The IndiaAI Mission

The Union Cabinet approved ₹10,371.92 crore for the IndiaAI Mission on 7 March 2024. By 9 February 2026, IndiaAI, an Independent Business Division under MeitY and the Mission's implementation agency, had received ₹400.94 crore. That is 3.9% of the outlay, 23 months into a 60-month plan. The cash figure comes from a Rajya Sabha annex, Unstarred Question 1668, answered on 13 February 2026.

The same reply said more than 38,000 GPUs had been "onboarded." Empaneling GPUs and receiving cash are different acts. One is a count of machines on the empanelment list. The other is money that reached IndiaAI.

Four ledgers, not one budget

Mission money sits in 4 ledgers. Public statements often mix them.

1. Outlay. The Cabinet ceiling of ₹10,371.92 crore, fixed in March 2024. This is the 5-year outer limit. It is not a promise that cash will arrive in year 1.
2. Estimate. The annual Budget Estimate and Revised Estimate. The Budget Estimate is the figure in the Union Budget at the start of the year. The Revised Estimate is the mid-year revision. In both years published here, that revision was a cut.
3. Receipt. Cash that reached IndiaAI. Parliament has published this for IndiaAI as a whole. The annex labels it "Amount received to IndiaAI."
4. Spend. What IndiaAI then pays to providers and institutions. Pillar-wise spend has not been published in the same annex that lists pillar-wise outlay.

The 13 February 2026 annex publishes estimates and amounts received for IndiaAI as a whole, and outlay pillar-wise. It does not publish spend. Without spend, a reader can see the ceiling and the cash in, but not how the cash was split across compute, models, skills, datasets, and safety.

The outlay split is in the annex and in MeitY's earlier Lok Sabha disclosure.

Tap to enlarge

Safe & Trusted AI gets about 1/5 of the overhead line. The sub-project breakdown inside Compute Capacity's ₹4,563.36 crore has never been published, as MediaNama noted in February 2025. Compute is the largest pillar. The annex does not break it down further. Slow receipts do not, by themselves, prove mismanagement. They show that ₹10,372 crore is the ceiling, not the cash IndiaAI had in its first 2 years.

The cash, year by year

The estimates were cut in both years that had a Revised Estimate. Receipts then ran behind those reduced estimates.

Tap to enlarge

MediaNama covered the FY 2024-25 cut in February 2025 and the same cash figures in April 2026.

FY 2024-25 was trimmed 69%, from ₹551.75 crore to ₹173 crore. FY 2025-26 was trimmed 60%, from ₹2,000 crore to ₹800 crore. Those original Budget Estimates are from the Union Budget papers, not from the Q.1668 table, which prints Revised Estimates and cash received. FY 2026-27 asks for ₹1,000 crore, half of the previous year's Budget Estimate.

Against the original Budget Estimate, FY 2024-25 delivered 3.9% of the opening figure, and FY 2025-26 delivered 19%. The 5-year ceiling was never meant to be released in 1 year. Receipts still lagged even the reduced estimates both years.

MediaNama's February 2025 coverage listed 3 possible reasons for the FY 2024-25 cut: cheaper GPU bids, phased rollout, and money moved to other priorities. The public record shows the cuts. It does not settle which reason dominated.

The Indian financial year starts on 1 April. The 9 February 2026 cut-off sits inside FY 2025-26. FY 2026-27 had not begun. The dash in that row is not a mid-year snapshot of an ongoing year. Any cash after 9 February 2026 needs a fresh parliamentary reply or Budget paper.

What officials said, and what the cash showed

Officials talked about GPU numbers. The cash shown to Parliament was much smaller.

The cash cut-off (9 February 2026) is earlier than the June 2026 GPU figure of 45,000+. Both series can be true. Empanelment can grow while cash stays thin if the programme rents capacity on demand and pays only for approved use. "45,000+ GPUs" is a shared-capacity count. It is not a count of machines in use. The 13 August 2026 PIB backgrounder does not reprint a new cash-received total.

How the compute model works

The main compute model is rental. IndiaAI empanels cloud providers, approves users, and subsidises access. The user pays a subsidised hourly rate. The provider is paid for hours used.

The 30 May 2025 PIB release published the lowest bid (L1) prices from 7 bidders including Cyfuture, Netmagic, Sify, Locuz, Vensysco, Ishan Infotech, and Yotta.

Rates in ₹ per hour. Source: PIB 2132817, Table 1.

