विमर्श Vimarsh · Arthanīti
India Bought the GPUs. Now Switch Them On.
India has onboarded more than 38,000 subsidised GPUs. Reports say about one in five is in use. Idle compute is the most expensive compute. Here is a ninety-day plan to fill it, and a clock that is already running.
Nikhil Sharma · अर्थनीति Arthanīti · 3 November 2026 · 4 min read
I work with founders who buy AI compute. Every figure is sourced at the end. Utilisation figures are reported, not official.
India did the hard part. It bought the compute. Now it has to switch it on.
Building capacity makes headlines. Using it makes a country. On India’s most important AI asset the second half of that sentence is not yet true.
संख्या · What the numbers say
India has onboarded 38,231 GPUs through the IndiaAI common compute facility. Startups and academia pay about ₹65 per GPU hour. The Mission’s outlay is ₹10,372 crore. Minister of State Jitin Prasada gave these figures in the Lok Sabha in March 2026 along with 190 approved projects (Lok Sabha reply).
Use is the problem. The Economic Times reported in March that only about a fifth of the GPUs on the IndiaAI Compute Portal were in use (ET via TelecomLive). In October Rhodium Group reported that nearly 80% of the Mission’s GPUs sit unused despite below-market prices (Rhodium Group, Elets CIO). These are reported figures. The government has not published a live utilisation number.
कारण · Why subsidised compute sits idle
The reasons are ordinary. That is good news because ordinary problems have ordinary fixes. The March report blamed long deployment, slow allocation and the finances of the companies meant to use it (ET via TelecomLive).
Business Standard’s June survey of users added detail (Business Standard). Short contract cycles and pricing uncertainty worry startups. A seven-day lease cap makes fine-tuning risky. Vouchers usually cover 40–60% of the cost so the user still needs cash. More than 60% of startups asking for access want inference while much of the Mission was designed around training.
Capacity without demand is not sovereignty. It is a warehouse.
अलार्म · The ninety-day alarm
An alarm is useful only if it has a time on it. So here is one. Ninety days from the day this was published, to 1 February 2027. That is enough time to change rules, not enough to build anything new. Everything below can be done with the GPUs India already has.
मार्ग · Five fixes for ninety days
1. Open credits to every registered startup, college and MSME. A DPIIT-recognised startup, an AICTE college or an Udyam-registered MSME should get a starter credit by default. Demand that has to apply is demand that gives up.
2. One-click access. UPI-simple. No paperwork. Sign in with an existing identity. Pick a GPU. Pay from the credit or by UPI for the rest. If a student can buy a train ticket in a minute they should be able to rent an hour of compute in a minute.
3. Move government AI workloads home first. Ministries, PSUs and state governments run models too. Every rupee of that work should land on Indian GPUs before it lands abroad. In 2025 the portal already routed allocations through Yotta, E2E Networks and NxtGen (Inc42). Government demand can anchor them.
4. Publish a live utilisation dashboard. GPUs onboarded, GPUs in use, hours sold, by provider and by user type, updated daily. What gets measured in public gets fixed in public. It also ends the argument over whether the 20% figure is right.
5. Pay providers on time. Providers who are paid late price in the risk or leave. Longer leases for fine-tuning and multi-year commitments for model builders belong here too. Predictable money is cheaper compute.
None of this is new thinking. The Minister himself framed Mission 2.0 around diffusion to MSMEs and a UPI-style playbook (TelecomLive). The point is the clock.
श्रेय · Credit where due
India set out to buy 10,000 GPUs and has onboarded more than 38,000 at rates well below global markets (Business Standard). In February the Minister announced 20,000 more (Dataquest). Few countries have put sovereign compute in reach of students at ₹65 an hour. The Mission now has a new chief executive in Saurabh Vijay who also leads UIDAI (AIM). The team that built Aadhaar at population scale knows how to drive adoption.
Rhodium’s own report found a bright spot in services. Indian engineers close the gap between a model and a working business ([Samachar](/samachar)). That is exactly the demand that should be filling these GPUs.
प्रश्न · Who moves first
The Prime Minister’s Office set the ambition. MeitY and IndiaAI hold the levers. The GPU providers hold the capacity. India’s model builders, startups, colleges and MSMEs hold the demand. Any one of them can move first. The alarm is for all of them.
मेरा दांव · My wager
By 30 June 2027 an official figure from MeitY, IndiaAI or a reply in Parliament will put utilisation of IndiaAI compute portal GPUs at 50% or more. If it does not, India will still own the warehouse and rent the future.
What I am not saying
I am not saying the Mission failed. Building 38,000 GPUs of shared capacity in two years is a real achievement.
I am not saying the 20% figure is final. It is reported by a newspaper and a research firm. That is why the first fix is a public dashboard.
Gurugram, 3 November 2026.
Sources, checked 3 November 2026. GPUs onboarded, price, outlay and projects: Lok Sabha reply by Jitin Prasada, 25 March 2026, via NewKerala. One-fifth utilisation and its causes: The Economic Times via TelecomLive, 4 March 2026. Nearly 80% unused: Rhodium Group, 27 October 2026 and Elets CIO. Lease cap, vouchers, inference demand and the original 10,000 target: Business Standard, 30 June 2026. 20,000 more GPUs: Dataquest, 18 February 2026. Mission 2.0: TelecomLive, 18 February 2026. Provider allocations: Inc42, April 2025. IndiaAI chief executive: Analytics India Magazine, 1 June 2026.