India wants AI sovereignty, but its data-centre boom risks socialising resource costs while foreign cloud firms retain the most profitable layer.
On August 6, I opened a parliamentary reply from Delhi. AI data centers in India do not, by themselves, require separate environmental clearance under the 2006 EIA notification. Large projects still face clearance once construction thresholds are crossed. From Karachi, the wording caught my attention because the physical appetite of AI has stopped looking like a software question.
New York had moved first, weeks earlier. Texas followed on August 3 by ordering an audit before projects could advance through the grid connection process. India was opening a regulatory door while two American jurisdictions were slowing the queue to examine who would carry the infrastructure cost. I started reading a subsidy ledger instead.
AI Data Centers in India and the Regulatory Gap
The ministry’s answer needs careful reading. A data centre does not automatically enter the EIA process merely because servers consume large amounts of power or cooling resources. Water use remains governed by existing laws and groundwater rules. The legal test therefore starts with project classification rather than the computational load inside the building.
India’s scale is changing quickly. MeitY says installed data-centre capacity rose from 375 MW in 2020 to about 1,575 MW by August 2026. The Central Electricity Authority now estimates data-centre power demand at about 17 GW by 2031-32. A sector that still looks small beside India’s national generation fleet can become locally dominant when several gigawatt-scale campuses cluster around the same transmission corridors.
| Metric | India | Regulatory comparison |
|---|---|---|
| Environmental trigger | No standalone EIA solely for being an AI data centre | New York ordered a one-year hyperscale moratorium |
| Grid exposure | CEA estimates about 17 GW demand by 2031-32 | Texas requires an audit before projects advance |
| Fiscal treatment | Eligible foreign cloud providers can receive a tax holiday through 2047 | New York plans to revisit large data-centre tax breaks |
| Current scale | About 1,575 MW installed capacity | Texas is screening a very large grid-connection queue |
The comparison proves nothing by itself. American states still want AI investment, and India has good reasons to expand domestic compute. The gap lies elsewhere: mature host jurisdictions are starting to price local externalities before granting the next connection. Delhi is still framing speed of build-out as a strategic advantage.
The Subsidy Ledger Matters More Than the Slogan
Google’s Visakhapatnam AI hub shows why the ledger matters. Google and Adani announced an investment of about $15 billion over 2026-2030, including gigawatt-scale data-centre operations plus new network infrastructure. Andhra Pradesh later raised the linked land allotment to about 601 acres. State incentives approved for the project carry an overall cap of roughly ₹22,002 crore.
The Union government has added another powerful attraction. Budget 2026-27 offers a tax holiday through 2047 to eligible foreign companies that serve global cloud customers through data centres in India. Indian customers must buy through an Indian reseller. The structure tells me that New Delhi wants India to become an export base for compute, not merely a domestic hosting market.
Subsidies are not automatically giveaways. A government can accept lower near-term tax revenue if new infrastructure raises productivity and creates a deeper supplier base. Google’s plan also includes subsea connectivity and clean-energy investment, which could strengthen Andhra Pradesh beyond one campus. The real issue is retained economic rent.
I would not dismiss the employment case either. Reuters reports that the wider Visakhapatnam project is expected to create up to 188,000 jobs, though that headline figure does not reveal how many positions will remain in direct data-centre operations. Construction and surrounding services can create substantial work even when server halls themselves employ relatively small technical teams. The jobs argument therefore needs accounting, not ridicule.
Karachi Teaches Me to Read a Gigawatt Differently
Karachi gives me a useful scale. K-Electric reported that the city’s highest demand reached 3,563 MW in June 2025, with peak supply at 3,545 MW. A one-gigawatt campus equals roughly 28 percent of that peak demand, even before I ask where its electricity comes from. Numbers change character when I compare them with a city I know.
