India’s Sovereign AI Race will be Decided by Power and Access, not Chips

On Thursday, the Centre for Development of Advanced Computing opens a seven-day window for bids to supply the IndiaAI Mission with up to 10,000 sovereign AI compute units. It is one of the largest public compute procurements anywhere this year, and the terms are exacting: clusters of at least 512 units in a single Indian data centre, bidders with an average annual turnover of at least 120 crore rupees, and the control plane, data, logs, backups and disaster recovery kept in India. Companies from countries sharing a land border with India are excluded.

The hardware nationalism is real, and it is now enforceable contract language. Separately, the Economic Times reports that New Delhi is drawing a sovereign AI map, identifying critical sectors and data categories for infrastructure that would be on-premises, entirely government-owned and air-gapped, with state-owned hardware run by private operators in less sensitive tiers.

But a tender is not a strategy. Sovereignty in AI will be decided less by whose chips sit in whose racks than by three operating questions India’s bid documents leave open.

The first is power. Minister of State for Electronics and IT Jitin Prasada said on October 14 that India is rolling out a data-centre policy so that consumers do not end up paying more for electricity, a concern Reuters linked to resident protests near a planned Amazon site. He gave no scope or date, and the official release for the conference where he spoke does not mention the policy at all. Yet the empanelled providers will price compute services for five-year terms, absorbing whatever tariff, siting and grid-cost rules arrive later. The tender’s scoring tells the same story in miniature: installed capacity earns 30 of 100 marks, while renewable energy and liquid or immersion cooling earn five each.

Power is treated as a bonus. It is in fact the binding constraint. A compute unit that cannot be powered affordably is not sovereign capacity; it is inventory. Every country that has scaled data centres quickly, from Ireland to northern Virginia, has discovered this in its electricity bills before it discovered it in its strategy documents. as India’s neighbours concede freely. Sri Lanka‘s digital economy adviser puts national data-centre capacity at about 5 megawatts against a short-term need of 25 to 50, with energy pricing unresolved. Bangladesh’s government has pledged a sovereign AI capability within a year. Only India has written sovereignty into enforceable bid terms, which means India’s rules will become the region’s template, for better or worse.

The second question is access. Public compute is public money. The IndiaAI Mission’s promise is that a student in Guwahati or a startup in Coimbatore will train models on capacity their own market would never finance. The tender itself is silent on who gets in. Empanelment, the document states plainly, confers no guarantee of work allocation, and the second stage, where scope, service levels and prices are set, will run through the Government e-Marketplace later. If access rules are written after the contracts, the incumbents who clear the 120-crore turnover bar will write them in practice.

Good access design is not mysterious. It means published rates for public users, a reserved allocation for universities and research institutions, a transparent queue, and reporting on who actually used the capacity. None of these requires new law. All of them are easier to write into a mission document now than into a crowded market later.

The third is price discipline in a rush. The same tender asks bidders still building their clusters for a guarantee of 10 crore rupees in one clause and 30 crore in another, with no corrigendum published as of October 19. A threefold gap in a security instrument, days before bids open, is what haste looks like on paper.

None of this argues against sovereign compute. The case for it is strong: a country of India’s scale should not have to ask a foreign cloud for permission to build its own AI capacity, and the IndiaAI Mission’s demand estimate of 10,000 GPUs, with a first tender of 5,000, is a serious start. The argument is about sequencing. Rules written before pricing are policy. Rules written after are retrofitting.

Three fixes, all available before the window closes on October 29. Publish the data-centre policy, or at least its power-pricing principles, before stage-two contracts are priced. Write access rules for public compute into the mission now, naming the universities, research labs and small firms the capacity is meant to serve, and the rates they will pay. And make power and cooling core scoring criteria in future rounds, not bonus points, so that providers compete on the cost that actually decides whether sovereign compute is affordable.

Sovereignty in AI is a promise about who benefits. India is the only country in South Asia in a position to define it as more than geography: as compute a researcher can actually reach, powered without asking households to pay more. The bid window is open for seven days. The rules can still be written first.

Jackie Omotalade

Jackie Omotalade is a Chief Programs & Impact Officer at a global nonprofit, where she works on artificial intelligence and climate programs. She advises executives and boards on climate finance, responsible supply chains and technology policy, and has moderated a main-stage panel at Greentech Festival Berlin.