The Hidden Cost of Air-Gapped: Ops Headcount

Most air-gap notes in 2026 still open with a GPU line and close with a sovereignty paragraph. The middle is empty. That empty middle is where programmes die: not because the model was weak, but because nobody was rostered to move a patch across a USB ceremony on a Saturday, or to explain a wrong retrieval to the secretary on Monday.

Cloud vendors hid this cost inside a subscription. Someone, somewhere, pulled images, rotated licences, watched crash dumps and opened a bridge at 2 a.m. An air gap does not delete that work. It relocates it. If you do not put names against it, the work lands on the one NIC or SDC engineer who already runs mail, VPN and the state portal.

This is a field study of that relocated work. It is written for CIOs and CTOs who have to defend a financial concurrence. It is not a bid to inflate staff. It is a bid to stop pretending that a disconnected agent is cheaper than a connected one because the invoice no longer shows a foreign APM.

What the invoice hides

A connected stack outsources five jobs you only notice when they stop: image pulls, model and tokenizer updates, licence and entitlement checks, crash and usage telemetry, and a follow-the-sun incident desk. Each of those jobs has a person behind it. The person is not on your establishment. That is why the SaaS quote looks lean.

Disconnect the stack and those five jobs become yours. They do not become 'zero'. They become a courier protocol, an artefact library, an offline licence file, a local SIEM view, and an on-call rota that can survive casual leave. Committees that compare only rupees-per-token will pick the disconnected bid and then wonder why the pilot never left the lab.

There is a sixth job the invoice never showed: purpose control. In a SaaS tenant the vendor's acceptable-use policy quietly refuses some workflows. Behind an air gap, your officers can point the agent at any share they can mount. Someone has to refuse that. That someone is staff.

These are roles, not necessarily six new posts. They are six named duties.

Work the cloud hid	What it becomes behind the air gap	Who usually inherits it if you do not plan
Image and chart pulls	Internal registry + signed bundle import	Whoever has sudo on the jump box
Model / tokenizer updates	Artefact library, hash verify, rollback	The data scientist who demoed once
Licence / entitlement	Offline licence file and clock-skew watch	Vendor TAM, until the TAM changes
Crash dumps and traces	Local observability + SIEM export	Nobody, until the first outage
Follow-the-sun incident desk	Department or SI rota with leave cover	The NIC nodal officer's personal phone
Acceptable-use / purpose	Tool catalogue and written refusals	Nobody — the agent grows a new purpose

An honest FTE map — ranges, not theatre

Do not put a fake 2.0 or a fake 0.2 in the note. Headcount depends on how many workflows, how many isolated networks, whether you share an SDC pool, and whether the SI is actually on the bridge. The ranges below are order-of-magnitude planning figures from programmes we have sat in, not a staffing circular. Treat them as a conversation starter for the establishment officer.

Ranges, not entitlements. A campus with one research agent is not a CPGRAMS desk. Write which one you are.

Duty	Sandbox / one workflow	Production citizen or plant workflow	Notes
Application / process owner	0.2 FTE named	0.4–0.7 FTE named	Cannot be 'the cell'
Platform (K8s, GPU, registry)	Shared SDC contact	0.5–1.0 FTE or a paid SI seat	Leave cover mandatory
Artefact / model librarian	Collateral of platform	0.2–0.5 FTE	Hashes, licences, rollback
Offline observability + SIEM	Borrow SOC one day a week	Standing SIEM mapping	CERT-In 180-day logs still apply
Patch courier / change manager	Ad hoc	Standing CAB slot	Two-person integrity for media
On-call (after hours)	Best effort	Rota of at least three names	One name is not a rota

Two anti-patterns show up in every review. The first is the hero: one competent engineer who can do all six duties on a good week. Heroes take earned leave. Heroes get promoted. Heroes become vendors. The second is the committee: a steering group of twelve that can approve a purchase and cannot restart a node. You need named operators, not more chairs.

Shared pools beat lonely posts

A district cannot hire a GPU SRE. A state can fund a small air-gap cell inside the SDC that several departments share: registry, artefact library, patch ceremony, GPU scheduling. The department still names an application owner. The cell names a platform owner. That split is how you keep headcount honest without stranding a tehsil.

Universities have a similar split: the computer-centre or HPC cell owns the rack; the registrar or controller of examinations owns the workflow. When those two names are the same person, the syndicate will one day discover that admissions week and a CUDA driver upgrade were planned for the same Saturday.

Cost is not only salary

Establishment cost is the obvious line. The hidden lines are training, dual control, and time-to-repair. An air-gapped change that needs two officers, a sealed bag, a hash printout and a CAB minute is slower than a cloud pipeline. Slowness is a cost. So is the overtime that appears when a festival week coincides with a CVE.

• Training: two days hands-on for every name on the rota, repeated when the SI changes. A seminar is not training.
• Dual control: media import and key ceremonies need two people. Budget the second person or you will skip the ceremony.
• Idle capacity: a second node or a documented manual path so leave does not equal darkness.
• Vendor access: if support must enter, someone from your side sits with them. That hour is yours, not theirs.
• Eval and regression: every model bump needs a person to run the gold set offline. That is not 'the data scientist if free'.

