Electricity Cost of On-Prem AI, State by State

Someone will offer you a spreadsheet with twenty-eight state tariffs and a single rupee-per-unit for data centre load. File it in the bin. Retail electricity in India is set by State Electricity Regulatory Commissions against petitions from named DISCOMs. The number that lands on a GPU hall is a stack: energy charge, demand charge, time-of-day, fuel surcharge, electricity duty, tax, sometimes a data-centre-specific rebate under a state policy, sometimes an open-access landed cost that has nothing to do with the DISCOM's HT industrial slab.

This is a method article with example ranges drawn from publicly discussed HT industrial and commercial bands — often a few rupees either side of ₹5–₹9 per kWh for energy charges in many large states, with commercial slabs and some city licensees running higher, and with well-structured renewable open access sometimes landing lower. It is not a 28-state table. A 28-state table that is not footnoted to the current tariff order is fiction with extra rows.

Use the method on your hall. The example arithmetic below is an illustration so a finance officer can see the shape of the bill. Replace every tariff with the order that applies to your connection. Prcept will not sell you a hall. We will refuse to hide power in an on-prem proposal.

The bill is a stack, not a unit rate

Start with the connection you actually have. A NIC room on a state-government HT bulk supply is not a commercial mall LT meter and not a hyperscale park with a dedicated feeder and a policy rebate. Ask the executive engineer for the category name as it appears on the last twelve bills. That name is the row you will look up in the SERC order.

Energy charge (₹/kWh) is the line people quote. Demand charge (₹/kVA/month) is the line that punishes you for sizing a transformer to a GPU peak you hit once a day. Time-of-day loadings can make evening inference more expensive than a night batch. Fuel and power-purchase adjustment charges move quarterly in some states. Electricity duty and tax sit on top. Power factor penalties are how a badly corrected rack becomes a moral story in the accounts section.

If you have open access or a group-captive renewable contract, the landed cost is a different arithmetic — wheeling, banking, deviation, and the residual DISCOM demand charge. Industry notes in 2025–26 have described well-run renewable structures in a broad ₹4–₹7 per unit band against DISCOM HT energy charges that often sit higher. That is a reason to ask your energy department, not a reason to write ₹4.98 into a DPR as if it were your tariff.

Method sheet — fill from the last bill and the current SERC PDF. Ranges are order-of-magnitude, not your rate.

Component	Where to read it	Illustrative shape (not your number)
Energy charge	SERC tariff schedule, your category	Many HT industrial energy charges discussed in public orders sit roughly ₹5–₹9/kWh; some commercial slabs higher
Demand charge	Same schedule, ₹/kVA/month	Often hundreds of rupees per kVA; dominates a peaky GPU hall
ToD / night rebate	ToD annexure	Can swing a batch window more than the base energy charge
FPPAS / PPAC / surcharge	Quarterly DISCOM circular	Treat as a moving adder, not a surprise
Duty and tax	State duty notification + GST treatment	Ask accounts. Do not guess the rate.
DC / IT policy rebate	State data-centre or IT policy + SERC adoption	Exists in some states; not automatic because your rack has a GPU

From nameplate watts to a rupee you can defend

A datacentre GPU's thermal design power is not the hall. Take the GPU TDP, add the CPU, memory, NICs and the server's conversion loss, then multiply by the hall's PUE. A single 700 W-class GPU in a 1,500–2,000 W server in a hall at PUE 1.5 is a very different animal from a marketing slide that multiplies 700 W by the energy charge and stops.

Idle is not zero. A card waiting for an officer prompt still draws a large fraction of loaded power, especially if you never set a lower power cap. The honest annual energy is closer to (weighted average watts × hours the node is powered × PUE / 1000) than to (TDP × 8760 × utilisation). Utilisation of the model is not utilisation of the PSU.

Worked illustration, not a quote. Suppose a two-GPU inference node draws 2.4 kW at the wall, the room PUE is 1.5, and the node is powered 24×7. Annual energy is 2.4 × 1.5 × 24 × 365 ≈ 31,500 kWh. At an illustrative landed stack of ₹8/kWh that is about ₹2.5 lakh a year for electricity on that node alone, before demand charges. At ₹5 it is about ₹1.6 lakh. At ₹12 it is about ₹3.8 lakh. Demand charges can add a line of the same order if the contract demand was sized to a training peak you no longer run. Replace every input.

