IndiaAI Compute Capacity: empanelled GPU clouds and the IndiaAI Compute Portal

NITI Aayog's 2018 national AI strategy proposed AIRAWAT, a pooled AI cloud of about 100 petaflops linking research centres. When the Union Cabinet approved the IndiaAI Mission in March 2024 it included a compute pillar, which the government's February 2026 reply shows to carry Rs 4,563.36 crore, about 44% of the outlay.¹ The pillar aims to provide high-end compute (GPUs) to all, including MSMEs and startups, "at an affordable cost". MediaNama reported in early 2025 that AIRAWAT, the government's earlier AI platform, had suffered a three-month downtime that disrupted GPU and storage access for startups.¹⁸

The design avoids a government-owned data centre. MeitY empanels private AI service providers through requests for empanelment, and the Mission subsidises approved users' consumption. By February 2026 the Rajya Sabha was told that "National AI compute capacity is operationalized through empaneled AI service providers offering GPU access on cloud at subsidized rates", all of them MeitY-empanelled cloud service providers.¹

First round (2024-25). MeitY issued a request for empanelment for AI compute and cloud services in August 2024 and held a pre-bid meeting with about 50 entities. Bidding opened on 2 December 2024 and drew 19 bidders, among them Jio Platforms, Tata Communications and Yotta Data Services.⁶ The original conditions required an Indian legal entity with average annual turnover above Rs 100 crore (Rs 50 crore for main consortium members), at least 1,000 AI compute units on a cloud platform with a self-service portal, and services delivered from Indian data centres. After feedback that the conditions excluded startups, an October 2024 amendment reduced the turnover thresholds to Rs 50 crore for primary bidders and Rs 25 crore for other consortium members and lowered the minimum performance thresholds for compute units.⁷ Selected providers are empanelled for 36 months, extendable by 12.⁶

Capacity announcements. The Mission's initial target was 10,000 GPUs. In late January 2025 Minister Ashwini Vaishnaw said a further 8,693 GPUs would be added, taking planned capacity to 18,693. The reported average winning bid was Rs 115.85 per GPU hour; a 40% government subsidy was meant to bring costs to about Rs 100 per hour, against a global range of USD 2.5 to 3. Bidders included Jio, E2E, NxtGen, Locuz and CtrlS.⁸ MediaNama reported in May 2025 that bids for a second round proposed 18,000 GPUs, of which about 15,000 were expected to be offered after technical qualification, and that seven companies had been shortlisted: Netmagic IT Services, Vensysco Technologies, Cyfuture India, Yotta Data Services, Locuz Enterprise Solutions, Ishan Infotech and Sify Digital Services.⁵

By 31 August 2025 the governance guidelines counted more than 38,231 GPUs "made available to startups, researchers and developers at subsidised rates", and said a secure cluster for 3,000 next-generation GPUs was being built for sovereign and strategic uses.¹¹ At the India AI Impact Summit in February 2026 the government said it planned to add more than 20,000 GPUs to the existing base of 38,000.¹⁰ In its February 2026 Rajya Sabha reply MeitY reported 14 empanelled AI service providers, with data centres in Mumbai, Navi Mumbai, Hyderabad, Bengaluru, Noida and Jamnagar.¹

The IndiaAI Compute Portal was launched by Minister Ashwini Vaishnaw on 6 March 2025, the Mission's first anniversary, offering AI compute, network, storage and platform services at discounted rates, with up to a 40% subsidy on cloud AI compute.⁹ The portal describes the process as follows:³

• Users log in with DigiLocker, e-Pramaan or Jan Parichay, register, and wait for verification.
• They submit a compute request with a draft bill based on the providers' rate list and a subsidy request, between the 1st and the 25th of the month.
• Requests of under 5,000 GPU hours are auto-approved; larger requests go to a committee (PMEC). The approved list is published on the 10th of the next month, and approvals are valid for 30 days.
• IndiaAI pays the subsidy and users pay the remainder to the provider.