CERT-In still wants specified ICT logs retained in India for 180 days. An air gap does not waive that. It means your SOC, not a foreign APM, holds the traces. If the SOC says they have no parser for the agent's logs, you have just found another half-FTE of integration work. Put it in the BoQ or it will land as a surprise after go-live.

Four places the maths changes

State data centre, many tenants

The SDC already has shift staff, a change calendar and a visitors' protocol. That is an asset. It is not a blank cheque. GPU nodes change power, cooling and the blast radius of a bad image. If the SDC absorbs the platform duty, the tenant department must still fund the application owner and the eval owner. Otherwise every tenant will open tickets that say 'the model is wrong' and the SDC will correctly reply that they only run the box.

District office

A collectorate will not get a Kubernetes engineer. Design for a managed appliance or a thin client into the SDC, plus a local officer who can disable the agent and fall back to the existing file movement. If your architecture needs a district sysadmin at 2 a.m., you designed a city system and shipped it to a tehsil.

PSU plant or OT-adjacent network

OT engineers are already scarce. Do not make them the LLM on-call. Keep the agent on a segmented IT VLAN with its own rota, and treat any diode or historian hop as a change that OT signs. Shared heroes across DCS and RAG are how a retrieval bug becomes a plant incident in the telling, even when it was not one in the physics.

Campus

Faculty will volunteer for the demo and vanish in exam week. The computer centre will refuse to own a model they did not choose. Write the split before the syndicate: HPC owns power, images and backup; the user department owns prompts, gold sets and student-data purpose. Student internships are not a rota.

Objections you will hear — and what to do with them

We will use the existing NOC.

Ask the NOC manager to initial the six duties. If they initial them, attach the page to the note and add a leave-cover name. If they will not initial them, you have your answer. A verbal 'we can keep an eye' is how Saturday becomes a war room.

The SI is in the BoQ, so headcount is their problem.

The SI can take platform and courier duties. They cannot be the fiduciary, they cannot sit in your CAB forever, and they will not be there on day 1,091. Write which duties revert on exit. If the answer is 'all of them' and you have no names, you have bought a subscription with extra dust.

Automation will handle it after month three.

Put the automation in the plan as a reduction, not as a premise. Month three is when the first real CVE arrives and the first officer asks for a new tool. Scripts that do not exist yet are not a reason to write zero in the concurrence.

Other departments run chatbots with no extra staff.

Ask whether those chatbots are on a public model with unofficial paste, or on a production record with a trace in the SIEM. Unofficial paste has a headcount of zero and a DPDP problem of one hundred. Do not use it as a benchmark.

A thirty-day staffing playbook

1. Days 1–5: list the six duties against this workflow. Mark each as department, SDC pool, SI, or vacant.
2. Days 6–10: name humans, not cells. Add a leave-cover for every production name. If a name is vacant, either redesign the architecture or open a post / SI seat.
3. Days 11–15: walk the NOC and SOC through a tabletop: image import, model rollback, SIEM silence, SI resignation. Minute who actually typed.
4. Days 16–22: price training, dual-control hours, and on-call. Put those lines next to the GPU line in the draft concurrence.
5. Days 23–30: get initials. No initial, no go-live on personal data. Sandbox may proceed on synthetic data while names are found.

How this shows up in the file

The financial concurrence should carry a one-page duty map: six rows, named humans, cover, and who pays. The architecture PDF can stay in the annexure. The duty map belongs on page one, next to power and purpose. A later CAG or internal-audit query will not ask which model you used. It will ask who was supposed to be there.

If you cannot fill the map, do not air-gap yet. Run a tightly proxied on-prem pilot with a vendor desk you still control, and write the date when the desk must disappear. An honest hybrid with names beats a patriotic diagram without them.

How to cost a year without fake precision

Finance will ask for a rupee figure. Give them a range built from named duties, not a made-up fully-loaded FTE published as if it were a circular. Take each named human, take the fraction of their year this workflow will eat, add training days, add dual-control hours, add on-call. Leave GPU and power on their own lines. The point of the exercise is that the people line is not zero.

A shared SDC cell serving four departments will look expensive until you divide it. A lonely hero in one department will look cheap until they take earned leave in the week a CVE lands. Show both pictures. Committees pick the cheap picture if you only print one.

Line	What to write	What not to write
Named owner	0.4–0.7 of an existing post or a new one	To be managed internally
Platform / SI seat	Shared cell or paid seat, with cover	Existing NOC will absorb
Training	Two days × names on the rota, twice a year	Knowledge transfer as part of go-live
On-call	Three names or an SI night clause	Best effort
Vendor accompany	Hours for escorted support	Free TAM forever

If your finance department uses a standard fully-loaded cost, use theirs. Do not invent one in a vendor blog and paste it into a sanction. The honesty is in the fractions and the names, not in a national average salary you cannot defend.

What 'managed air-gap' still leaves you

Managed services move platform toil. They do not move fiduciary duty, purpose control, or the right to refuse a new tool. Write the residual list on the day you sign the SI. Residual lists that say 'nothing' are how you discover on exit that nobody in the department can admit a bag.

This article is a field guide, not legal, procurement, electrical or engineering advice. Confirm numbers, duties and designs against the current Gazette, CERT-In directions, your SDC / NIC / campus standards, a site survey and your counsel before you file them.