The Billion-Person Algorithm

A scene from a fabric market in Surat, before getting to the argument

Imagine Meera, a small-time saree shopkeeper in Surat's textile market. It is the week before the wedding season and Meera is staring at her spreadsheets — last year's sales figures, this year's advance bookings, a supplier who has been quoting her three different rates in three different conversations. She needs to know how much stock to order before her working capital runs out, but she has no one to ask who actually knows the market besides her.

She pulls out her phone and types a query in Gujarati: “Aa season ma kevu stock rakhvu joie?” — how much stock should I keep for the season?

Within seconds she has an answer — a demand forecast, derived from thousands of queries from other small-time shopkeepers, extrapolated to apply to her own mix of products, factoring in the local wedding season and the supplier's known inclination to hike prices at this time of the year.

Nothing about this requires Meera to know English, or to know what a neural network is, or to be able to afford a consultant who could have helped her. What it does require is a system that has learned — from thousands of shopkeepers who have asked similar questions with similar spreadsheets — what Meera's question actually means.

This system does not exist yet. Everything necessary to build it does — and this is an article about how we could build it, and why it should not look much like anything being built anywhere else in the world.

The core argument, stated plainly

India keeps trying to participate in the AI race by asking the wrong question. The wrong question is “how can we catch up to the countries that have already built AI capabilities?” It is the wrong question because it presumes, incorrectly, that we are on the same playing field as the US, China and the other early adopters. It is a question about catching up to countries that have more capital, more compute and more infrastructure than anyone else — and it therefore assumes that we should be thinking about how to replicate their existing approaches.

The right question is actually “what do we have that they don't?” The answer to that question is people — not people as a resource to be exploited, but people as a source of diversity of lived experience and of problems that no one else has thought about solving yet.

The rest of this article is about unpacking this idea — about exploring what it would mean to think of people as more than a resource, and how an AI-assisted world might be built around serving their needs rather than being built to extract from them. Before getting to that, there is one critical distinction to be made, which is also easy to conflate with the wrong question.

Scale is not the same as surveillance

When people hear “India should turn its population into an asset”, the reaction that often follows is a reasonable and valid concern: “does that mean we want to know everything about everyone?”

It does not — and it shouldn't — but the reason for that is not obvious at first. There is a difference between a government that wants to know everything about everyone and a government that wants to know enough about enough people to build useful models that can help solve real-world problems. The former is a surveillance state — and, ironically, it may also be an innovation desert, because there is no incentive to let innovation take root if it can all be controlled from the top. The latter is something akin to a public good — roads, rail, electricity — but made of information and built by many people, not just one.

The goal should not be to know everything about everyone. The goal should be to treat the experience of a billion people — in all its messy, inconvenient, wonderful variety — as something that can be turned into knowledge that makes their lives better. The consent of the governed should be a prerequisite to that process, as should be the transparency of the methods used — without those things, this whole endeavour is another form of surveillance that India would be better off avoiding.

With that said, here are a few reasons why this sort of approach should be desirable — if not inevitable — in India.

Chapter one: from one complaint to a pattern

Why a single data point is not much, but a million of them can be.

Consider the value of a single data point. One shopkeeper has miscalculated her inventory for the season. One citizen has filed a complaint about a road. One student is struggling with a lesson and has not understood the explanation.

On their own they are not particularly valuable — they simply exist. No one has acted on them beyond perhaps a customer service rep acknowledging the complaint or a teacher spending a minute flagging the student for extra help. They are incidents, and as such they are not likely to lead to much beyond their immediate effect.

Now imagine incidents — not identical, but similar — occurring thousands of times over. The patterns they form can be more valuable: they can be used to predict which categories of goods are most likely to sell in a given season, or what time of year supplier prices tend to peak. They can be used to determine which roads are most likely to need repair, or which areas to prioritize for maintenance. They can be used to identify weaknesses in a curriculum, or what sort of explanation works best for a given category of students.

This is the value of building models from patterns. The reason such patterns are especially valuable to build in India is the diversity of experiences they are likely to be built from. One shopkeeper's inventory might be more informative in isolation than another's, but in combination with similar experiences from other shopkeepers they can begin to reflect an underlying reality. The same applies to all manner of problems and their solutions.

It is crucial to remember that no individual's experience is being mined for these patterns — each one is simply being used as part of a larger set. That is the point of using scale: to get beyond the vagaries of any single experience and identify something more universal. It is also the reason such an approach would work best in India — because those universal patterns are more likely to emerge from the variety of experiences that India has to offer. A pattern from one region and one Indian season will not be much use in another region and another season, but together they can begin to reflect something akin to a national-level pattern that can be useful to many.

Chapter two: why it has to be for everyone, not the few

The danger of an AI-assisted future in India is that it will be dominated by a few large companies and institutions.

There is no question that such an outcome would be preferable to the alternative — an AI-assisted future dominated by large companies and institutions elsewhere. It is still an undesirable outcome, and one that should be actively resisted. The reason is simple: the ability to solve problems at scale should not be a privilege of the few — it should be available to anyone who has the inclination and the ability to use it.

That means an AI-assisted future in India should be decentralized, at least in its ability to participate in and benefit from it. It means a shopkeeper should be able to use an AI-assisted tool without having to understand much about how it works, and a student should be able to build something with it without having to understand much about how it works. It means a small-town startup should be able to rent the infrastructure it cannot afford to build, and a local government office should be able to deploy working solutions to its backlog of problems without outsourcing much of its operations to consultancies. It means researchers should be able to experiment on problems using data without having to prove their institutional affiliations first, and citizens should be able to contribute to those problems without having to understand how they are being solved.

