AI Data Centres, EVs & Air Conditioners: A Mexican Standoff for India's Power Grid

India's next electricity challenge is not producing enough energy—it is delivering reliable electricity when AI, EVs and cooling demand converge.

Central Thesis

India's power sector has traditionally planned for growing annual electricity consumption. The next challenge is different in kind: rapidly growing demand from AI data centres, electric vehicles and air conditioning is increasingly interacting with the system's high-stress hours, creating a severe capacity and flexibility requirement rather than simply an energy-availability problem. The binding constraint is not annual energy; it is coincident peak demand. Without market signals and regulatory reforms that price flexibility and reliability explicitly, the cost of meeting that peak will ultimately be socialised onto consumers who did not cause it.

On 21 May 2026, India's all-India peak demand reached 270.8 GW for the first time. The reflexive response to a number like that is to ask whether the system has enough capacity. That is the wrong question, or at least an incomplete one. The right question is whether the system has enough reliable capacity, in the right place and in the right hour, to serve loads that are becoming steadily less tolerant of interruption.

For thirty years, Indian power planning has been organised around energy sufficiency — building enough generation, in aggregate units, to match a growing economy. That worked when the dominant new loads were agriculture, general industry and residential lighting: loads either diffuse through the day or are already accepted as schedulable. It no longer works. Three loads — air conditioning, EV charging and AI-oriented data centres — are growing rapidly, but their load shapes are different. AC demand is strongly weather- and daytime-coincident; EV charging can concentrate in the evening if unmanaged; and AI data-centre demand is comparatively flat. Individually, each is manageable. Their interaction with the 6–10 PM window, particularly as solar output falls away and EV charging can rise, turns “how much energy do we need this year” into “who gets served, at what price, when there is not enough flexible or firm capacity for everyone at once.” This is not hypothetical—it has already played out in different forms, in California, Texas, Ireland and Singapore, and this piece sets out why India is walking into the same conflict faster than its regulatory architecture is built to handle.

A Standoff, Cast

Sergio Leone's The Good, the Bad and the Ugly ends with three gunslingers standing in a graveyard, each capable of drawing on the other two, none willing to move first, all circling the same buried gold. It is a useful lens for India's evening peak, because three loads are now interacting across the same daily stress window, although they do not necessarily peak simultaneously, each with the potential to increase pressure on the other two — and, as this piece will argue, one of the three is considerably more reasonable than the other two.

Blondie, “the Good” (Clint Eastwood) — the electric vehicle. The Man with No Name is the character in the trio capable of being reasoned with, of holstering his gun when the incentives are right. EV charging behaves the same way: given a price signal, a managed-charging default, or simply a cheaper midday tariff, it can move out of the way without a fight. It is the one load in this standoff actually worth calling good.

Angel Eyes, “the Bad” (Lee Van Cleef) — the AI data centre. Van Cleef's mercenary is a professional: cold, contracted, and indifferent to anyone else's plans once he has taken the job. He does not negotiate and does not stand down. That is the AI load in miniature — a 24x7, 85–95% load-factor draw with no intention of flexing for the grid's convenience, because flexing was never part of the contract.

Tuco, “the Ugly” (Eli Wallach) — the air conditioner. Wallach's Tuco is chaotic, appetite-driven and everywhere at once — unpredictable, prone to sudden violent bursts, impossible to fully discipline, and dangerous mostly because there are so many of him. Millions of individually modest room ACs switching on in near-unison the moment a heatwave hits behave in much the same way: no single unit threatens the grid, but the appetite-driven crowd of them does.

None of the three is likely to disappear from the load mix voluntarily. The only real question—and the subject of the rest of this piece—is whether India's regulators write the rules of the standoff in advance, or find out the hard way, some summer afternoon or evening, which load proves hardest for the system to accommodate.

1. Air Conditioning — India's Tuco: The Biggest, Least-Discussed Load on the System

India's room-AC penetration is low by global standards — roughly 8–10% of households, against 60–90% in China and the US. That is usually read as a growth opportunity for manufacturers; for a grid planner it means India has barely begun adding its single largest new coincident-peak load. Room AC sales grew 20–25% year-on-year in FY25 to roughly 12.5 million units, and modelling published in Applied Energy/ScienceDirect (2025), drawing on Bureau of Energy Efficiency sales data, projects 130–150 million new room ACs added between 2025 and 2035.

