India opens 10,000-unit AI compute tender with 512-unit clusters

What the tender changes

The tender lets providers offer capacity in clusters of 4,000, 1,000 or 512 units, with each cluster at a single data centre in India, Lapaas Voice reported. The eligible hardware list covers GPUs, tensor processing units and other AI accelerators, not GPUs alone. The compute pillar is run by C-DAC under MeitY.

The lower threshold is meant to widen participation among infrastructure providers. It does not remove other financial, technical and operational eligibility requirements, so smaller companies may still face significant barriers to entry. Bidders must show average annual turnover of at least Rs 120 crore across FY2023-24, FY2024-25 and FY2025-26, and positive net worth as of 31 March 2026, according to the tender details.

MeitY has adjusted the terms to secure uptake for at least five years of GPU usage, a second official told ET Datacenters. Bidders in the latest round may also sell capacity globally, a concession the official said was accepted to keep them profitable until Indian demand picks up.

The pool India already has

More than 38,000 GPUs are already available through the shared facility at Rs 65 per hour, against an original first target of 10,000. The IndiaAI Mission was approved in March 2024 with an outlay of Rs 10,371.92 crore over five years. In an August 2026 parliamentary update, MeitY said 15 compute service providers had been empanelled across four rounds, and reported 237 projects approved for subsidised support with 93.18 lakh GPU hours sanctioned. ET Datacenters put the number of companies that have bid to onboard GPUs at 14.

The government says the shared model avoids duplicative capacity. It is also meant to improve utilisation of expensive hardware.

Why access, not price, is the bottleneck

Access, not price, is the constraint, according to ET Datacenters. Rahul Agarwalla, managing partner at SenseAI Ventures, said early-stage startups often receive cloud credits, so affordability is less of an issue in their first year. Juspay cofounder and chief operating officer Sheetal Lalwani said GPU allocation commitments take three to six months, even though the Bengaluru payments firm can front-load investments. Deliveries of GPU cards can run three to four months, and growX ventures noted that such a delay can cost an engineering team a full quarter.

Memory is a second constraint, ET Datacenters reported. Only SK hynix, Samsung and Micron produce high-bandwidth memory at scale.

Micron committed its entire 2026 HBM supply by December 2025, and SK hynix had secured demand for all its 2026 memory chips by October 2025, ET Datacenters reported. Some buyers need more than a shared pool. Robotics firm CynLr and defence technology firm Armory require dedicated or local compute for security or real-time operational reasons.

Lalwani warned that larger Indian players could monopolise data centre capacity and power, a risk that remains even if chip procurement improves.

The hardware gap behind the subsidy

The pool runs on imported silicon, since India does not yet make these AI chips at scale. MeitY secretary S.Krishnan told ETGovernment in a 4 November interview that India targets an indigenous GPU chip in about 24 months under India Semiconductor Mission 2.0. He named no manufacturer, chip supplier or builder for the project. Krishnan described five layers needed to support AI development: energy, accelerator chips, computing infrastructure and servers, AI models and applications.

What happens next

A chip count is an input, not a result, Wion noted. The questions that follow are whether the 10,000 units are procured and deployed, and how much of the capacity is genuinely used by the startups it is meant for.