DM Explains · No. 018 · Sovereign AI
Sovereign compute: why India wants to own the GPUs, not just rent them
India is building a government-controlled pool of GPUs so startups and researchers aren't entirely dependent on foreign clouds — but the chips inside that pool are still imported.
· 12 min read
Read me first
India's promise of 100,000 GPUs is real as a target and partly real as hardware, but the pool that startups can actually rent remains smaller than every headline number suggests; renting beats owning only once the queue to use it is short enough to matter.
The IndiaAI Compute Portal — a startup requests subsidised GPU-hours → 1 draws from the ₹65/hr pool across 15 empanelled providers → 2 unmet demand triggers tenders for more capacity → 3 C-DAC and UAE-backed clusters add government-owned GPUs when tenders lag → 4 the stated target climbs toward 100,000 GPUs by December 2026 → 5 usage is logged against 93.18 lakh GPU-hours already drawn by 237 projects as of August 2026
A startup founder in Bengaluru reads that India now has 58,000 GPUs for AI, then tries to book time on the national compute pool and finds a waitlist. Both things are true. The government has announced, tendered and built real GPU capacity through the IndiaAI Mission, but the number quoted in a press release and the number sitting idle and bookable on a given Tuesday are not the same number, and the gap between them is where most of the actual story lives.
Everyday picture. Amma tells the family she's "got the kitchen stocked for the month" after placing an order with three different vendors. Papa checks the fridge and finds half the order hasn't arrived yet, some of it arrived late and spoiled, and what's actually there is enough for maybe two weeks, not four.
Where the picture breaks. Amma's shopping list is fixed once written. The government's GPU targets keep being revised upward mid-month, with new numbers announced before the previous promise has even been delivered.
Why it exists now
Training and running large AI models needs GPUs, and GPUs are expensive, scarce and mostly sold by a small number of foreign chipmakers and cloud providers. A startup or university lab in India that wants to build its own models either pays international cloud prices in dollars or does without. The IndiaAI Mission, approved in March 2024 with an outlay of ₹10,371.92 crore over five years, exists to close that gap by renting out GPU capacity to Indian researchers and companies at subsidised rates, treating compute as a utility the state helps provision rather than a luxury only well-funded firms can buy (One Big Future, 2026).
What changed to put this at the centre of attention is the pace of upward revision. At the India AI Impact Summit in February 2026, the government announced it would add 20,000 GPUs beyond its existing 38,000, in the coming weeks (Digital India, MeitY, 2026). IT Minister Ashwini Vaishnaw separately said more than 50,000 GPUs would be deployed within six months, more than doubling the then-current base of 38,000 (Times of India, 2026). IndiaAI Mission CEO Abhishek Singh went further, suggesting capacity may triple from 38,000 to roughly 100,000 by the end of 2026 (Indian Express, 2026). None of these numbers contradict each other exactly, but they describe different moments in a fast-moving build-out, which is exactly why anyone trying to use the pool needs to separate the target from the delivered capacity.
How it works
By August 2026, 237 projects had drawn 93.18 lakh GPU-hours from the IndiaAI Compute Portal — against a mission that promised 58,000 GPUs, scaled to 100,000, but whose shared pool reached only around 30,000 GPUs against a 45,000-GPU target.
Subsidised pool
Rent compute like a utility, not a luxury
What it is. A shared stock of GPUs, listed on the IndiaAI Compute Portal, that approved projects can book by the hour at a government-set price well below open-market cloud rates.
Why it exists. Most Indian AI startups and research labs cannot afford sustained access to high-end GPUs at international prices; a subsidised pool lets them train and test models without that upfront barrier.
Where it breaks. A shared pool has a queue. If demand outruns supply, "subsidised" access means waiting, not instant access.
How that's handled. More than 38,000 high-end GPUs have been made available to startups, researchers and public institutions at ₹65 per hour (PIB, 2026).
In technical terms
In technical terms: the portal allocates GPU-hours against project applications, with usage metered and billed per hour rather than leased as dedicated hardware.