State by state, without a fake table

The honest state-by-state work is a folder of PDFs, not a row of invented rupees. For each state in which you operate a hall, download the current year tariff order from that SERC — MERC for Maharashtra, KERC for Karnataka, TNERC, UPERC, PSERC, and so on — and highlight the category that matches your connection. Forum of Regulators and CERC pages help you find the door. They do not replace the order.

Public discussion in 2025–26 has repeatedly noted that industrial and commercial tariffs are not uniform: some southern DISCOMs have sat toward the lower-middle of the industrial energy-charge band, some urban licensees and commercial categories toward the higher end, and policy rebates for certified data centres exist in a subset of states. Maharashtra HT industrial energy charges in recent MSEDCL-related discussion have often been described in a mid-single-digit rupee band plus material demand charges; Karnataka industrial retail has been discussed in a similar mid band after recent revisions. Those sentences are orientation. They expire when the next order lands.

If someone hands you a 28-row table with one number per state and no DISCOM, no category, no year and no demand charge, ask them which SERC PDF each cell came from. If they cannot answer, the table is a prop.

What to do with the number once you have it

Put power on the same TCO sheet as AMC, SI and idle GPU, not in a facilities annex nobody reads. A cheap on-prem node in a high-tariff commercial category can lose to a reserved IndiaAI hour for burst training and still win for standing inference that cannot leave the building.

Use night ToD windows for evals and fine-tunes when the order rewards them. Cap idle power. Do not heat a 70B all year for a desk that works 10 to 5. Shared compute across departments only helps if the metering will survive an audit; see the companion piece on shared GPUs.

How to attribute the hall when AI is not the only tenant

Most government halls already cool storage, identity and a MIS. Do not dump the entire DISCOM bill on the GPU. Attribute by metered PDU if you have one. If you do not, attribute by nameplate watts of the AI node over nameplate watts of the hall, then say so. A crude attribution that is written beats a precise attribution that is invented.

If the AI node forced a contract-demand increase, that increment is the node's, not the hall's historic load. Write the old kVA and the new kVA. Demand charge arithmetic is where on-prem TCO quietly doubles. It is also where a night-only training window pays for itself without a new card.

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.

Facilities will handle power; IT should ignore it

Facilities pay the bill. Finance will still ask why the IT node doubled the hall's night load. Put a power annex in the AI file so the argument happens once.

We have a DG and solar, so tariff does not matter

Diesel is not free and solar has a capacity factor. Credit them only if the metering attributes kWh to this node. A rooftop array that already serves the hostel is not a GPU subsidy.

PUE is a hyperscale vanity metric

PUE is how you stop pretending the CRAC does not exist. Even a crude 1.4–1.8 band is better than 1.0. Measure if you can. Bound if you cannot.

Give us the 28-state table anyway

No. A stale invented table is worse than a method. We will sit with your last bill instead.

A ten-day power file

1. Days 1–2: collect twelve months of bills for the hall. Note category, demand, ToD, surcharges, duty.
2. Days 3–5: download the current SERC order. Highlight your category. Write energy, demand and adders on one page.
3. Days 6–8: measure or bound node watts and room PUE. Compute annual kWh two ways: powered-hours and a utilisation fantasy. Keep the powered-hours number.
4. Days 9–10: put the rupee on the TCO sheet next to AMC and idle GPU. If a rebate is claimed, attach the order page that grants it.

How this shows up in the file

Subject: Electricity cost estimate for on-prem AI node [id] at [hall].

Connection category as on bill: [name]. SERC order: [citation, date]. Energy charge, demand charge, ToD and adders copied in Annex A. Node wall watts [x], PUE [y], powered hours [z]. Annual kWh [n]. Landed ₹/kWh stack used for estimate: [figure], marked as estimate. Demand-charge impact shown separately. No 28-state average was used.

This note is not a tariff order and not legal advice. Replace every tariff input before concurrence.

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.