Empanelled GPU Providers Compared

MeitY, through the IndiaAI Independent Business Division, empanelled cloud agencies to sell GPU hours into a common compute facility. PIB has described that facility in successive notes — more than 34,000 GPUs onboarded by mid-2025, later updates speaking of thirty-eight thousand and a still-moving portal. The list of names on the RFE and on subsequent allocations is public enough to be useful and unstable enough that this article will not pretend to be the live catalogue.

A department that treats empanelment as a quality ranking will buy the wrong thing. Empanelment says a bidder cleared a technical and commercial process to sell into the mission. It does not say the VM is inside your SDC, that your DPDP roles are mapped, that the subsidy applies to your entity class, or that the hour you need next March will be there. Those are comparison columns. Price is one column, and it is the one that goes stale first.

This comparison is a method. Names below are illustrative of classes that have appeared on official and DIC notes — E2E Networks, Yotta, NxtGen, Jio Platforms, CtrlS, CMS and others. Confirm the current list and the current rupee-per-hour on the IndiaAI compute portal before you file a number. We will not invent a 2026 rate card and call it official. prcept.com is not a GPU landlord page. It is the agent we will still run if you rent someone else's hour.

What empanelment is — and what it is not

The August 2024 request for empanelment, described by Digital India Corporation, asked agencies to provide AI compute and cloud services to support a mission that originally spoke of making 10,000 GPUs available, with later procurement rounds expanding the pool. Empanelment was described as valid for 36 months with a possible 12-month extension. That is a contracting fact. It is not a security accreditation of your workload.

Eligibility on the portal has been described, in official FAQ language, as extending to researchers, faculty, fellows, students, startups and MSMEs, with government entities appearing on allocation tables. Your department is not automatically in. Subsidy — publicly described as up to about 40 percent for eligible users on some mission communications — is not a right that attaches to every GeM user. Write the entity class and the approval path before you write the rate.

Empanelment is also not on-prem. An empanelled GPU is a cloud VM in a provider data centre. It can be an Indian data centre. It is still a shared facility with a provider control plane. If your file needs air-gap, this comparison is the wrong article. Read the air-gap TCO notes and stop. A subsidised hour that phones a licence host is still a connected hour.

Columns that survive a price change

Compare providers on facts that will still matter when the portal updates the rupee. Residency of the VM, the disk, the snapshot and the support consoles. Whether the contract can name an Indian legal entity and an Indian dispute seat. Whether you can reserve capacity for an exam window or a budget close. Whether logs can be exported to a store you own. Whether the image can be brought up without a foreign licence ping. Whether the exit includes weights you fine-tuned and the object store you filled.

Then compare operations. Indian-hour support with a named escalation, or a global queue. A written RTO for a dead node. A way to prove that a support engineer in another country cannot open your tenant. These are the columns that decide whether a ₹67-looking hour becomes a six-week outage.

Price still matters. Public reporting around the early mission bids put winning unsubsidised H100-class hours in a broad band that later commentary compressed to roughly the low hundreds of rupees before subsidy, with effective subsidised hours often described in the ₹65–₹100 region for some user classes. Treat every rupee in that sentence as a rumour until the portal shows it against your login. Order-of-magnitude is enough to know that subsidised IndiaAI hours are cheaper than list hyperscaler H100s and more constrained than a card you already own.

Comparison sheet — fill from the live portal and the draft MSA. Do not paste last quarter's tweet.

Column	What good looks like	Fail if
Entity eligibility + subsidy	Your legal name and class are on an approval, with the subsidy percentage written	Sales says government is always 40 percent off
Residency of VM, disk, snapshot, keys	Named Indian DC, keys in your HSM or a documented Indian KMS	Region is India and snapshots replicate for DR to an unnamed geography
Reservation	You can hold a block through a gazetted window	Best-effort queue during board exams
Export / exit	Weights, objects, logs leave in a documented format within N days	Exit is we will try or a proprietary snapshot
Support identity	Named Indian-hour roster; foreign break-glass is logged and exceptional	Follow-the-sun with no jurisdiction map
Meter you can audit	GPU-seconds, idle, storage and egress on a bill an accounts officer can reconstruct	A single platform fee with no grain

Do not turn this into a brand rank

Yotta, E2E, NxtGen, Jio, CtrlS, CMS and the rest of the DIC-listed names are not a league table. A provider that is excellent for a startup training run can be the wrong tenant for a department that must produce CERT-In-resident logs and a CAG packet. Another provider that looks expensive per hour can be cheaper if they will sign a reservation through your budget session.

Do not score Make in India as a synonym for empanelment. Preference under the Public Procurement (Preference to Make in India) Order is a separate legal instrument with its own definitions. IndiaAI empanelment is a mission procurement. If you need both, write both. If you write neither and hope the brand will carry the file, the technical committee will invent a score.

Prcept is not an IndiaAI GPU landlord. We run agents on your rack or in your gap. We will consume an empanelled GPU when that is the honest compute for a burst, and we will tell you when a rented H100 is a worse fit than a smaller card you already own. We will not reprint a portal table and call it our benchmark.

What to ask in the pre-bid week

Send the same sheet to every empanelled name you are actually willing to use. Ask where the hypervisor management plane sits, whether snapshots replicate, whether a support session can be opened from outside India without your ticket, and whether the image first-boots without a licence ping. Ask for a sample invoice with GPU-seconds, idle, storage and egress on separate lines. Ask for the reservation instrument in writing, including pre-emption. Ask for the exit format of weights, objects and logs, and the number of days after termination.

If two providers give the same answers, the live portal rupee can break the tie. If one provider cannot answer without a sales adjective, they are not cheaper at any rupee. Keep a dated screenshot of the portal next to the answers. Empanelment lists move. Your file should show which list you read.

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.

Empanelled means already evaluated, so we can skip the technical committee

Empanelled means evaluated for the mission's cloud service. Your workload, your data class and your reservation still need a committee. GFR does not vanish because a portal exists.

We should pick the lowest rupee on the public price list

L1 on a stale screenshot is how you buy a queue. Score reservation, exit and residency, then price. If two providers are equal on those, the live portal L1 can decide.

If it is on IndiaAI it is sovereign

It is an Indian mission facility operated by private data-centre companies. That is a useful procurement fact. It is not a statutory synonym for sovereignty, DPDP compliance, or air-gap.

We will lock the rate in this article into the DPR

Do not. Attach a dated portal extract and a validity. This article is a method. A DPR that copies a blog's rupee will be wrong by the time it is sanctioned.

A two-week comparison you can attach

1. Week 1: confirm your entity class and subsidy path on the IndiaAI portal. List two or three currently offered SKUs that fit the model class you actually run. Write residency, reservation and exit questions as a one-page sheet.
2. Week 2: send the sheet to those providers, or extract the answers from the live MSA and SLA. Fill the comparison table. Price last, from a dated screenshot. Put the sheet in the file. Re-pull prices the week of financial concurrence.

How this shows up in the file

Subject: Comparison of IndiaAI-empanelled GPU options for [project].

Empanelment is treated as eligibility to quote, not as a quality rank. Providers compared: [names as on the portal on date]. Columns: eligibility/subsidy, residency of VM and keys, reservation through [window], export/exit, support identity, auditable meter. Price extract dated [date] attached. Recommendation: [provider / SKU / reserve vs on-demand] for [reason]. On-prem remains the path for [air-gapped / standing production] work.

This note is not a rate card and not legal advice.

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.