Tendered capacity
Ask the market to build what the state can't
What it is. The government invites private compute providers to bid for contracts to supply GPU capacity into the national pool, rather than building every data centre itself.
Why it exists. Building and operating GPU clusters at scale is capital-intensive and specialised; tendering lets the state pay for capacity as a service from companies that already run it.
Where it breaks. Providers can pledge capacity they then struggle to deliver. The common compute pool reached only around 30,000 GPUs against a targeted 45,000, a roughly 15,000-GPU shortfall, as private providers struggled to deliver pledged capacity amid rising equipment costs (The Hindu BusinessLine, 2026).
How that's handled. By August 2026, 15 compute service providers had been empanelled under the IndiaAI Mission to supply the pool, spreading delivery risk across more than one vendor (One Big Future, 2026).
In technical terms
In technical terms: empanelment is a formal approval process; an empanelled provider has passed government vetting and sits on an approved supplier list before any GPU-hour it offers can enter the public pool.
Government clusters
If renting fails, just build your own
What it is. Alongside rented and tendered capacity, the state is pursuing GPU infrastructure it owns outright or receives as a government-to-government arrangement.
Why it exists. Relying only on private providers leaves delivery exposed to their costs, supply chains and willingness to build. Owned capacity is a hedge against exactly the shortfall described above.
Where it breaks. Government-run clusters take longer to stand up than a tender, and a single large project concentrates risk rather than spreading it.
How that's handled. The government is working with C-DAC to install 3,000 GPUs of sovereign compute capacity, and is finalising an 8-exaflop AI supercomputer, equivalent to about 10,000 GPUs of computing capacity, with the UAE via G42 and Cerebras in partnership with C-DAC, to be owned by the Indian government as a gift (The Hindu BusinessLine, 2026).
In technical terms
In technical terms: an exaflop measures one quintillion floating-point calculations per second; translating that into "GPU-equivalent" capacity is an approximation used to compare supercomputers against standard GPU clusters.
Scaling targets
The number that keeps climbing before it's real
What it is. Public figures stating where total national GPU capacity is headed, announced well ahead of the hardware existing.
Why it exists. Targets signal intent to investors, providers and other governments, and create political pressure to deliver.
Where it breaks. Different officials and outlets have cited different totals in the same window: 58,000 GPUs reported following the February 2026 summit (ETGovernment, 2026), a plan to reach 100,000 GPUs by December 2026 up from shared capacity that had crossed 45,000 by June 2026 (One Big Future, 2026), and the triple-to-100,000 figure from the Mission's own CEO (Indian Express, 2026).
How that's handled. Treat any single total as a snapshot of an announcement, not a verified inventory, and check the date attached to it.
In technical terms
In technical terms: none of these figures are independently audited hardware counts; they are self-reported totals from government officials and ministries at specific announcement dates.
Usage tracking
A pool only matters if someone's drawing from it
What it is. Records of how many projects have actually accessed the subsidised pool and how many GPU-hours they consumed.
Why it exists. Capacity that exists but isn't booked tells you nothing about whether the scheme is working for the people it's meant to serve.
Where it breaks. Usage figures lag behind capacity announcements by months, so a researcher reading a hardware number has no way to tell how much of it is actually reachable today.
How that's handled. By August 2026, 237 projects had accessed subsidised computing through the IndiaAI programme, totalling 93.18 lakh GPU-hours across model development, training, testing and research (One Big Future, 2026).
In technical terms
In technical terms: a GPU-hour is the metering unit, one GPU occupied for one hour, that converts raw hardware counts into comparable usage and billing figures.