Under a business-as-usual efficiency trajectory, that same study estimates room ACs alone could add more than 180 GW to India's peak demand by 2035 — close to 30% of projected national peak. An accelerated Minimum Energy Performance Standards (MEPS) path (ISEER 5.0 by 2027, 7.4 by 2033) could cut that by more than 60 GW and avoid an estimated ₹7.5 trillion in system investment—a policy choice not yet made. India's cooling outlook points to a substantial increase in space-cooling demand through 2037–38, reaching as much as 44% of peak load by 2050. Cooling demand is highly weather-sensitive and can become strongly coincident across large regions during heatwaves. Conventional room AC systems also offer limited direct controllability once switched off, although building thermal mass and smart controls provide some flexibility. The key system issue is that cooling demand can coincide with solar-hour peaks during severe heat, while also contributing to evening stress in some regions.

2. Electric Vehicles — India's Blondie: A Manageable Load, If Managed

India registered roughly 2.55–2.6 million EVs in FY2025–26 (IESA data), up about 26% year-on-year, with two- and three-wheelers dominating the installed base. NITI Aayog's stated ambition — 30% of private cars, 70% of commercial vehicles, 40% of buses and 80% of two- and three-wheelers electrified by 2030 — implies a cumulative fleet approaching 80 million EVs by decade-end.

The electricity math is not alarming by itself: an EV fleet of the order of 80 million vehicles has been cited in Indian policy and industry projections, but its peak impact depends heavily on vehicle mix, battery size, charging power and charging behaviour. The annual energy requirement can be accommodated much more easily than a concentrated charging peak. The risk is behavioural. Left unmanaged, EV charging clusters exactly where AC demand clusters, as people arrive home around 6-9 PM; California studies consistently show that unmanaged EV charging can amplify evening peaks; the specific 20-GW-to-260-GW figure in the earlier draft could not be independently verified and has therefore been removed. EV charging is nonetheless the most manageable of the three loads, because batteries can absorb a nine- to twelve-hour window without inconveniencing the driver — California's ChargeWise dynamic-pricing pilot achieved 98% off-peak charging simply by making the solar-rich midday window more attractive through pricing. India, adding solar and EVs in the same years, has a real chance to build the right habit before the fleet locks in the wrong one.

3. AI and Data Centres — India's Angel Eyes: Small Today, Structurally Different Tomorrow

India's data-centre sector is the newest and fastest-growing of the three loads, and the one where forecasts diverge most — itself a sign of how early-stage this is. Installed IT capacity stood at roughly 1.1–1.5 GW through 2025 (JLL, Savills, CBRE), with strong recent absorption and continued expansion driven by cloud, BFSI and AI demand. 2030 projections range from a conservative ~4.5–6.5 GW (Colliers, Rubix, JLL) to a bullish 8–10 GW (Jefferies), with forecasts varying materially by source and assumptions. Annual consumption is estimated to rise from roughly 10-15 TWh today to 40-45 TWh by 2030, concentrated in Mumbai, Chennai, Hyderabad, Delhi-NCR and Bengaluru.

Three things distinguish this load. Load factor: a hyperscale AI facility runs at 85–95% of nameplate capacity continuously, versus 55-70% for typical industrial load, because idle GPU capacity is capital sitting unused. Reliability expectation: interruptions can impose substantial commercial costs, and hyperscale operators generally require much higher availability than ordinary retail customers; however, the exact reliability requirement varies by facility, workload and contractual architecture. Forecast uncertainty: unlike ACs or EVs, which track demographic and vehicle-fleet data reasonably well, a single hyperscaler's siting decision can add or remove hundreds of megawatts in a single investment-committee decision — exactly the dynamic now playing out in Texas.

4. The Evening Timeline: How the Loads Actually Stack

This is India's version of the “duck curve neck” California has managed for a decade — but India's pattern is more complex because solar withdrawal interacts with weather-driven cooling, potentially coincident EV charging and relatively flat data-centre demand. The data-centre row is deliberately flat: AI load does not fall away in the evening the way commercial load partly does. A high-utilisation data centre can therefore behave more like baseload in its load shape while still creating a peak-planning issue because its demand does not naturally decline during system stress.

5. Estimating the Shift in Relative Contribution (Illustrative)

Triangulated against CEA's National Electricity Plan peak-demand trajectory (277 GW by 2026–27, 366 GW by 2031–32, with a newer draft reportedly targeting ~458 GW by 2032; this figure should be retained only if the underlying draft is cited directly):

No public source reviewed here publishes an official bottom-up sectoral breakdown at this granularity — a gap that CEA/Grid-India should close. The table should therefore be read as an illustrative scenario, not as an observed sectoral allocation. The broader direction is nevertheless important: the two fastest-growing load categories (cooling, AI) have the least native flexibility, while the categories historically used to balance the grid (agriculture and industry) are shrinking as a share — exactly as the system's own balancer, solar, disappears each evening.

