Search interesting materials

Monday, September 01, 2025

Powering AI with Reliable Grids and Networks

by Renuka Sane.

There is much action and anticipation related to the AI decade, and especially about the potential of building large data centres, in India. As of April 2025, at least five hyperscale data centres were in the works. It is expected that India's data centre capacity will surpass 4,500 MW by 2030, backed by $25 billion in investments. Recently, OpenAI has indicated its interest in setting up a data center in India. Before we celebrate these investments, we should ask if the economics of locating them in India adds up? Even leaving aside issues such as land and taxes, do power prices, network quality, and the reliability of both electricity and the internet make us globally competitive? Some investment will come anyway because of data-localisation rules. But that is compulsion, not strategy. Real scale will only occur when a rational firm would choose India even without a localisation mandate, because the numbers and the policy risk both make sense. The AI story has two prerequisites: abundant, reliable electricity and reliable connectivity.

Reliable electricity

Let's start with electricity consumption. A small data center requires about 1-5 MW of power, while a "hyperscale" data centre draws about 100 MW of power at full load. Assuming a power usage effectiveness (PUE) of a data center of 1.2, one hyperscale data center will require 100 MW * 1.2 * 8,760 h = 1.0512 TWh. over a year.

Where is this power going to be sourced from, and how much will it cost? There are three supply options: (1) grid supply,(2) round-the-clock renewable energy plus storage (RTC-RE+storage) contracted from a developer under green open access or (3) a captive plant powering the data centre.

  1. Grid supply: If we assume an industrial tariff of Rs. 7.5 per kWh (which is close to the tariffs in Maharashtra and Tamil Nadu,the two states at the forefront of the data center business), the electricity bill for uninterrupted grid power is roughly Rs. 7.9 billion (US$ 90 million) per year. Except that power supply is not guaranteed, and significant power outages imply that data centres have to build alternatives to ride through grid outages. At Rs. 25-30/kWh, diesel generators routinely cost an order of magnitude more to almost US 236-300 million.

  2. Round-the-clock renewables with storage (RTC-RE + storage) via green open access: Here we have to consider three cost items. The first is the cost of the RTC power purchase agreement (PPA), which will be higher than plain solar/wind because it includes storage and portfolio diversity. The second are the network charges and losses that include intra-state transmission, wheeling (distribution) charges, and transmission + wheeling losses before it reaches the meter. If this is inter-state then one has to start considering the prevailing ISTS regime. The third is the cross-subsidy surcharge (CSS) and additional surcharge (AS) to the DISCOM. In Maharashtra, for example, these are quite high, potentially making the final price above the grid, even if the landed cost of the RTC PPA is significantly lower than the grid price. An approx overall price of Rs.9/kWh hour, gives us a total cost of around US$110 million.

  3. Group-captive RTC: In this case, CSS and AS charges wouldn't apply and the price may come close to (or be lower than) the grid price. The data centre, however, would need to hold the required equity and off-take, and would have its own governance challenges. It is these that can become the binding constraints, not just price. A group-captive cost of Rs.7/kWh leads to a cost of US$83 million.

How do these numbers compare to say the US or Germany? The table below gives some indicative answers.

Country Unit cost (US$) Annual costs (US$)
India (grid average) 0.085 89.4 m
India (RTC, group captive) 0.076-0.080 80-85m
India, (RTC, third-party) 0.104-0.112 109-118 m
Germany, (grid average) 0.196 205.5 m
United Kingdom (grid average) 0.249 261.6 m
US, (grid average) 0.0886 93.1 m
US (Texas) 0.063 66.6m

As the table shows, despite the difficulties in India, it remains competitive vis-a-vis countries such as Germany and the UK as far as electricity prices are considered. However, India is not competitive vis-a-vis the US, where prices in places such as Texas and Virginia (US$0.091) are lower than the third-party open access option in India. The economics of electricity will further change based on the amount of cooling required, which will be relatively higher in India, requiring more energy than in countries with ambient temperatures much lower than India.

Reliable connectivity

There is much more awareness regarding the impediments of the electricity sector for our AI ambition. The issue of reliability of connectivity is less talked about - reliability that gets compromised because of our world-leading record of internet shutdowns. India sees routine network blackouts for reasons related to mobile data bans during exams, to internet suspensions in conflict for months on end. In 2024 India accounted for 28% of all government ordered shutdowns globally - the highest by any country. There have already been 28 shutdowns in 2025. One study estimates the economic cost of shutdowns to the Indian economy at $968 million.

One could argue that data centres used leased lines, and a mobile only shutdown would not matter much. Except that mobile only bans knock out end-user access potentially affecting any product whose customers are on these networks. Shutdowns can affect logistics, field engineering, and remote operations. A shutdown freezes the demand for AI inference in that region. Such vanishing demand equals loss of demand of electricity, leading to idle capacity, another indirect cost for an entrepreneur to handle. One could also argue that shutdowns primarily occur in conflict zones, or districts that are on the periphery of economic activity, and therefore not likely to materially affect the AI story. While that may be true at the moment, routine ad-hoc shutdowns undermine trust in India as a location for latency-critical workloads and cross-border data partnerships. They can lead to a sovereign reliability discount, raising the hurdle rate for capital. Even without shutdowns, India routinely has to deal with connectivity problems owing to frequent power cuts causing internet outages.

Way forward

There are two ways to look at the issue of data centres. The first is whether India can become the regional hub and service clients across Asia and the Middle East. On this question, the answer is clear. Firms will make rational decisions - unless it makes economics sense, firms will prefer to rent equipment or make API calls to the cheapest data centres elsewhere in the world. The second is what it costs firms in India to be forced to place data centres here owing to data localisation mandates. If the cost of training and inference is lower in the US (or other overseas markets) than in India, then firms in India will be at a considerable disadvantage if forced to use local facilities.

Blackouts and shutdowns are not compatible with the way in which AI services evolve and deploy. Foundation models, hyperscale data centres, and exportable AI services demand 24x7 supply and connectivity. India needs to get its grid electricity to world-class levels. It needs to reform its charge structure that makes firm green more expensive than grid making exit difficult. While nuclear energy remains an option, there is no clarity yet on the resolution of supplier-liability laws, making its future still uncertain. Additionally, India needs a radical overhaul of its policy on shutdowns. They should be the absolute last instrument; a rule-of-law state first exhausts narrower tools such as content take-downs, site-specific throttling, geofenced blocks, and targeted law enforcement, and only then even contemplates turning off the network. Our AI strategy should focus on building on two pillars: dependable power and dependable networks.


The author is a researcher at the TrustBridge Rule of Law Foundation. I thank Ajay Shah and Anand Venkatanarayanan for useful comments.