This is not only a matter of fairness — though that alone would be reason enough, if only as a counterweight to other trends. It is also a matter of practicality: the moment an AI-assisted future becomes the domain of the few, it becomes harder to solve problems at scale, because those few have much less incentive to distribute their capabilities freely. That is where the value of diversity begins to intersect with the value of scale — and why an AI-assisted future in India should be built to benefit from both.

Chapter three: every sector is sitting on a data set

How to think about AI as something that underlies everything, rather than as an industry.

One of the easiest mistakes to make when thinking about AI is to mistake it for an industry. It is not — not yet — and it is much more akin to electricity or the internet, which underlie industries rather than being one themselves.

Every industry in India — and indeed every individual — is sitting on top of a data set. Some of these data sets are obvious: they exist in the form of problems being reported and attempts being made to solve them. Others are latent: they exist as patterns in how problems are solved or avoided, visible only to people who know where to look.

That does not mean every one of these data sets is worth building on — in fact, most probably are not. What it does mean is that the first step towards building something is recognizing that these data sets exist, and beginning the process of documenting and formalizing them. It means building an honest, updated record of the problems that people encounter — how often, where and with what effect they occur. It means building a map of what is happening, where and to whom. From there, it becomes much easier to think about what might need to be done about it.

This distinction matters because it is the difference between building something that only serves to document problems and building something that can actually help solve them. It is also the reason this approach should be more viable in India than anywhere else: no other country has such a diverse, distributed set of problems, and therefore no other country has an equivalent set of potential data sets to document and formalize.

Chapter four: sovereignty is not about an Indian model

The layers of sovereignty that are rarely discussed when talking about “our own model”.

When people talk about “sovereign AI” in India, they often mean one specific thing — whether or not India will be able to build its own foundational model. This is a real consideration: it is difficult to have true AI sovereignty if one has to rely on other countries to host one's large language models. But it is only one layer of a much larger set of considerations.

Sovereignty is much more than a foundational model — it spans compute, data, models, people, applications, research and even standards and governance. A country can have a model but not the compute to run it, the data to train it or the people to operate it. It can have all three and lack the applications to make any of it useful to everyday citizens. It can have everything and lack the research and development to improve upon it, or the governance to ensure that it is being used responsibly.

There is a layer of sovereignty that is even more difficult to discuss: the ability to understand one's own country. A model that has been trained on data from around the world can be incredibly valuable, yet know very little about any one country in particular — and therefore have little ability to respond to the specific needs of any one group of citizens. It may be able to provide advice, but it will have much more difficulty providing specific solutions to local problems or understanding the nuances of local culture and context. It is much easier to have an AI that is available in India than one that understands India — and only the latter is truly sovereign.

Chapter five: the danger of arriving at someone else's race

Why copying the strategies of other countries will always put India behind them.

This may be the most difficult argument of all to accept, but it is also the most important — India cannot win the AI race by following everyone else's lead. It does not help to look to other countries to see what we should do — not because they are doing everything wrong, but because they are arriving at many of the same ideas from different directions, and it is much more difficult to follow than to lead.

If the goal is to rebuild an existing platform with a flag next to it, then the race has already been lost — India would be following in the footsteps of others, building on their infrastructure, and would therefore always be playing catch-up. If the goal is to build models on the same assumptions, the same data sets and the same design principles as everyone else, then the best that can be hoped for is competence. It is far more difficult to build something truly novel when everyone has already built something similar.

The more useful question is not “how can we do what they are doing?” but “what can we do that they cannot?” And for India, that list is not as short as one might think. It includes everything from scale and diversity to engineering talent and a unique set of problems that the rest of the world has not had to solve yet. It includes a set of constraints that have led India to develop entirely new approaches to many different problems, as well as a much larger set of languages to work with and a much larger set of people who are used to working within those constraints.

Constraints become design features. Diversity becomes a training set. A billion people become a billion participants rather than a billion potential customers. Problems become data sets. Data sets become models. Models become platforms that other countries will have to study and attempt to replicate if they want to keep up.

Closing: to Surat

We opened this essay with Meera, a small-time saree shopkeeper in Surat who asked a question in Gujarati that an AI was able to answer in the same language. Let us close this essay by talking about what it would take to get her the answer she needs.

It would take treating her question as data that can be used to train future models, rather than as an inconvenient interruption to someone's day. It would take building the infrastructure to actually collect that data — her question, and the similar questions that other shopkeepers are likely to ask — and using that data to build something that can be useful to everyone. It would take resisting the temptation to simply license someone else's model and calling that a sovereign AI capability. It would take a government, a research community, a startup ecosystem and ordinary citizens who all believe that it is worth creating a map of India's problems before anyone tries to solve them.

It would take accepting that none of these things are easy, and that most of them do not exist yet. But it is also true that none of these things exist anywhere else — which is exactly why they are worth pursuing in India.

The AI race is still being run. The outcome is still being decided. It will not necessarily be decided by the country with the most capital or the most compute — it will be decided by the country that has learned to combine what it has in unique ways and to apply those combinations to new problems before anyone else has had the chance to copy them.

Bharat does not need to catch up to anyone — Bharat needs to build the thing only Bharat could have built, in the place only Bharat occupies.

Written August 2026. The strategic environment in which this article is written — in terms of compute availability, data governance and the overall development of the Indian AI ecosystem — is shifting rapidly. Many of the recommendations detailed in this article may well change in the years to come. That is not a criticism of the argument; rather, it is an observation that the map is being drawn as we walk the path.