India's GPU Tender Strains
Nine companies advanced past the technical evaluation in India’s GPU tender, but rising hardware costs and short contract durations are weakening the economics of leasing compute domestically. That suggests national AI ambitions may face commercial limits unless contract terms and financing models change. (telecom.economictimes.indiatimes.com)
India’s plan to rent out artificial intelligence chips at bargain rates just hit a wall that looks very ordinary: the machines got more expensive, but the contracts stayed short. In the fourth round of the IndiaAI Mission tender, nine companies cleared the technical stage, and several are already warning that the math no longer works cleanly for domestic providers. (economictimes.indiatimes.com) The chips in question are graphics processing units, which are the heavy-duty engines used to train and run large artificial intelligence models. India’s government wants to make those engines available through approved cloud providers so startups, researchers, and students do not have to buy the hardware themselves. (indiaai.gov.in) (pib.gov.in) This is not a side project. IndiaAI was approved in March 2024 with a budget of ₹10,371.92 crore over five years, and one of its core promises was to build national compute capacity through public-private partnerships instead of waiting for every lab or startup to raise its own capital. (static.pib.gov.in) (indiaai.gov.in) The first rounds moved fast. By January 31, 2025, the government said more than 18,000 compute units were available at subsidized rates, with eligible users getting up to 40% lower prices under the mission. By August 2025, the pool had climbed to 34,333 graphics processing units after additional tender rounds. (indiaai.gov.in) (economictimes.indiatimes.com) The fourth round was supposed to widen that pool again. The nine technically qualified bidders are Paradigmit Technology Services, Tata Communications, RackBank Datacenters, Netmagic Information Technology Services, E2E Networks, Yotta Data Services, Cyfuture India, Sify Digital Services, and UrsaCompute. (economictimes.indiatimes.com) Some of the offers are enormous on paper. Yotta said it is offering 17,000 Nvidia Blackwell B300 graphics processing units in this round, while Cyfuture bid 1,024 Nvidia Blackwell B200 units, and RackBank committed a mix of 1,024 Blackwell B300, Blackwell B200, and Hopper H200 units. (economictimes.indiatimes.com) The squeeze starts one layer below the chips. RackBank’s chief executive told The Economic Times that dynamic random-access memory modules, which act like the short-term working memory attached to these servers, jumped to nearly $2,000 from about $300 late in 2025 as supply tightened among Samsung Electronics, SK Hynix, and Micron. (economictimes.indiatimes.com) That matters because cloud providers are not buying one box at a time. They are financing whole clusters, and hyperscale companies that buy in much larger volumes can lock up components first, which leaves smaller domestic bidders paying more and waiting longer for the same parts. (economictimes.indiatimes.com) Then comes the contract problem. Providers say the current IndiaAI structure effectively gives them about a one-year revenue window on equipment that is usually financed over five to six years, and one recent IndiaAI order to E2E Networks was explicitly set for 360 days. (economictimes.indiatimes.com) (electronicsforyou.biz) That gap changes the return profile. RackBank said payback that once took about two years can now take up to four years, and its chief executive argued that contracts need to run at least two years before lenders will view these deployments as comfortably financeable. (economictimes.indiatimes.com) India is still pushing prices down hard. In earlier rounds, the mission said it was offering graphics processing units at less than $1 per hour in some cases, and the continuous empanelment model pressures new entrants to match or beat the lowest discovered rates. (economictimes.indiatimes.com) (moneycontrol.com) That is great news for startups shopping for cheap compute. It is harder news for the companies building the racks, wiring the data centers, and signing the hardware invoices, because every rupee cut from the hourly price has to be recovered from a machine whose components are getting costlier before it is even switched on. (economictimes.indiatimes.com) (moneycontrol.com) So the headline is not that India lacks bidders. The headline is that India has bidders, hardware demand, and a national mission, but the tender now depends on whether contract length, financing terms, and pricing rules can catch up with the cost of the machines the mission wants companies to deploy. (economictimes.indiatimes.com) (pib.gov.in)