For most of the past three years, the binding constraint on AI work in India was not talent or ideas but access to graphics processing units. Startups queued for overseas cloud capacity, researchers rationed experiments, and enterprises that wanted to keep sensitive data in the country had few local options for serious workloads. That picture has changed quickly. The IndiaAI Mission now reports more than 38,000 GPUs onboarded to a common compute facility, data centre capacity has passed 1.5 gigawatts, and Indian GPU cloud providers sell on-demand capacity by the hour.
Yet "AI compute India" is not one market. A startup applying through the IndiaAI portal, a bank running models on its own premises, a research lab and a manufacturer buying capacity from a hyperscaler are drawing on different pools of compute, at different prices, under different rules. Headline GPU counts say little about who can actually use them, at what cost, or with what conditions attached.
This article sets out what is known about public and commercial AI compute in India, separates the routes through which different users reach it, and offers an access map that shows who can use which compute, at what price signal, and with what limits. The economics of on-premise deployment and the sovereign AI stack are covered in separate PanScience Research pieces and are referenced rather than repeated.
The public pool: what the IndiaAI Mission has built
The IndiaAI Mission was approved by the Union Cabinet in March 2024 with an outlay of ₹10,371.92 crore over five years. Compute is its largest pillar. A PIB document of 12 October 2025, "Transforming India with AI", states that "from an initial target of 10,000 GPUs, India has now achieved 38,000 GPUs", and that these "are available at a subsidized rate of just ₹65 per hour." A National e- Governance Division release of 2 April 2026 repeats that more than 38,000 GPUs "for common compute facility have been onboarded through the AI compute portal", provided to "Indian start-ups and academia."
The figures need careful reading. Onboarded capacity is not the same as capacity that is installed, running and allocated. Inc42 reported in September 2025 that the country had deployed more than 17,300 GPUs as of June 2025, from providers including Yotta Data Services and NxtGen Cloud Technologies. Business Standard reported on 30 June 2026 that over 18,000 GPUs were empanelled in the early phases and that more than 38,000 were "now available through the broader compute ecosystem at subsidised rates". Communications Today, citing a Lok Sabha reply of August 2026, reported 93.18 lakh GPU hours sanctioned across 237 approved projects and 15 compute providers empanelled for subsidised support. The same publication noted that some reported figures on deployed versus committed GPUs had not been confirmed by MeitY, IndiaAI or C-DAC.
A simple calculation shows what the sanctioned hours mean. 93.18 lakh GPU hours is about 9.3 million hours. A single GPU running continuously for a year provides 8,760 hours. The sanctioned hours are therefore equivalent to roughly 1,060 GPUs running for a full year. That is substantial for the startups and researchers who receive it, and it is a reminder that sanctions, not the onboarded fleet, are the better measure of how much public compute has been put to work. The sanctions may also span more than one year, which would lower the annual equivalent further.
The ₹65 figure is a headline rate. Neither the PIB document nor the NeGD release specifies the GPU types or how the rate varies across them. Business Standard quoted a cloud executive saying the voucher system "typically covers 40-60 per cent of compute costs", with users paying the rest. Communications Today cited a PIB factsheet referring to a ₹100 per hour tier. The honest summary is that subsidised public compute in India costs a fraction of commercial rates, but the exact rate depends on hardware and scheme terms that applicants must check.
Eligibility: the public pool is not for everyone
The public pool is targeted. The NeGD release names start-ups and academia. Business Standard lists eligible applicants as startups, MSMEs, researchers, academic institutions, government agencies and other approved entities. Large private enterprises are generally not the intended users. For a bank, insurer, manufacturer or IT services firm, the IndiaAI portal is mostly relevant indirectly, through the Indian models and startups it supports, rather than as a source of their own capacity.
The pattern of demand is shifting too. Business Standard quoted the same cloud executive estimating that training still accounts for about 80% of GPU hours consumed, while more than 60% of startups requesting access want GPUs for inference. These are an industry participant's estimates, not official data, but they point to a mismatch that matters for planning: a pool built and allocated around training projects will increasingly be asked to serve steady inference loads, which need different contracts, availability guarantees and pricing.
