State Data Centre Readiness Assessment

Vendor meetings have a ritual sentence: the State Data Centre cannot do AI. Sometimes the next sentence is a public-cloud quote. Sometimes it is a request to put a GPU under a secretary's conference table. Both sentences insult people who already run treasuries, land portals and scholarship systems on that floor. They also dodge the real checklist. GPU, power, ops, egress. If those four are ugly, say so. If they are fixable, fix them. If they are fine, stop performing contempt.

This is a BOFU checklist for a state CIO who has to choose a home for an agent in the next budget cycle. It is dated 17 August 2026. It is not an official NIC or MeitY audit instrument. It is the walk we do with SDC managers who are tired of being a slide villain.

Score the floor, not the brand

An SDC is a building, a VLAN design, a change calendar, an identity tie, and a group of people. 'NIC cloud' and 'state cloud' are related but not identical objects. Ask which floor, which cage, which hypervisor, which backup site, which on-call. Geography of the city is not the score. A well-run SDC with no GPU is a better parent for a small clerk than a shiny colo whose outbound you cannot see.

Be fair in the minutes. Write what the SDC already does well: physical security, visitor logs, often a working SIEM, often a culture of change tickets. Then write the AI-specific gaps without sarcasm. People who have been mocked in the last steering committee will not open a rack for you.

Readiness rows to walk with the SDC manager. Fill the third column on site. Do not invent numbers in this article.

Row	What good looks like	Fail if
Power and cooling headroom	Documented spare kW and cooling for the proposed enclosure, with summer derate	The only answer is 'we will see' or a diesel story without a number
GPU or equivalent capacity	Named SKU, count, tenancy, and who may not share the box with an open website	A promise of GPUs 'when the IndiaAI allotment arrives' as the only plan
Ops people	A named on-call who can restart the runtime, plus a vendor window that is not 3am WhatsApp	The only operator is a project intern
Identity	Official IDs, no shared root, break-glass recorded	A vendor standing password on a sticker
Egress	Deny-outbound VLAN, packet capture, licence and crash-reporter pinned or removed	'The container needs GitHub' as a production fact
Logs	Department-owned export, CERT-In 180-day floor, Indian jurisdiction	Logs only in a vendor console
Change calendar	You can book a window and a rollback	Silent vendor patches on a Friday night

GPU, power, ops, egress — the four arguments that are fair

GPU density in many SDCs is low because the original programme was not a training cluster. That is history, not incompetence. If you need inference, say how much, at what batch size, for which model class. A scheme-eligibility clerk is not an LLM pretrain. Do not requisition a research cluster to extract income certificates.

Power and cooling are why honest SDC managers say no. Believe a no that comes with a number. A yes without a summer derate is how you trip a breaker in May and take down someone else's treasury batch.

Ops is the scarce resource. A model that needs weekly weight swaps and a CUDA dance is a staffing request. Prefer a runtime the existing Linux team can restart from a runbook of two pages. If your vendor cannot write that runbook, they are selling you a laboratory.

Egress is where modern AI software is rude. Licence pings, tokenisers, telemetry, feature flags. On a health or welfare VLAN that rudeness is a transfer. Test it. The working-with-NIC guide in this batch is about how to ask for that test without starting a war.

Two assessments of the same building

Objections you will hear — and what to do with them

IndiaAI compute will arrive and save us

Treat allotments as a possible later parent, not as a present architecture. If you cannot name the queue, the tenancy and the egress rule, you do not have a home. Build the clerk where you can see it.

The SDC will never allow GPUs

Ask them, with the power number, not with a manifesto. Many will allow a dedicated enclosure with the right tenancy. Some will not, for good reasons. A documented no is progress.

Public cloud is cheaper than honesty

Cheaper until the first health or welfare packet lands in a review queue you do not control. Price the egress risk and the exit. Then compare.

We need Kubernetes on day one

You need a restartable runtime and a deny-outbound VLAN. If the SDC already runs VMs well, start there. Orchestration is not readiness.

A one-week assessment playbook

1. Day 1: walk the floor with the SDC manager. Fill the table. No vendors in the aisle.
2. Day 2: packet-capture a representative runtime in a sandbox VLAN. List every outbound name.
3. Day 3: write the ops runbook test — who restarts, in how many minutes, from what document.
4. Day 4: decide CPU-now versus enclosure-next versus a documented NIC national option. Write the temporary-public-API ban.
5. Day 5: file the note. Invite NIC / SDC to correct any number before it goes to the secretary.

File note you can paste

Subject: Readiness of the State Data Centre to host an agentic workload — assessment, not a smear.

This department records that NIC and SDC teams already host critical public systems. AI-specific readiness will be scored on power and cooling headroom, GPU or equivalent capacity, named ops, official identity, deny-outbound egress, department-owned logs, and a change calendar. Public model APIs will not be used as a temporary home for production personal data. Figures in the annex were agreed with the SDC manager on the date below. This note is not an official MeitY audit.

This note is an internal aid. It is not legal or engineering certification.

Prcept AI expects to be hosted where you can see us. We will not win an argument by insulting NIC. If the SDC is the right parent, we will write the two-page runbook. If it is not, we will say so with numbers, not with a slogan.

Tenancy and the open-website problem

Many SDCs correctly refuse to put a model runtime on the same hypervisor cluster as a public brochure site. Believe them. Prompt-injection and noisy-neighbour stories are not only cloud folklore. Ask for a dedicated VLAN and, if GPUs arrive, a dedicated enclosure. Write who may not share it. A 'temporary' share with a student-project VM is how a welfare clerk inherits someone else's outbound.

Backup is part of readiness. If the only snapshot lives on the same floor, you have a hall, not a design. Ask where restore was last tested and how long it took. A model store that cannot be restored is a go-live you will explain after a disk dies on a Saturday.

Invite the SDC manager to correct every number in your note before it goes upstairs. A CIO who surprises the host with a cabinet figure will not get the next cage. Fairness is operational, not decorative.

This article is informational field guidance for Indian public institutions, not legal, data-centre-certification or procurement advice. Confirm against SDC / NIC capacity, CERT-In directions, your security policy, and counsel before you file it.

How to sequence this in a state, not a slide

“State Data Centre Readiness Assessment” is a department problem. A P1 CIO/CTO should name the legacy system, the officer who owns the file, and the citizen charter clock before buying “state data centre AI readiness”.

An SDC is often the right home for a state agent. It is not a magic GPU hotel. Score power, cooling, capacity, ops, identity and egress honestly — and be fair to the NIC and SDC people who already keep the rest of government up. Do not invent league tables of states. Read tenders and policies. Election Model Code of Conduct can freeze a rollout. NIC is a partner, not a villain. SDC readiness is GPU, power, ops and egress — not a logo.

• Audit the legacy store first.
• Keep mutation and money as officer actions.
• Map SLAs to the citizen charter.
• Budget change requests after go-live.

Close this loop before the next CAB

Put “State Data Centre Readiness Assessment” 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. “state data centre AI readiness” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.