Deployment over frontier: how India should measure its AI progress

Most conversations about India and artificial intelligence begin with a scoreboard. Who has the largest model, the most GPUs, the biggest training run. On that scoreboard India looks behind, and the usual prescription follows: build a frontier model of our own.

I think the scoreboard is wrong. A frontier model is an input. What a country gets from AI depends on where that input ends up: in a district hospital's triage, a farmer's credit decision, a teacher's lesson plan, a tax officer's case file. That is deployment, and it is where India has a real head start.

What we already know how to do

India has built digital systems that serve more than a billion people. UPI did not win because India invented the fastest payment protocol. It won because the rails were open, the cost to participate was low, and banks, apps and merchants could all plug in. Aadhaar and DigiLocker followed the same pattern. The skill on display is not research. It is putting technology into daily use at a scale few countries attempt.

AI should be judged the same way. A model that is good enough, cheap to run and available in twenty languages will do more for India than a model that tops a benchmark and stays in a lab.

Where deployment actually breaks

At Nuclear Edge we spend most of our time on the physical side of technology: tablets for field executives, panels in government classrooms, surveillance across railway stations. That work teaches a plain lesson. Programmes rarely fail because of the software. They fail at the last mile.

• The device arrives but nobody is trained to use it.
• The school has the panel but the power is unstable.
• The branch has the system but no engineer within a day's travel.

AI will meet the same obstacles, only faster. If we want AI to change how a block office or a primary health centre works, we need to plan for devices, connectivity, local-language interfaces, support contracts and training as seriously as we plan for compute.

A different set of measures

In our working paper, Deployment over Frontier, we propose that India track a small set of public measures:

1. The number of government services where an AI system assists a decision, and the share of citizens those services reach.
2. The cost per AI-assisted transaction, so that programmes are judged on affordability as well as accuracy.
3. Coverage in Indian languages, measured service by service rather than model by model.
4. The time from a successful pilot to state-wide rollout.

None of these require India to win the frontier race. All of them require India to be very good at something it already does well.

What this means for industry

For Indian companies, the opportunity is not only to build models but to become the best in the world at putting them to work. That means systems integration, field operations and long-term support, the unglamorous work that turns a pilot into a programme. It is the work we have done for years with classrooms, branches and substations, and it is the work AI now needs.

India does not have to be first to the frontier. It has to be first to the people.