The IndiaAI Mission Has Over 38,000 GPUs at Rs 65 an Hour and 12 Funded Model Builders, but Its 2025-26 Budget Was Revised Down From Rs 2,000 Crore to Rs 800 Crore

The Union Cabinet approved the IndiaAI Mission on 7 March 2024 with an outlay of Rs 10,371.92 crore. The government’s own documents now say what that bought: more than 38,000 GPUs, the chips that train AI models, rented out at Rs 65 per hour, and twelve organisations funded to build Indian models. The same documents show that the mission’s line in the Union Budget was revised for 2025-26 from Rs 2,000 crore to Rs 800 crore.

Bhashini’s request counts and the Krutrim story are in an earlier tuput piece. This one covers the money, the compute, the models and the people who dispute the approach.

What the Cabinet signed off on

PIB’s release of 7 March 2024 lists seven pillars: compute capacity, an innovation centre for home-built foundation models, a datasets platform, an application development initiative, FutureSkills (courses and data and AI labs), startup financing, and Safe & Trusted AI. It set the compute goal at 10,000 or more GPUs, built through public-private partnerships. A PIB backgrounder of 30 December 2025 gives the period as five years and says the 10,000 target has been passed. A division called IndiaAI, which sits under Digital India Corporation, runs the mission.

38,000 GPUs at Rs 65 an hour

The count grew in stages. PIB reported on 30 May 2025 that 15,916 GPUs had been added to 18,417 already empanelled (approved as suppliers), making 34,333. PIB’s backgrounder of 10 February 2026 says more than 38,000 GPUs are onboarded at Rs 65 per hour, which it calls nearly one third of the global average cost, along with 1,050 TPUs, a second kind of AI chip. It frames the strategy as following UPI and Aadhaar: open, affordable access.

On the second day of the India AI Impact Summit, IT Minister Ashwini Vaishnaw said 20,000 more GPUs would be added “in the coming weeks”, according to the National e-Governance Division. A Business Standard report of 30 June 2026 still gave the figure as more than 38,000.

Use is harder to see. A Lok Sabha reply reported by ANI on 6 August 2026 said 15 compute providers are empanelled, 237 projects have been approved for subsidised compute and 93.18 lakh GPU hours have been sanctioned. One GPU running nonstop for a year logs 8,760 hours, so that total equals about 1,060 GPUs for a year, by tuput’s arithmetic. The reply gives hours sanctioned, not hours used. It also said a purchase order has been issued for a system of about 1.1 exaflops (a billion billion calculations a second) at the NIC data centre in Shastri Park, Delhi.

Business Standard’s June report quotes Abhijeet Kate, co-founder of digiCloud Solutions, saying more than 60 per cent of startups asking for access want GPUs to run finished models, not to train new ones. He said the voucher system typically covers 40 to 60 per cent of compute costs, and that a seven-day lease cap on fine-tuning runs “creates real operational risk”.

Twelve organisations and where the money went

PIB’s backgrounder says more than 500 proposals came in for the foundation-model pillar. Twelve were selected: Sarvam AI, Soket AI, Gnani AI, Gan AI, Avataar AI, BharatGen (an IIT Bombay consortium), Zenteiq, Genloop, Intellihealth, Shodh AI, Fractal Analytics and Tech Mahindra’s Maker’s Lab. The government pays the actual cost of compute the teams use, and earmarks a further 25 per cent of that compute spending for datasets and staff.

Annexure I of the PIB release of 13 February 2026, tabled in the Rajya Sabha by Minister of State Jitin Prasada, gives each team’s support. BharatGen is far ahead at Rs 1,058.52 crore, of which Rs 990.92 crore is compute. Then come Sarvam at Rs 246.72 crore, Zenteiq at Rs 206.49 crore, Gnani at Rs 177.27 crore, Soket AI at Rs 177.08 crore, Fractal at Rs 137.91 crore and Gan AI at Rs 110.03 crore. The other five get Rs 49.50 crore or less. Adding the column gives Rs 2,194.29 crore, and BharatGen’s share is 48 per cent.

The sizes MeitY announced on 30 May 2025 were a 120-billion-parameter open model each from Sarvam and Soket AI, a 14-billion-parameter voice model from Gnani and a 70-billion-parameter multilingual model from Gan AI aimed at text-to-speech. Parameters are the adjustable numbers inside a model, a rough measure of its size. By the August Lok Sabha reply, 20 proposals had been identified for support, 12 large multimodal models and eight small language models, with the applicants keeping the intellectual property. ANI’s summary lists launches from Sarvam, Gnani, BharatGen and Avataar.