The October 2025 PIB explainer says GPUs are available at a subsidised rate of ₹65 per hour. In January 2025, MeitY had announced a 40% subsidy with a target of ₹100 per hour. The 1X rows in the May 2025 table run from ₹29 to ₹323 an hour depending on the model and term. An H100 SXM is ₹153/hour on-demand. A B200 SXM is ₹323/hour. Multi-GPU rows in the same PIB table go higher. The February 2025 tender round averaged ₹115.85 per GPU-hour.

These numbers sit next to each other without a public reconciliation. ₹65 per hour is a headline subsidy rate. The L1 table is a bid sheet by model. A researcher using an L4 at the month rate pays far less than a lab using a B200 on demand. Averaging them into 1 number hides the product mix. A mature scorecard would publish hours by model class, not a single rupee-per-hour slogan.

GPUs pass through several stages before anyone uses them. The Economic Times, citing an IndiaAI letter to providers, reported these numbers.

Source: Economic Times, citing an IndiaAI letter to providers. Press-reported. Not a parliamentary utilisation table.

On those press-reported figures, 7,418 of 33,099 committed GPUs were in use (22.4%). That is arithmetic on the letter ET cited, not a parliamentary utilisation table.

Separately, in May 2025, before that letter, Economic Times reported that IndiaAI CEO Abhishek Singh said bids in the second tender offered 18,000 GPUs, that about 15,000 would clear technical checks, that more than 30,000 GPU requests were in the pipeline, and that a third tender round was open. Those figures are Singh speaking to ET. They are not in a PIB release. The later official count is in the 30 May 2025 PIB note: 34,333 GPUs empaneled (18,417 + 15,916). The January 2025 target had been 18,693 GPUs, announced when the initial 10,000 figure was expanded, before that May expansion.

MediaNama, covering the same tender, named some bidders' partners. Vensysco and Locuz with AWS. Ishan Infotech with Oracle. Netmagic with Google Cloud. That article does not give a share of onboarded GPUs by owner. The empanelment model rents access. The Mission does not, under that model, take title to the rented machines.

Lok Sabha Unstarred Question 2807, answered on 5 August 2026, added that 15 compute service providers had been empaneled over 4 rounds, and that a purchase order had been issued for a high-performance AI compute system of approximately 1.1 EFLOPS at the NIC data centre in Shastri Park, Delhi. The same reply said 237 projects had been approved for subsidised compute and 93.18 lakh GPU hours (9.318 million) had been sanctioned. The 13 August 2026 PIB backgrounder repeats those two totals with different verbs: projects that had accessed subsidised computing, covering those hours. Sanctioned, accessed, and consumed are not the same word. The August annex still does not print GPU-hours used, or hours per rupee received. 93.18 lakh is an approval or coverage total. MeitY has not published a consumption series next to it.

The National AI Compute Scorecard places India's rental model next to the US donation model, the UK purchase model, and the EU co-federation model. Consumed hours per rupee still do not appear in the published annexes.

Who is using the machines

Parliament also got user counts.

Tap to enlarge

The annex lists those 7 categories as separate rows. They sum to 305. The 33,099 committed GPUs come from the later Economic Times letter, not from that annex. More than 100 committed GPUs for every named user slot. The capacity is pooled and shared, so the ratio does not mean each user holds 100 GPUs.

The single largest named category is academic researchers, 114 of them. IndiaAI Fellows number 8. Takshashila warned in April 2025 that strict eligibility rules and slow processes would leave Mission compute underused, and placed India last among the US, China, and the EU on transparency and accountability.

IndiaAI has not published a wait-time series from application to first GPU-hour.

The Sarvam deal, and a larger line in the annex

Press reported a compute package for Sarvam. A later annex listed a larger support line for BharatGen. The company had already shipped 2 open-source models before the government selected it. Sarvam-1 at 2 billion parameters and Sarvam-M at 24 billion. The selection, announced on 26 April 2025, asks Sarvam to build a sovereign open-source model at 120 billion parameters, with use cases named "2047: Citizen Connect" and "AI4Pragati" (PIB, 30 May 2025). The company also works with UIDAI on generative AI for Aadhaar services (PIB, 12 Oct 2025). 12 teams had been shortlisted for foundation models by the February 2026 Rajya Sabha reply.

MediaNama, citing Economic Times, reported the terms of Sarvam's compute package.

Source: Economic Times, via MediaNama. Press-reported. Not in the Rajya Sabha PDF.

The subsidy is 40% of the bill, matching the subsidy level MeitY announced in January 2025. The support is compulsorily convertible debentures, not a grant. If the CCDs convert, the government ends up with a 1-2% ownership stake in Sarvam.