6. Why These Loads Don't Fit the Old Planning Categories

This is not merely a generation-capacity problem, even though it is usually described that way. Serving these loads reliably through the evening ramp stresses generation (dispatchable capacity in the right hour), transmission (corridors sized for new load geography), distribution (feeder loading in dense AC/EV pockets), flexibility (ramping as solar falls away), operating reserves (sized for a steeper ramp), ancillary services (frequency response and reactive power as synchronous generation gives way to inverter-based renewables), and resource adequacy (whether firm capacity was contracted years in advance for a peak that no longer follows historical trendlines). A system can have ample annual energy and still fail on any one of these dimensions — which is why “we have enough units” is not a reassuring answer.

7. The Tariff Design Gap

Indian retail tariffs—even where Time-of-Day rates exist—largely recover energy consumed, contracted demand, and connected load. They do not systematically price a consumer's coincident contribution to system peak, their flexibility, their draw on ancillary services, or their contribution to capacity adequacy. A data centre and a textile mill drawing identical MW at 7 PM face similar demand charges in most states, despite their very different reliability burdens — the mill can plausibly shed load in an emergency; the data centre, by design, cannot.

The fix exists, but only in fragments. CERC's Ancillary Services Regulations, 2022 created the first national gateway for demand-side resources as grid-balancing assets. Maharashtra notified a Demand Flexibility Portfolio Obligation in 2024; Karnataka released a draft framework in 2025; Rajasthan notified its Demand Flexibility/Demand Side Management Regulations in June 2026, providing a formal state-level framework for demand flexibility. The regulation itself should be cited for the policy point rather than tying it to a specific solar-capacity figure. But as a recent Florence School of Regulation analysis notes, India still lacks mature, nationwide distribution-level flexibility markets and aggregators capable of systematically bundling small loads into dispatchable resources. State-level frameworks are emerging, but a nationwide market-based demand-response mechanism is not yet operational at scale.

8. Introducing “Competition for Reliability”

Electricity markets built around energy need to become markets that also allocate reliability — firm capacity, fast ramping, reserve margin — as a distinct, priced product, rather than a free attribute bundled into every unit of energy. Today, an AI operator, an EV fleet, an industrial consumer and a household all notionally receive identical reliability from a common tariff pool, regardless of how much each contributes to peak-hour stress or how much it values uninterrupted supply. A hyperscaler that would gladly pay a premium for firm capacity and a household that would accept managed, interruptible service for a lower tariff, are both forced into the same undifferentiated product.

The instruments are not exotic. Dynamic retail tariffs let consumers respond to real-time conditions. Managed EV charging converts a potentially peaky load into a dispatchable one. Capacity markets or capacity remuneration mechanisms pay generation, storage and demand-side resources for verified availability during stress hours, separately from energy sold — separating the grid's insurance function from its commodity function. Demand response turns consumers into grid resources. Explicit reliability products let a consumer choose, and pay for, the firmness it actually needs. Together these let scarcity be allocated by price rather than by outage.

9. Difficult Regulatory Questions That Need Answers, Not Avoidance

Indian regulation today centres on least-cost procurement, renewable purchase obligations, and tariff determination. A regime built for an AI-EV-AC load mix also has to price reliability, flexibility, resilience and demand responsiveness explicitly rather than treat them as free background conditions. That raises questions regulators have mostly been able to defer:

• Should genuinely life-critical loads — hospitals, water/sewage pumping, emergency services — keep unconditional, unpriced priority even as every other class increasingly pays for differentiated reliability, or does the essential-services carve-out itself need formalising for a grid where “priority” is otherwise becoming a purchased attribute?

• Should AI data centres be allowed to buy premium, firm reliability if they pay its full incremental system cost, following emerging approaches in Ireland and Texas?

• Should residential consumers implicitly cross-subsidise the network and capacity costs imposed by AI infrastructure, as happens today under connected-load-based tariffs?

• Should EV charging be interruptible by default, with an opt-out at a price, given how cheaply it can be shifted relative to AC or data-centre load?

• Should flexible consumers receive an explicit rebate reflecting the capacity they help the system avoid building?

• Can reliability be unbundled into a tradable, tiered product at the retail and distribution level, extending the wholesale ancillary-services logic CERC already applies?

10. What Other Systems Are Already Doing

Texas (ERCOT). By August 2026, ERCOT was dealing with roughly 474 GW of large-load connection requests, around 90% of which were attributed to data centres—more than five times the state's record peak demand of about 85.5 GW. Senate Bill 6 (June 2025) imposes disclosure and curtailment obligations on loads above 75 MW, empowers ERCOT to direct curtailment in emergencies, and (under a June 2026 directive) requires data centres to bear the full cost of dedicated infrastructure rather than socialising it. Its long-standing cost-allocation methodology is now under formal review.

