India’s GPU tender advances nine bidders

India’s plan to build a national pool of artificial intelligence chips just hit a strange milestone: more companies cleared the technical gate, even as some bidders warned the deal may not make financial sense on the terms offered. Nine companies moved ahead in the latest stage of the government tender, after an earlier round had already narrowed a larger field of 19 bidders. (telecom.economictimes.indiatimes.com, indiaai.gov.in) This is not a factory tender. It is a government effort to rent out graphics processing units, the chips used for training and running artificial intelligence models, through cloud providers so startups, researchers, and public agencies can buy computing time instead of buying entire servers. (indiaai.gov.in, indiaai.gov.in) India started this push with a target of 10,000 graphics processing units under the IndiaAI Mission, then expanded the pool as more suppliers came in. By May 2025, the government said common compute capacity had risen to 34,000 graphics processing units with support from industry partners. (indiaai.gov.in, telecom.economictimes.indiatimes.com) The sales pitch from New Delhi has been simple: make advanced computing cheap enough that an Indian startup can rent it by the hour like electricity. In March 2025, Information Technology Minister Ashwini Vaishnaw said some graphics processing units on the IndiaAI portal would be available for 67 rupees an hour after government support. (telecom.economictimes.indiatimes.com, indiaai.gov.in) The problem is that the chips underneath that promise have become much more expensive. In March 2026, a senior electronics ministry official said graphics processing units now account for almost 90 percent of the cost of artificial intelligence servers, up from about 70 percent earlier. (telecom.economictimes.indiatimes.com) That shift changes the math for every bidder. If a cloud provider has to lock in prices for a short government contract while chip costs, memory costs, power bills, and financing costs keep moving, the provider can win the tender and still lose money over the life of the deal. (telecom.economictimes.indiatimes.com, telecom.economictimes.indiatimes.com) India’s own policymakers have been flagging the same pressure from another angle. A parliamentary committee said in March 2026 that high hardware costs, supply-chain delays, and the power and water demands of data centers were major risks in the country’s plan to build large graphics processing unit clusters. (telecom.economictimes.indiatimes.com) The bidder list shows why the government wants this market to work. Earlier stages included companies such as Jio Platforms, Tata Communications, Yotta Data Services, E2E Networks, CtrlS Datacenters, and NxtGen Datacenter and Cloud Technologies, which are the kinds of operators that already run data centers, cloud systems, or both. (indiaai.gov.in, indiaai.gov.in) What this tender is really testing is not whether India can find enough chips. It is whether the government can design contracts that let private operators recover the cost of machines that depreciate fast, consume huge amounts of electricity, and can sit idle if demand forecasts are wrong. (indiaai.gov.in, telecom.economictimes.indiatimes.com, telecom.economictimes.indiatimes.com) If the contract terms stretch too little and the price ceilings sit too low, the state gets a cheap headline but not durable capacity. If the next round fixes duration, utilization assumptions, and pricing visibility, India could end up with something more valuable than a one-time subsidy: a market where artificial intelligence computing is actually worth supplying. (telecom.economictimes.indiatimes.com, telecom.economictimes.indiatimes.com)