A later Lok Sabha annex, Unstarred Question 5417, answered on 25 March 2026, lists "support extended" to 12 organisations under the Foundation Models pillar. Sarvam's compute support worth is ₹246.72 crore, matching the press bill to a rounding. BharatGen, an IIT Bombay consortium, is listed at ₹1,058.52 crore (₹990.92 crore compute, ₹67.60 crore non-compute). That is the largest line in the annex, more than 4 times Sarvam's. The 12 rows sum to ₹2,194.29 crore. The Foundation Models pillar outlay is ₹1,971.37 crore. The annex labels the columns "Compute Support Worth," "Non-Compute Cost," and "Total Cost." It does not say those sums had been paid. IndiaAI had received ₹400.94 crore by 9 February 2026. The March annex does not reprint a received-cash total.

The structure can be read 2 ways. As upside protection, the public keeps a share of the value it subsidised. As dilution, the startup gives up equity for compute it might have bought outright. Some IndiaAI participants have said they do not want to give up equity for compute support. The Sarvam terms, reported in June 2026, offer a concrete case of that trade-off. Neither reading is proven by the press table alone. The conversion price, licence terms, and any converted equity need a primary document.

Sarvam closed a $234 million round in June 2026, with HCLTech leading at $150 million and Khosla Ventures and Peak XV participating, at a valuation of about $1.5 billion (MediaNama, June 2026). MediaNama wrote that US export controls on Anthropic's frontier models came days before Sarvam announced the first close. Timing does not prove cause.

In January 2026, Tamil Nadu signed a ₹10,000 crore MoU with Sarvam for a sovereign AI park. That MoU figure is close to the central Mission outlay of ₹10,371.92 crore. An MoU is not cash received. Private funding rounds are a third published envelope. Counting only the central cash understates other published figures. Counting MoUs as cash overstates spend.

A related piece, Sovereign AI Debt, treats the Sarvam CCD as the smaller cousin of sovereign funds that finance AI compute with equity, convertibles, and bank debt.

What the Mission has actually delivered

Foundation models. MeitY received 506 proposals by 30 April 2025 (PIB, 30 May 2025). The 5 August 2026 Lok Sabha reply says that, against those 506 applications, 20 indigenous foundation-model proposals had been identified for support: 12 large multimodal models and 8 small language models, with intellectual property retained by the applicants. The same reply lists "notable outputs already launched": Sarvam AI's 30 billion and 105 billion parameter models, and Indus; Gnani.AI's speech-to-speech model; BharatGen's Param2-17B, Patram-7B, and Shrutam-2; and Avataar AI's video generation model. Those names and sizes are the ministry's list as of that answer. They are not an independent eval table. An earlier PIB note described Sarvam's selection as a 120 billion parameter effort. The August 2026 annex lists 30 billion and 105 billion as launched. This essay does not collapse those figures into one model. They may be successive releases, parallel products, or a change of target. The annex does not say which.

A fair public scorecard would list model name, parameter count, open weights or not, evaluation set, and licence. That table does not yet exist in the parliamentary annex.

Applications. 30 applications were approved for India-specific AI use cases by July 2025, in healthcare, agriculture, climate change, governance, and assistive learning (PIB, 12 Oct 2025). A 29 July 2026 PIB note, from a Lok Sabha submission the same day, said the Mission had launched 11 national-level hackathons and innovation challenges. Of the original Innovation Challenge, 30 solutions were shortlisted for pilot implementation and 5 teams, one per sector, were selected for national-level deployment, with pilots underway. After the India-AI Impact Summit, held in New Delhi from 16 to 21 February 2026, the Cabinet Secretariat collected 762 problem statements from 62 Ministries and Departments and shared them with MeitY. The 5 August 2026 Lok Sabha reply cites 62 AI prototypes developed and 20 AI solutions deployed across public-sector institutions. Problem statements, shortlisted pilots, prototypes, and deployments are different counts. The annexes do not list live users per application.

Centres of Excellence. The same August reply said 58 AI Centres of Excellence had been approved, of which 22 had been initiated across 13 States/UTs. Neither that reply nor the 29 July PIB note publishes a spend line for the CoEs.

AIKosh. MeitY said 367 datasets were on the platform in May 2025 (PIB, 30 May 2025). By July 2025, the platform reported over 3,000 datasets and 243 AI models across 20 sectors, with 265,000 visits, 6,000 registered users, and 13,000 downloads (PIB, 12 Oct 2025). The 13 August 2026 PIB backgrounder puts the July 2026 count at over 14,000 datasets and 331 AI models. Count alone is not quality. MediaNama has noted that MeitY has not published a scrub protocol matching the privacy risk of residual identifiers in "AI-ready" public datasets.