The commercial pool: Indian GPU clouds and data centres
Outside the subsidised pool, compute in India is sold commercially by Indian GPU cloud providers, by global hyperscalers operating Indian regions, and through colocation for enterprises that buy their own hardware.
Commercial prices give a reference point for the subsidy. E2E Networks' public pricing page lists on- demand rates per GPU-hour, excluding taxes, of US$3.99 for an H100 80GB, US$5.29 for an H200, US$2.10 for an A100 80GB and US$1.20 for an L40S. At ₹95 to the dollar, the H100 rate is about ₹379 per GPU-hour. Comparing that with ₹65 is indicative only, since the public rate's hardware mix is not published, but the gap explains why the portal is oversubscribed relative to its sanctions.
The physical base is growing fast. A JLL report summarised by W.Media in June 2025 found that India's data centre capacity "surged past the 1 GW mark at the end of 2024", with about 1,030 MW of live IT load, of which 993 MW was occupied, and projected capacity of up to 1.8 GW by the end of 2027. JLL found Mumbai accounted for nearly half of supply in the second half of 2024, and expected Mumbai and Chennai to account for 64% of new additions. CBRE, in a report published on 12 November 2025, put operational capacity at about 1,530 MW as of the first nine months of 2025, with about 260 MW added in that period, nearly 90% of capacity concentrated in Mumbai, Chennai, Delhi- NCR and Bengaluru, and about US$94 billion of investment commitments between 2019 and September 2025. Gartner's November 2025 forecast for India expected data centre systems spending to grow 20.5% in 2026, the fastest of any IT segment.
Data centre megawatts and GPUs are not the same thing, and the published reports do not say what share of capacity is AI-ready. An illustrative calculation gives scale. If each high-end GPU, with its share of server, networking and cooling, draws between 1 and 2 kW, an assumption rather than an observed figure, then 38,000 GPUs would require roughly 38 to 76 MW. That is a small share of the 1,530 MW CBRE counts. The constraint on AI compute in India is therefore less about total data centre space and more about high-density, liquid-cooled halls, power availability at specific sites, and the capital to buy accelerators.
The AI compute India access map: who can use which compute
The table below is this article's analytical contribution. It sets out the five main routes to AI compute for users in India, who can use each, the price signal, the main limits and the workloads each suits. Price signals are drawn from the sources cited above; where no reliable public figure exists, the table says so rather than estimating.
Three readings follow from the map.
First, the subsidy reaches a specific population. The public pool is most valuable to startups and researchers whose alternative is commercial on-demand pricing. For them, it changes what is affordable to train.
Second, for large enterprises, the decision is between commercial cloud, own hardware and APIs. The deciding factors are utilisation and data control. Steady high-volume inference on sensitive data favours owned or reserved capacity within India; spiky or experimental work favours on-demand cloud or APIs. A separate PanScience Research piece on on-premise AI economics sets out the break-even reasoning.
Third, data rules shape the map as much as price does. The Digital Personal Data Protection Rules, notified on 14 November 2025, subject significant data fiduciaries to restrictions on transferring specified personal data outside India, and the RBI's 2018 circular requires payment system data to be stored only in India. For enterprises in these categories, the "global model APIs" route is constrained for some workloads regardless of cost, which moves demand towards domestic compute.
An illustrative cost comparison
To show what the price difference means in practice, consider a fine-tuning run that uses 8 GPUs for seven days, or 1,344 GPU-hours. This is an illustrative scenario, not an observed project.
The comparison assumes the subsidised rate applies to comparable hardware, which the published sources do not confirm, and ignores storage, networking, taxes and engineering time. Even so, a gap of roughly six times explains why eligible startups pursue the portal and why the allocation process, rather than the price, becomes their binding constraint.