Power scarcity in South Asia rarely stays inside an engineering spreadsheet. A new load can require transmission investment, generation contracts and tariff choices that eventually reach households through the political system. My banking instinct is to follow settlement, not announcement. A $15 billion investment figure tells me the size of the commitment, not the domestic value retained after power subsidies and imported hardware.
Water sharpens the same question. Reuters reported that Visakhapatnam receives about 410 million litres a day against a requirement of 480 million, while activists have challenged the Google project over water concerns. Andhra Pradesh says residential or rural supplies will not feed the data centres, and Google says it plans advanced air cooling to protect local water resources. I cannot treat either the protest claim or the corporate assurance as the final ledger entry.
Who Owns the High-Value Layer?
Digital colonialism becomes a useful test only when defined precisely. Foreign ownership alone does not qualify, because India gains real infrastructure when cables arrive and grid capacity expands. The colonial pattern would emerge if India repeatedly subsidised scarce physical inputs while foreign firms retained the most profitable intellectual property and cloud margins. Ownership of the high-value layer matters more than the nationality painted on the gate.
Global cloud concentration makes the question concrete. Q2 2026 market data estimates that Amazon held 28 percent of cloud infrastructure spending, while Microsoft held 20 percent. Google adds 15 percent. Together, the three leaders control 63 percent; India can host their compute without capturing a similar share of platform economics.
India is not standing still. The IndiaAI Mission had access to about 38,000 GPUs by February 2026, and the government announced another 20,000 to expand national compute capacity. The mission carries an outlay above ₹10,300 crore and also backs indigenous foundation-model development. Those investments make the digital-colony argument less comfortable, which is exactly why the argument deserves precision.
My test therefore stays narrow. A data centre strengthens Indian sovereignty when local firms can use the compute to build valuable models and businesses on top of it. A campus looks more extractive when public incentives mainly reduce the cost base of foreign cloud exports while domestic capability remains dependent on the same platforms. The difference will appear in balance sheets long before it appears in patriotic slogans.
America Is Repricing the Local Cost
New York’s July 14 order did not declare war on artificial intelligence. Governor Kathy Hochul imposed a one-year moratorium on new hyperscale data centres while the state develops rules meant to protect ratepayers from transmission costs. Her administration also wants a community investment framework tied to future projects. The political message is blunt: a private compute boom cannot quietly become a public utility bill.
Texas makes the contrast harder to ignore. On August 3, Governor Greg Abbott ordered the Public Utility Commission and ERCOT to audit data-centre projects before any could move forward in the interconnection process. ERCOT was handling more than 474 GW of connection requests of all kinds, more than five times the state’s record peak demand, when the order arrived. Abbott’s later guidance says data centres should pay for their own electric infrastructure and conserve water.
The American turn does not make India foolish for building. Richer economies can afford tighter filters, while India still needs capital and computing capacity. Yet the sequence should trouble any political economist: jurisdictions that accumulated data centres first are now writing rules around grid cost and local consent. India is offering tax relief deep into the future.
The Bill Has Not Yet Arrived
I keep returning to Visakhapatnam because the argument remains unresolved there in physical form. Earthmovers are cutting into red soil while a legal challenge asks whether the project received adequate scrutiny around water and nearby wildlife. The Andhra Pradesh High Court is scheduled to hear the public-interest case again on August 24. Google says the development complies with Indian law.
Perhaps the project will deliver the infrastructure gains its supporters expect. Perhaps new clean-energy capacity will prevent the feared burden on residents, and the wider AI ecosystem will justify the fiscal concessions. My concern begins before either forecast wins. India has already decided that attracting the compute is strategically valuable, while the method for pricing its local resource cost remains much less settled.
Banking taught me to distrust a transaction whose settlement path I cannot see. The AI boom presents a similar problem at national scale: capital enters, incentives move and electricity gets committed under long contracts. The server racks will show up on schedule. Who captures the economic rent is still an open account.
AI Transparency Statement: “This analysis was drafted under editorial direction with AI technical assistance, then verified and edited by Munaeem Jamal.”