FutureSkills. More than 200 students had received fellowships by July 2025, according to PIB. 26 institutes had onboarded PhD students. The same note said 27 labs had been identified with NIELIT, and that states and Union Territories nominated 174 ITIs and polytechnics. The Rajya Sabha reply of 13 February 2026 says more than 8,000 undergraduates, 5,000 postgraduates, and 500 PhD students "are supported" for talent development, and that 27 labs have been established with 543 more identified. Those larger figures are the government's claim of students supported under the Mission, not a separate verified enrolment census published with names and institutes.

The 29 July 2026 PIB note gives a later awarded count: 686 fellowships nationally across 178 institutions. Awarded is not the same word as supported. The same note says 543 Data and AI Labs at ITIs and polytechnics have been approved for establishment. Q.1668 said 27 established and 543 more identified. Approved, identified, and established are different words.

A later account, in ThePrint on 7 August 2026, citing the Thirty-First Report of the Standing Committee on Communications and Information Technology (presented to Lok Sabha and laid in Rajya Sabha on 6 August 2026), gives selection counts rather than "supported" counts. According to that reporting of MeitY's action-taken replies: in 2024 the ministry selected 150 undergraduate fellows against a target of 5,000, 140 postgraduate fellows against 2,000, and 3 PhD fellows against 100. In 2025 the UG target rose to 8,000 with 159 selected, the PG target rose to 5,000 with 101 selected, and PhD selections were 199 against a target of 167. Those selection counts are ThePrint's account of the parliamentary report. "Supported" and "selected" are different words, not two parts of one total.

ThePrint, citing the same report, also wrote that MeitY had spent 32% of 2025-26 revised-estimate funds as on 31 December 2025. That is a spend ratio against RE, with a December cut-off. It sits beside, and is not the same as, the ₹379.15 crore received by IndiaAI as on 9 February 2026 in Q.1668. Received and spent can differ. The 2 dates differ. Enrolment is not a labour-market outcome. The Mission can count seats. Employers will count skills.

Startup Financing. The IndiaAI Startups Global programme launched in March 2025, sending 10 Indian startups to Europe with Station F and HEC Paris (PIB, 12 Oct 2025). The Startup Financing pillar outlay is ₹1,942.50 crore. The October PIB note does not publish a disbursement list for that pillar. Without a spend table, the public record shows the programme name and the 10 startups, not how much of the ₹1,942.50 crore line was paid.

Safe & Trusted AI. 8 projects were selected in the first round, covering machine unlearning, bias mitigation, and explainability. More than 400 applications came in for the second round. An expression of interest for partner institutions in the IndiaAI Safety Institute was published on 9 May 2025 (PIB, 12 Oct 2025). The August 2026 Lok Sabha reply puts the selected-project count at 13. The 29 July 2026 PIB note repeats 13 Responsible AI projects and says India AI Governance Guidelines were released on 5 November 2025. It also says an IndiaAI Safety Institute is being established. The pillar's outlay is still ₹20.46 crore in the February annex, the smallest line in the Mission.

AIRAWAT at C-DAC was reported down for about 3 months in late 2024, disrupting startups that depended on it. MediaNama treated the duration as reported, not as a ministry confirmation. Bhashini, launched in July 2022, has crossed 1 million downloads and supports 20 Indian languages with more than 350 translation models (PIB, 12 Oct 2025). Neither replaces the IndiaAI money. Adding their GPU numbers to IndiaAI's would count the national total twice. Prior programmes show both delivery and operational risk.

Private pledges, public metrics

Private pledges are larger than the Mission cash. In December 2025 Amazon said it would invest over $35 billion in India by 2030 across its businesses. Microsoft committed $17.5 billion over 2026-2029 for cloud, AI infrastructure, operations, and skilling. Combined, that is over $50 billion in company pledges. Not all of it is AI compute. It is still separate from, and far larger than, the ₹400.94 crore IndiaAI had received. Under the empanelment model, the Mission pays for access. The rented machines stay with the providers.

India's most successful digital programmes reported completed transactions: UPI payments and Aadhaar authentications. That argument is developed in the API States and bureaucracy essays on this site. IndiaAI reports onboarded GPUs, Cabinet outlays, and sanctioned GPU-hours. Transaction metrics force a programme to show use. Capacity and sanction metrics show readiness and approval.