
4. Days 71–90: put both sheets and the residual-risk row in front of finance. Decide. Stop quoting a mythical premium.

What goes in the file

• Two costing sheets for the same workflow (SaaS or hosted vs operated).
• The people-hour assumptions, named, so they can be challenged.
• The hardware quotes with dates, not a blog range.
• The trigger that would justify a larger model.
• A sentence that no single sovereign premium was used.

Prcept AI will not hand you a fake 30 percent study. We will help you size a small honest cluster and tell you when a larger one is theatre.

How to defend this in the file

A P1 CIO/CTO will be asked to explain “The Cost of Sovereignty, Quantified” to a secretary who has ten minutes. Do not start with the model. Start with the store, the hop, the clause, or the residual risk. “cost of sovereign AI” is a search phrase. The file needs a decision.

Sovereignty has a price, and most of it is people. Treat GPU list prices as the visible third of a bill that also includes SRE, evaluation and the air-gap tax. DPDP does not define sovereign AI. Transfers can be lawful and still be a bad idea. Sector circulars can be stricter than DPDP. Write which instrument you are using.

If you cannot name the Data Fiduciary, the processor, the location of traces, and the erasure method, you are not ready for production personal data — whatever the architecture PDF says.

• One sentence on lawful basis or the procurement rule you are invoking.
• One sentence on where prompts, embeddings and logs live.
• One sentence on who can compel the operator.
• One artefact: packet capture, DPA schedule, or deletion certificate template.

Close this loop before the next CAB

Put “The Cost of Sovereignty, Quantified” 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. “cost of sovereign AI” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.