India advances GPU tender

India has moved nine companies into the next stage of a government tender for artificial intelligence chips, and the list includes Tata Communications, Yotta Data Services, E2E Networks, Netmagic IT Services, Sify Digital Services, RackBank Datacenters, Cyfuture India, Paradigmit Technology Services, and UrsaCompute. The tender is part of the IndiaAI Mission, which is trying to buy computing power the way a government buys roadwork: through approved vendors, fixed bids, and public contracts. (economictimes.com, indiaai.gov.in) The thing being bought is not software. It is access to graphics processing units, the specialized chips that train and run large artificial intelligence models, and IndiaAI says its compute program is meant to create a public-private pool of more than 10,000 such units for startups, researchers, and public projects. (indiaai.gov.in, dic.gov.in) India started this push because buying time on top-end chips has become a national bottleneck. In April 2025, the Ministry of Electronics and Information Technology assigned the first subsidized workloads to Yotta, E2E Networks, and NxtGen Cloud Technologies, with access priced at less than $1 per hour and an average discount of 42 percent from market rates. (economictimes.com) That first wave was already larger than the original headline target. Yotta, E2E Networks, and NxtGen together were set to supply more than 11,600 graphics processing units, while later rounds and continuous empanelment were designed to keep adding vendors that could match or beat the lowest discovered prices. (economictimes.com, moneycontrol.com) Now the strain is showing up in the math. Yotta told The Economic Times it is offering 17,000 Nvidia Blackwell B300 chips in this round, while Cyfuture bid 1,024 Nvidia Blackwell B200 chips and RackBank committed a mix of 1,024 B300, B200, and Hopper H200 chips, which means bidders are promising expensive hardware before they know whether the contract terms will let them earn the money back. (economictimes.com) One bidder said memory prices have blown out fast enough to wreck older assumptions. RackBank chief executive Narendra Sen told The Economic Times that dynamic random-access memory modules had jumped to nearly $2,000 from about $300 late in 2025, and he linked that surge to geopolitical disruption, supply shortages, and bulk buying by hyperscale cloud companies. (economictimes.com) The contract length is the other pressure point. Sen said the current IndiaAI structure runs for one year, while financing for a graphics processing unit cluster usually stretches across five to six years, and he said payback periods that once took about two years can now take up to four. (economictimes.com) That mismatch turns a chip tender into a balance-sheet gamble. A company can win by offering low prices today, but it still has to buy servers, memory, networking gear, power capacity, and data-center space up front, then hope the government keeps enough demand flowing through the portal long enough to cover the build-out. (economictimes.com, moneycontrol.com) India is trying to solve that with a rolling system instead of a one-shot auction. Since February 2025, the government has allowed continuous empanelment, refreshed every quarter, and new vendors can join only if they meet the technical rules and match or undercut the lowest existing rates. (moneycontrol.com) That keeps prices low for startups and universities, but it also keeps squeezing suppliers just as the hardware gets pricier. The result is a familiar problem in infrastructure: the state can accelerate access to scarce machines, but it cannot repeal the cost of capital, the chip cycle, or the time it takes for a giant cluster to pay for itself. (economictimes.com, indiaai.gov.in)