What this means for enterprises
For Indian enterprises, the expansion of domestic compute has two practical effects.
The first is choice. An enterprise that must keep regulated data in India can now run serious workloads on domestic GPU clouds or in colocation, rather than choosing between overseas APIs and nothing. The second is a supply chain of Indian models and tools. Public compute subsidises the training of models and products by startups that enterprises may later buy. The Mission's support for Indian foundation models, covered in a separate PanScience Research piece on the sovereign AI stack, is one route by which public compute reaches enterprises indirectly.
Within the PanScience Innovations portfolio, Eigen works on this enterprise side of the map. PanScience reports that Eigen provides on-premise AI for regulated industries (company-reported). Its relevance here is as an example of the "own hardware" route: for an enterprise whose data cannot leave its environment, the relevant compute question is not access to a national pool but sizing, procuring and keeping busy a modest on-premise or colocated cluster, and choosing models that run efficiently on it. A separate PanScience Research piece on small language models covers why model size matters most on this route.
What to watch
Four indicators will show whether India's AI compute base is meeting demand.
Sanctioned and consumed GPU hours, not onboarded GPUs. Parliamentary replies that report hours sanctioned and used are more informative than fleet counts.
The share of inference in public allocations. If demand is moving towards inference, as industry participants suggest, allocation rules built around project-based training will need to adapt.
High-density capacity. Reports from JLL, CBRE and others that separate AI-ready, liquid-cooled capacity from general data centre space will be the best measure of physical readiness.
Price convergence. If commercial GPU prices in India fall towards the subsidised rate, the subsidy's role shifts from affordability to allocation for strategic priorities.
Conclusion
India's AI compute base has grown faster than most observers expected. The IndiaAI Mission reports more than 38,000 GPUs onboarded at a headline subsidised rate of ₹65 per hour, data centre capacity has passed 1.5 GW, and commercial GPU clouds sell capacity in India by the hour. But access is segmented. The public pool serves startups, researchers and public institutions; enterprises mostly choose between commercial cloud, their own hardware and global APIs, with data rules often deciding the answer before price does. Reading the market through the access map, rather than through headline GPU counts, gives a clearer picture of who can build what, where, and at what cost.
Methodology and sources
This article draws on sources opened by the author. IndiaAI Mission figures are from PIB's "Transforming India with AI" document of 12 October 2025 (static.pib.gov.in/WriteReadData/specificdocs/documents/2025/oct/doc20251012664501.pdf and doc20251015666901.pdf) and the NeGD release of 2 April 2026 (negd.gov.in/press_release/indiaai- mission-expands-ai-ecosystem-with-affordable-compute-and-startup-support). Deployment, eligibility, voucher and demand-mix figures are from Business Standard, 30 June 2026 (business- standard.com/technology/artificial-intelligence/india-ai-compute-race-subsidised-gpu-capability-gap- 126062900734_1.html), Inc42, 10 September 2025 (inc42.com/buzz/centre-to-set-up-600-data-labs- deploy-38k-gpus-indiaai-mission-ceo) and Communications Today (communicationstoday.co.in/indiaai-mission-faces-gap-between-capacity-and-deployed-gpus), which cites a Lok Sabha reply; industry executives' estimates are reported as such. Data centre figures are from CBRE's "India's data centre market in a new era" (cbre.co.in/insights/reports/india-s-data-centre- market-in-a-new-era) and JLL's "India Data Center Market Dynamics" as summarised by W.Media (w.media/indias-dc-industry-capacity-expected-to-touch-1-8-gw-by-2027-jll-report). Commercial GPU prices are from E2E Networks' pricing page (e2enetworks.com/pricing), read in October 2026. The Gartner forecast is editor-verified via CXOToday. Currency is converted at ₹95 to the US dollar. The access map, the GPU-hour equivalence, the power estimate and the fine-tuning cost comparison are the author's analysis; the power estimate and cost comparison are illustrative and rest on the stated assumptions. PanScience venture descriptions are company-reported from panscience.xyz.