Eligibility differs by category. Researchers and academics need an h-index of 5 or more, journal publications with an impact factor above 2.5, or 150 or more citations. Startups and MSMEs need DPIIT recognition, AI or ML experience and either annual revenue of Rs 50 lakh or funding of Rs 100 lakh. Early-stage startups need four years of AI or ML experience and a letter from the IndiaAI CEO or a senior official, and government entities need a letter from an authorised officer.³ The portal lists empanelled providers including CMS, CtrlS, Cyfuture, E2E Cloud, Ishan, JPL, Locuz, NxtGen, NTT, Neysa, Orient, Sify, Tata Communications, Vensysco and Yotta.³ IndiaAI has also notified an End User Policy governing GPU access and subsidy allocation.¹

A Rajya Sabha question in February 2026 referred to subsidised GPU access at a rate of "65 per hour" (the currency symbol is missing in the published text). The Minister's written answer did not state a rate, and it reported users rather than utilisation: 114 academic researchers, 47 startups and MSMEs, 36 early-stage startups, 10 early-stage researchers, 32 students, 8 IndiaAI Fellows and 58 government entities.¹

The compute pillar underwrites the Mission's foundation-model programme. In March 2026 the government told the Lok Sabha that funding for the twelve selected organisations is "provided in calibrated manner to cover compute usage costs". The annexure shows compute support of Rs 246.72 crore for Sarvam AI, Rs 177.27 crore for Gnani, Rs 88.02 crore for Gan AI, Rs 162.47 crore for Soket AI, Rs 12.88 crore for Avataar AI, Rs 990.92 crore for BharatGen, Rs 137.91 crore for Fractal, Rs 2.66 crore for Tech Mahindra, Rs 165.19 crore for Zenteiq, Rs 2.09 crore for Genloop, Rs 41.50 crore for Intellihealth and Rs 7.52 crore for Shodh AI.²

MediaNama reported in June 2026 that Sarvam received 4,096 Nvidia H100 GPUs for six months and the Mission's largest compute subsidy, Rs 98.68 crore on a compute bill of about Rs 246.7 crore (roughly 40%), and that the government might obtain a 1-2% minority stake through compulsorily convertible debentures issued in return for compute support.¹² The governance guidelines also cite access to compute as a requirement for safety evaluations, so that developers can "test and validate the effectiveness of guardrails" at population scale.¹¹

Size of the pillar. At a December 2024 roundtable some participants questioned whether 44% of the Mission's outlay should go to compute rather than datasets, skilling or research, while others argued that inference demand would grow and that 44% could be a floor.⁸ MediaNama noted that a breakdown of the pillar's sub-projects was not public.⁸

Under-use. MediaNama reported in May 2025 that the Takshashila Institution had warned that capacity risked being under-used because few projects would qualify for subsidies and because of slow bureaucratic processes.⁵

Eligibility for suppliers. The first tender's turnover and capacity conditions drew industry feedback that they were too high for startups and smaller companies, and were relaxed in October 2024; all empanelled providers are MeitY-empanelled cloud service providers.¹⁷

Opacity. The published data are counts of users and GPUs. Utilisation rates, per-provider capacity and the average subsidy per user have not been disclosed in the documents reviewed.¹

As of 15 October 2026 the pillar is ongoing. On or around 7 October 2026 C-DAC, under MeitY, issued a request for empanelment inviting AI service providers to supply up to 10,000 AI Compute Units, a term covering GPUs, TPUs and other accelerators, which C-DAC may scale beyond 10,000 during the empanelment's validity. Clusters would be of 4,000, 1,000 or 512 units, each cluster needing at least 512 units at a single data centre in India, and bidders must keep all infrastructure, data, logs, backups and disaster recovery in India and isolate C-DAC's data from other customers.⁴ On 8 October Vaishnaw described a 'first tender' of 5,000 of the 10,000-GPU estimate and said the government would use an 'empowerment approach' instead of traditional tenders. MediaNama's report did not give prices or a submission deadline, and the call did not specify how many units would be procured in the first round.⁴