Compared with what came before
| Before IndiaAI Mission | With the sovereign compute pool | |
|---|---|---|
| Who provides GPUs | Mostly foreign hyperscalers, billed internationally | Government-subsidised pool plus empanelled Indian and foreign providers |
| Price | Open international cloud rates | ₹65 per hour for high-end GPUs, as of February 2026 |
| Who decides allocation | Whoever can pay | Government-approved projects, via the IndiaAI Compute Portal |
| Delivery certainty | Market-driven, generally reliable at a price | Target-driven, with gaps between pledged and delivered capacity |
Renting from the sovereign pool is cheaper and more aligned with national priorities than paying hyperscaler rates, but it trades price for queueing and for a less certain delivery timeline than a straightforward commercial contract.
Where it breaks
Pledged capacity is not delivered capacity
The common compute pool reached only around 30,000 GPUs against a targeted 45,000, as private providers struggled to deliver amid rising equipment costs. A headline total should never be read as hardware sitting ready to book.
The scheme works well when a project's need matches what's actually online, not what's been announced. It breaks when the gap between pledge and delivery stretches a research timeline past its deadline, or when pricing for newly added capacity is still unsettled: the government has not yet decided on pricing for the new sovereign GPU pool and plans to float bids after finalising agreements, a process expected to take three to four months. It also breaks on provider concentration: the IndiaAI Compute Portal lists accelerators from Nvidia, AMD, Intel and Google among its suppliers, which spreads some risk, but a shortfall from any single large supplier still shows up as a dent in the national total.
Numbers to carry
| Figure | Value |
|---|---|
| IndiaAI Mission outlay | ₹10,371.92 crore over five years, approved March 2024 |
| Subsidised GPU price | ₹65 per hour, as of February 2026 |
| Existing base before latest addition | 38,000 GPUs |
| Planned addition announced at the 2026 summit | 20,000 GPUs beyond the existing 38,000, as of February 2026 |
| Minister's six-month target | more than 50,000 GPUs deployed within six months, as of February 2026 |
| CEO's end-2026 target | roughly 100,000 GPUs by end of 2026 |
| Reported total post-summit | 58,000 GPUs, as of February 2026 |
| Shared capacity by mid-2026 | crossed 45,000 GPUs by June 2026 |
| Delivered vs targeted pool | around 30,000 GPUs delivered against a 45,000 target, a roughly 15,000-GPU deficit |
| Sovereign C-DAC installation | 3,000 GPUs |
| UAE supercomputer gift | 8-exaflop, equivalent to about 10,000 GPUs |
| Projects using subsidised compute by August 2026 | 237 projects, 93.18 lakh GPU-hours |
| Empanelled providers by August 2026 | 15 providers |
Myths worth killing
| Commonly repeated | What is actually true |
|---|---|
| "India has 100,000 AI GPUs now" | 100,000 is a target cited for the end of 2026 or by December 2026, not a current inventory. |
| "Every announced GPU is bookable" | The pool fell short of its own 45,000 target, delivering only around 30,000, because providers struggled to deliver pledged capacity. |
| "The subsidised price applies to all new capacity" | Pricing for the newly tendered sovereign pool had not yet been decided as of the 2026 reporting; the ₹65-per-hour rate applies to the existing 38,000-plus GPUs. |
| "This is one government project" | It is a mix of a subsidised rental pool, tendered private capacity, a C-DAC installation and a UAE-gifted supercomputer, each with different delivery mechanics. |
Worked example
- A Bengaluru startup (illustrative) wants to fine-tune a mid-sized model and checks the IndiaAI Compute Portal for availability.
- The founder sees the national total reported at 58,000 GPUs as of February 2026 and assumes capacity is plentiful.
- The application queue reveals that only a portion of that total is live and bookable, consistent with the pool having delivered around 30,000 GPUs against a 45,000 target at an earlier point in 2026.
- The project is accepted into the subsidised pool at ₹65 per hour and joins the 237 projects already drawing on it as of August 2026.
- The founder budgets for GPU-hours the way Amma budgets for groceries bought from a vendor who sometimes delivers late: plan around delay, not around the catalogue price.
Should you use it?
