India’s GPU tender hits economics snag

India just got nine more companies through the technical gate in its latest government tender for artificial intelligence chips, but several of those same bidders are warning that the math may not work if contracts stay this short. The fourth round is part of the IndiaAI Mission, and the qualified names include Tata Communications, Yotta Data Services, E2E Networks, Sify Digital Services, RackBank Datacenters, Cyfuture India, Netmagic IT Services, Paradigmit Technology Services, and UrsaCompute. (economictimes.com) This is a government effort to buy access to graphics processing units, which are the specialized chips used to train and run large artificial intelligence models. India’s Ministry of Electronics and Information Technology opened the first empanelment in August 2024 to make at least 10,000 of these chips available to startups, researchers, students, and public agencies. (dic.gov.in) The pitch from New Delhi was simple: instead of every startup building its own expensive chip cluster, the state would line up cloud providers and subsidize access. Under the IndiaAI Mission, the government approved a budget of ₹10,372 crore and built the program around public-private partnerships rather than a single state-owned data center. (negd.gov.in) That first round moved fast. Nineteen bidders submitted proposals by November 28, 2024, ten cleared technical evaluation, and financial bids were opened on January 22, 2025. (dic.gov.in; negd.gov.in) By March 6, 2025, the government had launched the IndiaAI Compute Portal and said users would be able to access more than 18,000 graphics processing units, with prices starting at ₹67 per graphics processing unit hour. The state also said eligible users could get subsidies of up to 40 percent through a coupon system. (pib.gov.in; moneycontrol.com; economictimes.indiatimes.com) Now the problem is showing up in the fine print. RackBank chief executive Narendra Sen told The Economic Times that the current IndiaAI structure works like a one-year contract, while financing for chip-heavy infrastructure usually stretches across five to six years. (economictimes.com) That mismatch matters because these providers are not buying a few servers; they are building facilities around power, cooling, networking, and imported components. Sen said returns that used to come in about two years can now take as long as four years, and he argued that contracts should run at least two years instead of one. (economictimes.com) The hardware bill is also moving in the wrong direction. Sen said dynamic random-access memory modules, the high-speed memory chips that sit next to artificial intelligence processors, jumped to nearly $2,000 from about $300 in late 2025 as supply tightened among Samsung Electronics, SK Hynix, and Micron. (economictimes.com) Big bidders are still showing ambition. Yotta said it is offering 17,000 Nvidia B300 chips in this round and is already building a 21,000-chip cluster, while Cyfuture bid 1,024 Nvidia B200 chips and RackBank offered a mix of B300, B200, and H200 systems. (economictimes.com) But India is also asking these vendors to meet strict local rules at the same time. Tender terms reviewed by MediaNama require artificial intelligence services to run from Indian data centers and bar providers from sending user-uploaded data outside India, even in encrypted or anonymized form. (medianama.com) So the latest tender result is not really a story about whether India wants more artificial intelligence capacity. It is a story about whether providers can lock in low prices for users while chip costs rise, financing runs longer than contract terms, and policy rules push them to build locally instead of renting the cheapest global capacity. (economictimes.com; medianama.com; moneycontrol.com)