38,000 GPUs, one big problem: India’s AI Mission can’t find enough users

Nearly two and a half years after the Union Cabinet cleared the IndiaAI Mission with a Rs 10,371.92-crore outlay, the government’s flagship bet on “sovereign AI” has hit an uncomfortable milestone: it has built far more compute than anyone is using.

The numbers on paper look impressive. The IndiaAI Compute Portal now lists more than 38,000 GPUs, mostly Nvidia H100 and H200 units,  onboarded through empanelled cloud and data-centre partners, blowing past the original 10,000-chip target set in 2024. Roughly a dozen approved providers, including Jio Platforms, Tata Communications, Yotta, CtrlS, E2E Networks and NxtGen, were meant to rent out this capacity to Indian startups, researchers and academics at subsidised rates as low as Rs 65–92 an hour, a fraction of what hyperscalers like AWS or Azure charge for comparable chips.

But for the compute providers who put up the capital to buy and rack these GPUs, the mission’s own success has become a liability. Two of the original empanelled partners, Jio Platforms and CtrlS, have lagged badly on installation timelines, according to government disclosures, leaving idle capacity on data-centre floors that isn’t earning anyone money. Meanwhile, founders who do try to access the subsidised pool describe a portal that is long on paperwork and short on flexibility,  a seven-day lease cap on fine-tuning workloads, unpredictable renewal cycles, and an eligibility process that a Takshashila Institution report warned would leave much of the capacity “at risk of being underused,” thanks to bureaucratic friction and qualification hurdles that shut out exactly the early-stage startups the mission was designed for.

The disbursement numbers tell their own story. Of the headline Rs 10,000-crore-plus outlay, only about Rs 400 crore has actually been released so far, per government data reported this year, a gap that data-centre operators say reflects how much of the mission remains announcement rather than infrastructure that pays for itself.

For India’s compute industry, this is the uncomfortable part of being an early mover in a state-subsidised market: the GPUs are real, the capex is real, but the demand the government promised has been slower to materialise than the hardware. Industry executives argue the fix isn’t more chips,  India is already eyeing 100,000 GPUs by year-end, but a portal and procurement process that treats compute like infrastructure startups can actually plan around, not a grant they have to re-qualify for every few months. Until booking a GPU is as simple as renting one from a global cloud, empanelled firms warn, India risks having built one of the world’s cheapest AI compute pools with too few takers to fill it.
CT Bureau