What Sarvam shipped, and a score from outside

Sarvam unveiled a 30-billion and a 105-billion-parameter model at the summit. Its blog post of 6 March 2026 announced the open release. It says both were trained from scratch in India on IndiaAI compute, on 16 trillion tokens for the 30B and 12 trillion for the 105B, with weights on AIKosh and Hugging Face. It also claims state-of-the-art results on Indian-language benchmarks. The 105B is smaller than the 120 billion parameters in the May 2025 announcement, and the post says Sarvam is ready to scale to larger models.

Artificial Analysis, an independent benchmarking firm, scored both on 2 April 2026. The 105B got 18 on its Intelligence Index and the 30B got 12. Among open reasoning models of about 100 billion parameters, the 105B trailed GLM-4.5-Air (23), INTELLECT-3 (22), Mistral Small 4 (27) and gpt-oss-120B (33), though all four use more parameters per token. The 105B did better on an agentic-tasks index, 25 against 21 and 20 for the first two. It scored 1.5 per cent on Terminal-Bench Hard, an agentic coding test, against 20.5 per cent for GLM-4.5-Air. On AA-Omniscience the 105B scored minus 60 and the 30B minus 72, and Artificial Analysis says both models hallucinate often relative to their accuracy and attempt far more questions than they decline. Sarvam’s previous model had scored 8.

The article reports no Indian-language benchmark, which is where Sarvam makes its claim, so that claim rests on Sarvam’s own numbers. Artificial Analysis has also revised its index since April. Its current page for the 105B shows an estimated 9 under the new version, marked as awaiting a fresh evaluation, so the April figures cannot be set against today’s leaderboard.

Bhashini, AIKosh and the safety institute

PIB’s February 2026 backgrounder says Bhashini, the translation platform, supports 36 or more languages, integrates over 350 AI models and has 450 or more active customers. The December 2025 backgrounder gave 20 languages for the same platform, and neither document says what changed.

AIKosh, the datasets platform, held 7,541 datasets and 273 models across 20 sectors in February, with 11,000 registered users and 26,000 downloads by December 2025.

On safety, PIB’s 13 February release says 13 projects have been selected under the Safe & Trusted AI pillar. IndiaAI’s call for partners of 9 May 2025 describes the IndiaAI Safety Institute as established, on a hub and spoke model with academic and private partners. Digital India Corporation later advertised for its director, with applications closing on 2 June 2026. The notice lists finalising the institute’s scope of work among the director’s duties.

The summit

PIB’s reply of 11 March 2026 says the summit ran from 16 to 21 February at Bharat Mandapam, the first in the series held in the Global South. It reports more than 100 countries, 22 heads of state or government and about 6 lakh people attending in person. The declaration was endorsed by 92 countries and organisations, and 13 model providers signed the New Delhi Frontier AI Impact Commitments. PIB lists investment announcements with expected investments above USD 250 billion, and says the summit was not designed to estimate growth or jobs.

India also joined the Pax Silica coalition at the summit, which PIB describes as focused on securing the global silicon supply chain. The February backgrounder cites the India Semiconductor Mission’s Rs 76,000 crore outlay and ten approved projects as of December 2025, and the current state of those plants has its own piece.

The Budget line and the critics

The Union Budget’s Notes on Demands for Grants for 2026-27, MeitY’s Demand No. 27, show the IndiaAI Mission at Rs 19.24 crore actual in 2024-25, Rs 2,000 crore budgeted and Rs 800 crore revised for 2025-26, and Rs 1,000 crore budgeted for 2026-27. Those three years add up to Rs 1,819.24 crore, or 17.5 per cent of the Rs 10,371.92 crore outlay, by tuput’s addition. India’s other national technology mission, quantum, has its own money trail in our National Quantum Mission piece.

Bharath Reddy of the Takshashila Institution wrote on 3 February 2026 that the revised and budgeted figures look lower than the Rs 10,300 crore planned over five years. He wrote that more than 85 per cent of the compute subsidy so far had gone to projects building indigenous generative AI models, that compute subsidies were over 40 per cent of the mission’s budget, and that the other pillars did not seem to be getting much funding.

Anirudh Suri of Carnegie India argued in February 2025, before the 38,000 count and before Sarvam’s release, that the mission’s effort had gone mostly into hardware and, to some extent, Indic-language models, leaving data, talent and research short. He cited the MacroPolo talent tracker: nearly all Indian AI researchers of 2019 ended up working abroad, and by 2022 almost 20 per cent had stayed in India.

Business Standard’s June report quoted Omprakash Subbarao, chief executive of FSID CORE, the R&D arm of the Indian Institute of Science, on datasets, which get a small share of the budget while compute takes nearly half. “50,000 H100s will not save you if you have a thin training corpus for Indic language,” he said.