Ireland. A four-year de facto moratorium on Dublin-area data-centre connections (2021-2025), prompted by data centres reaching ~22-24% of national metered demand, has given way to a Large Energy User Connection Policy (December 2025): new large loads must supply dispatchable backup matching their own import capacity, source 80% of demand from renewables, and accept non-firm/curtailable status where needed — grow, but bring your own firmness.

California. The CAISO duck curve is the original diagnosis of this problem. The CPUC's move to mandate dynamic retail pricing by 2027, and pilots like ChargeWise (98% off-peak EV charging via price signals alone), show demand-side signals can flatten the ramp; grid-scale batteries have demonstrated discharges above 12 GW during some California evening periods.

Singapore and China. Singapore has used a competitive allocation approach for additional data-centre capacity, with resource-efficiency criteria. China's “Eastern Data, Western Compute” programme also seeks to steer compute capacity toward locations with stronger energy and renewable-resource characteristics. Nordic markets, meanwhile, have passed hourly wholesale prices through to retail consumers for decades — proof that dynamic pricing is administratively workable at scale, rather than merely theoretical.

The common thread across very different instruments — conditional connections, curtailment mandates, dynamic pricing, and central planning — is a shared premise: reliability for large, inflexible, fast-growing loads must be paid for and structurally provided, not assumed as a free extension of the existing grid.

11. Policy Recommendations

1. Make ToD tariffs binding and more granular, extending them to residential and EV-charging categories, rather than limiting them to industrial consumers.

2. Introduce Critical Peak Pricing for a limited number of forecast high-stress evenings a year.

3. Operationalise distribution-level demand-response markets, building on CERC's 2022 Ancillary Services framework and state DFPO models, with a national floor so that DISCOMs are not penalised for cutting peak.

4. Mandate managed EV charging as the default rather than opt-in, for subsidy-linked public and workplace charging infrastructure.

5. Introduce an AI/data-centre Grid Connection Code: large-load disclosure, network-cost contribution, and firm-backup or explicitly non-firm connection status.

6. Create tiered retail/distribution reliability products, extending wholesale ancillary-services logic downward.

7. Adopt probabilistic (loss-of-load-expectation) resource adequacy planning in place of deterministic reserve margins.

8. Introduce wholesale scarcity pricing for a small number of extreme-stress hours a year.

9. Accelerate AC efficiency standards (MEPS) to ISEER 5.0 by 2027 and 7.4 by 2033 — the single most cost-effective peak-avoidance lever identified in this analysis.

10. Establish distribution-level flexibility procurement, with a revenue mechanism that rewards rather than penalises DISCOMs for cutting peak demand.

11. Require CEA/Grid-India to publish a sector-disaggregated coincident-peak forecast (AC, EV, data centre, industry, agriculture) as a standing annex to the NEP.

12. Design an explicit capacity remuneration mechanism, paying resources for verified availability during stress hours, separately from energy sales.

13. Invest in AI-based, feeder-level demand forecasting, since none of the above can be well targeted without materially better short-horizon forecasting than most DISCOMs currently have.

 12. A Simple Numerical Illustration

Take a DISCOM territory with a current peak of 1,000 MW, growing at 6% a year to ~1,340 MW by 2030. If AC's share of that peak rises from 14% to 20%, AC-related peak demand would rise from about 140 MW today to about 268 MW in 2030 — an increase of roughly 128 MW. If unmanaged EV charging adds a further 4% of peak, that represents about 54 MW. A single 50 MW data-centre customer on the same feeder, at a 90% load factor, adds another 45 MW of average demand that remains comparatively flat. The combined incremental requirement is therefore roughly 227 MW, subject to the assumed coincidence and load shapes. through conventional peaking capacity would generally be expected to cost substantially more than a portfolio of accelerated AC efficiency standards, managed charging, and demand-response procurement to defer the same requirement by two to three years — but only if the tariff architecture exists to capture that flexibility—which, today, it largely does not.

Conclusion

India's next electricity challenge is not generating enough power. It is deciding who gets reliable power when everyone wants it at the same time. In the coming decade, electricity markets will compete less over energy and more over reliability. The sooner regulators recognise reliability as an economic product — not merely a technical obligation — the better prepared India will be for the AI age.

Left to itself, the graveyard standoff has only one likely ending: whoever the system operator asks to stand down last, at the worst possible moment, on the worst possible evening. Regulators who write the rules of the draw now — who decides, who pays, who flexes — get to choose that ending in advance. Regulators who don't will simply find out, some July evening, who was really the fastest gun on India's grid.