India’s AI compute subsidy should be judged by who gets to use it

India has a case for subsidising access to computing power for AI. A university lab or a small firm can have a promising project and too little capital to reserve the machines it needs. The public programme can lower that barrier. At the India AI Impact Summit on 17 February, the government said it would add 20,000 graphics processing units to the 38,000 already available under the initiative. That is a planned increase of about 53%. The test is how many credible projects get usable computing time on terms they could not otherwise obtain. A larger inventory can help. So can allocating the existing machines better. MeitY’s 17 February summit announcement · IndiaAI compute backgrounder, 17 February

Count usable computing hours

The Cabinet approved the ₹10,371.92 crore IndiaAI Mission in March 2024. Its compute pillar envisaged 10,000 or more GPUs through a public-private partnership and a marketplace through which users could obtain computing services. The outlay is for the wider mission, which also covers areas such as skills, data and startup finance. Comparing that entire figure with the GPU count would give a misleading cost per machine. The relevant public cost of the compute service includes the subsidy actually paid for usage and any financial commitment for capacity that remains idle. Those are the figures to compare with the volume and quality of access delivered. Cabinet approval of the IndiaAI Mission, March 2024 · IndiaAI compute backgrounder, 17 February

A count of GPUs treats very different machines as if they were interchangeable. Memory, the connection between processors, and the ability to reserve several together affect which jobs can run. A team adapting an existing model might need a modest amount of computing spread over several weeks. A larger experiment may need a cluster available at the same time. Both can encounter a shortage even when some capacity is idle. The market can have spare time on one type of machine and a waiting list on another. Publish available and used accelerator hours by hardware class, together with waiting times for different reservation sizes. That would reveal a constraint a headline count cannot. IndiaAI compute backgrounder, 17 February · MeitY’s 17 February summit announcement

The government’s summit backgrounder cites ₹65 per GPU-hour for available capacity. Use that as an illustrative rate. A project renting 32 GPUs for 200 hours uses 6,400 GPU-hours and pays ₹4.16 lakh in accelerator charges, before storage, networking, taxes or staff time. If only 70% of those reserved hours produce useful work, the effective accelerator cost is about ₹93 per useful GPU-hour. A team able to interrupt its job or debug on smaller machines has room to reduce that bill. A research group facing frequent failures or forced to book capacity in larger blocks may pay much more per completed experiment. The price schedule matters alongside the nominal hourly rate. IndiaAI compute backgrounder, 17 February

The subsidy should change who can attempt a project

Access can be undersupplied where a useful experiment has benefits that its sponsor cannot fully charge for. A language dataset, a public evaluation benchmark or a tool for a poorly served Indian language can be reused by others. Its developer may struggle to recover the cost through sales alone. A fixed allocation of compute can help a team demonstrate enough progress to attract funding or show that the project is unworkable before it spends more. The first decision should be whether the intervention changes what can be attempted. A discount on jobs that a well-capitalised company would have run anyway mostly transfers money to the user. Cabinet approval of the IndiaAI Mission, March 2024

The strongest case for granting access to larger teams is that some useful work genuinely needs scale. An experienced team can make productive use of a large cluster and produce a tool that many smaller firms use. Splitting every award into tiny allotments could therefore waste the ability to run certain experiments. A marketplace can accommodate both needs. Give newcomers short initial allocations and reserve a pool for larger, time-bounded requests evaluated against published criteria. Releasing follow-on time against verifiable milestones makes it possible to support ambitious projects without making permanent commitments based on the strength of an application alone. Cabinet approval of the IndiaAI Mission, March 2024 · MeitY’s 17 February summit announcement

There is also a case for keeping much of the service commercial. If subsidies are available at a single low price regardless of demand, high-value users can be crowded out by those willing to reserve time they may never need. Giving commercial users a way to pay for priority while protecting access for approved research helps manage that tension. Cancellation rules and scheduling that can reassign unused reservations can recover hours without buying more machines. Competition among providers can improve service if applicants can take their data and workloads to another provider at reasonable cost. Contracts should make portability and billing transparency part of what the programme buys.

Separate infrastructure from the cost of deploying a useful service

The summit discussion has included a common AI platform for small businesses. A shop, exporter or manufacturer will judge such a service by whether it improves a task at a reasonable ongoing cost. Cheap training time can help create a tool, but deployment can still require clean data, integration with existing software and people who can check its outputs. An invoicing tool that needs regular manual correction saves little time. A language interface that works in demonstrations but performs poorly in a local dialect may generate more support calls. The public return depends on how the tool performs outside the team that built it. Evaluation should include usage by firms after any introductory support ends. The Indian Express’s summit coverage · Cabinet approval of the IndiaAI Mission, March 2024

There are costs outside the budget too. Operating capacity requires dependable electricity, network connections and sites able to support cooling. A location with spare grid capacity at relevant hours can accommodate new demand more easily than a constrained location that needs network upgrades. The public compute programme can limit waste by procuring capacity in stages and asking providers to report energy per completed workload where measurement is comparable. The point is to compare different ways of delivering usable capacity, including access to existing machines, before paying for more. Additional demand on a power system should be considered alongside the gains to researchers and firms.

An evaluation can start with the applications themselves. If awards use a published scoring threshold, compare applicants just above and below it whose proposals and funding look similar. Follow both groups over the same period to see whether an award shortened the time to complete an experiment, increased the chance of finishing it or merely replaced a cloud bill the team would have paid anyway. Where access is allocated through other means, a staggered rollout or a lottery among equally ranked applications can provide a useful comparison. Count the costs paid by teams as well as the public subsidy. That would give a more credible estimate of additional work than dividing the total number of awards by the available machines.

Publish evidence on who gains access

The government should report active users and usage by type of organisation, the share going to first-time users, median wait times, used and reserved hours, subsidy per completed project and how concentrated awards are among recipients. Track whether supported projects go on to produce reproducible evaluations, tools that independent users adopt or commercial products that can sustain their own compute bills. An increase in genuine first-time access, with shorter queues and repeat use of the resulting tools, would strengthen the case for expansion. High utilisation dominated by a few teams that already had market access would weaken it. If the constraint turns out to be staff, data or adoption, the next rupee may do more there than on another processor. Cabinet approval of the IndiaAI Mission, March 2024 · IndiaAI compute backgrounder, 17 February