Use the sovereign pool when a project is research, academic or early-stage, timelines have slack for queueing, and ₹65 per hour against actual availability beats open-market cloud pricing. Avoid it when a deadline is fixed and tight, since delivered capacity has repeatedly lagged pledged capacity, or when the workload needs a provider mix or location commercial cloud already guarantees. The cheapest first experiment is checking current slot availability on the portal against the project's actual deadline before committing a research plan to it.
Say it in the room
"Treat every GPU number from this programme as an announcement with a date on it, not an inventory you can draw against today."
Lines to carry
- "The target is not the inventory."
- "Pledged capacity is not delivered capacity."
- "₹65 an hour only helps if there's a slot."
- "Usage figures, not headline totals, tell you if it's working."
- "Every number in this story carries a date."
FAQ
Is the IndiaAI Mission's GPU pool actually cheaper than commercial cloud?
Yes, on price: ₹65 per hour for high-end GPUs is well below typical international cloud rates. Whether it's cheaper in practice depends on how long a project waits for a slot, since queueing time has a cost too.
How many GPUs does India actually have right now?
It depends on which announcement and which date you're reading. As of February 2026, officials cited figures from 38,000 to 58,000 depending on the source and whether planned additions were included; treat any single number as a snapshot, not a running total.
Why did the pool fall short of its own target?
Private providers that had pledged capacity struggled to deliver it amid rising equipment costs, leaving the pool at roughly 30,000 GPUs against a 45,000 target at that point in 2026.
Who actually supplies the GPUs in the pool?
A mix of empanelled private providers, hardware from vendors including Nvidia, AMD, Intel and Google listed on the IndiaAI Compute Portal, a C-DAC-installed sovereign block, and a UAE-gifted supercomputer built with G42 and Cerebras.
Can any startup just sign up and get GPU time?
Access runs through the IndiaAI Compute Portal and requires an approved project application; by August 2026, 237 projects had been accepted and had drawn a combined 93.18 lakh GPU-hours.
Glossary
- GPU (Graphics Processing Unit) — a specialised chip, originally for graphics, now widely used to train and run AI models because it can do many calculations in parallel.
- Sovereign compute — government-owned or government-controlled computing infrastructure, especially GPUs, used for national AI development rather than relying solely on foreign private providers.
- IndiaAI Mission — India's flagship government programme to build AI compute, datasets, skilling and governance capacity, approved in March 2024 with a budget of ₹10,371.92 crore over five years.
- IndiaAI Compute Portal — an online platform through which startups, researchers and institutions in India can request and access subsidised GPU computing capacity.
- Empanelled provider — a private compute company formally approved and listed by the government to supply GPU capacity to the national pool.
- GPU-hour — a unit measuring the use of one GPU for one hour, used to track and bill compute consumption.
- Exaflop — a measure of computing speed equal to one quintillion floating-point calculations per second, often used to describe supercomputer performance.
- C-DAC (Centre for Development of Advanced Computing) — an Indian government research and development organisation involved in building domestic computing infrastructure.
- RFP (Request for Proposal) — a formal government tender document inviting companies to bid to supply goods or services, such as GPU capacity.
- Hyperscaler — a very large cloud computing company that operates vast data centres and sells computing capacity at scale.
- Data localisation — a legal requirement that certain data be stored and processed within a country's borders.
Sources
- India to Add 20,000 GPUs Beyond Existing 38,000 to Strengthen National AI Infrastructure — Digital India, MeitY, Government of India, 2026
- Centre fortifies India's AI mission with Sovereign Compute Pool amid rising GPU costs — The Hindu BusinessLine, 2026
- India's GPU capacity may triple to 100,000 by 2026-end: IndiaAI Mission CEO — The Indian Express, 2026
- Press Release Page — Press Information Bureau, Government of India, 2026
- India Scales AI Compute Infrastructure to 58,000 GPUs: Transforming Global AI Dynamics — ETGovernment, Economic Times, 2026
- India to more than double its GPU capacity in six months, says IT Minister Ashwini Vaishnaw at AI Summit — The Times of India, 2026
- India Wants To Expand Public AI Compute To 100,000 GPUs By December — One Big Future, 2026