Inferencing at Scale is How AI Starts Becoming Infrastructure 

India’s DPI story has always been misread as a technology story. Aadhaar, UPI, DigiLocker - all of them succeeded because they reduced the cost of coordination at scale. It was never about rails alone, it was about making critical decisions cheap, instantaneous, and universal.

AI in India is headed down the same path. 

In the first episode of our new series, Intelligent Indians! our ongoing conversations with AI builders and policymakers, one theme keeps surfacing: India’s AI moment will be defined by how cheaply and widely intelligence can be deployed. Abhishek Singh, CEO of the IndiaAI Mission framed the next phase with unusual clarity:  

"Building foundation models is just one milestone. Ultimately, those models need to develop into applications which benefit people. And for that, what we need is inferencing at scale."

Put bluntly, it’s not just about smarter models, or pretty demos. It is continuous, low-cost judgement running in the background of everyday systems.

That shift, from episodic intelligence to ambient intelligence, is where DPI^AI, or the idea of layering AI over India’s existing digital rails, becomes truly consequential. Here are five outcomes that fall out of that shift. 
‍

Intelligence becomes ambient 

Today, most business decisions are still event-based: A loan is assessed when someone applies, compliance is checked after something goes wrong, risk is reviewed at quarter-end. 

Inference at scale breaks that rhythm. With low-cost, continuous inference, these checkpoints start dissolving. 

Let’s look at an example of a small manufacturer. Instead of their creditworthiness being assessed once a year, it’s inferred daily: from cash flows, receivables, inventory movement, and payment behaviour. No new application. No new documents. Just a system that’s always updating its view.

The difference here is frequency. When decisions are updated continuously, fewer things escalate into “events” in the first place. 
‍

A Relationship Manager for every Indian, without the economics of one 

For decades, personalised financial advice in India has been gated by economics. Relationship managers existed, but only for a small slice of customers who were profitable enough to justify human time and attention.

DPI^AI makes that interpretative layer possible at population scale. 

As Abhishek Singh noted in our conversation, once inference runs continuously in the background, intelligence stops being something users actively seek and starts showing up when it’s needed. Combined with consented data flows through Account Aggregator rails, this enables personalised, automated guidance rooted in a person’s actual financial context — not generic advice.

This is where the idea of “an RM for every Indian” becomes real. Not as a human substitute or a chatbot bolted onto an app, but as an always-on layer that interprets cash flows, liabilities, and eligibility in real time, and surfaces the right prompts at the right moment. 

The result is broader participation. Financial products reach the mass-affluent more effectively, and credit access expands as AI-led underwriting builds richer profiles for thin-file customers. What changes isn’t just access, but confidence — and that’s what drives adoption.
‍

Trust becomes architectural, not declarative 

As Abhishek Singh pointed out, India’s digital public infrastructure was designed around consent-based data sharing and clearly defined purpose limitation. When AI is layered on top of this foundation, trust doesn’t have to be rebuilt from scratch, it is inherited from the underlying architecture. 

This changes the nature of the problem. Instead of asking users to believe that a system will behave responsibly, the system is constrained by design: what data it can access, for how long, and for what purpose are all predefined. 

For example: 

• A system that only accesses bank data for a specific credit decision, and then lets that permission lapse. 

• An AI agent that explains why it’s asking for additional data, in plain language. 

• A product that intervenes only when confidence crosses a clear threshold. 

As AI becomes more powerful, trust becomes less about persuasion and more about structure, not something products promise, but something they embed by default.
‍

Productivity shows up long before pricing does 

Another important observation from the episode was: 

"Initially, I don’t think there will be a business model in which people are willing to pay for these services."

That’s how many Indian platforms have scaled: UPI didn’t monetise early, and neither did Aadhaar. First they proved their efficiency and then let value accumulate across the ecosystem. 

With AI, we’re likely to see the same pattern: 

• Better credit decisions 

• Fewer compliance errors 

• Faster resolution of routine issues 

This is also why the state is intervening where it is. Through the IndiaAI Mission, the government is effectively absorbing the early cost of experimentation, subsidising compute and making datasets accessible, so builders can focus on proving real-world impact before viable business models fully emerge. 

The intent isn’t to pick winners, but to prevent intelligence from becoming wealth-gated at the outset. By lowering the cost of inference and iteration, the ecosystem gets time to demonstrate usefulness at scale. Monetisation can follow once outcomes are visible and demand is real.
‍‍

The real challenge? Restraint 

As inference gets cheaper, the temptation will be to use it everywhere. But at scale, too much intervention is as risky as too little. 

Think about it, a lending system that constantly changes limits creates confusion. Or a compliance system that flags everything creates noise. And an AI assistant that always responds creates dependency. 

The hardest problem will be using judgement about when intelligence should step in. 

DPI^AI will work in the background, reducing friction, smoothing decisions, preventing issues before they surface.  When intelligence becomes infrastructure, institutions look calmer, more predictable and reliable. 

And historically, those are the systems that scale the farthest. 
‍

‍

Many of these themes come alive in our conversation with Abhishek Singh on Intelligent Indians, where we discuss how DPI ^AI could play out at population scale. 

‍