There is a specific kind of silence you learn in India as a boy — the silence of knowing exactly what you need but not being able to reach it. You could see the library across town. The books were real. But the bus ran twice a day, the road flooded in monsoon, and by the time you arrived the librarian had gone home. The knowledge existed. The infrastructure to access it did not.
I left India carrying that lesson like a stone in my pocket. I came to America, built a career in data, built a company — FedMine — on the premise that the most valuable thing in any market is not the product being traded but the information gap surrounding it. I sold the gap. And for many years I believed that infrastructure was destiny: you could have all the talent in the world, but without the physical machinery to process and distribute what you know, you were still that boy standing at the bus stop with a library across town.
I no longer have to imagine what India's AI future looks like. The data is in. And it looks exactly like that bus stop.
The Numbers That Should Alarm Everyone
Here is the structural fact that should be alarming every policymaker in New Delhi and every technology investor in Palo Alto: India today generates approximately 20 percent of the world's AI training data. It has roughly 210 million active AI users — more than the entire population of Brazil. Indian users send an estimated 620 million consumer prompts per day, edging past the United States' 520 million.[1] By the only measure that feels like power — engagement, adoption, raw signal — India is a giant.
But India owns approximately 3.5 percent of global AI compute infrastructure. Roughly 38,000 high-end GPUs sit on Indian soil, against an estimated 1.2 million in the United States and 450,000 in China. Total data center power capacity in India runs to about 1.8 gigawatts. The United States runs more than 14 gigawatts.[2] The asymmetry ratio — data generated relative to compute owned — sits at roughly 5.7 to one. India generates nearly six times as much AI fuel as it has refinery capacity to process.
| Country | Active AI Users | Daily Prompts | Global Data Share | Compute Share |
|---|---|---|---|---|
| India | 210M | 620M | ~20% | ~3.5% |
| United States | 180M | 520M | ~16.5% | ~54% |
| China | 290M | 810M | ~18% | ~13.5% |
Track GPU rankings, data center capacity, and compute power by country — the live infrastructure numbers behind this asymmetry.
In any other industry, we would call this a structural crisis. In AI, we mostly call it a market opportunity. That distinction is worth examining carefully.
The Hardware Problem in One Sentence
A single NVIDIA Blackwell B200 NVL72 rack draws 120 kilowatts of power.[3] India's entire data center sector draws 1.8 gigawatts. You could fit fifteen thousand Blackwell racks into India's current power envelope — and they would represent the country's entire data center footprint. In practice, a meaningful AI training cluster requires hundreds of such racks running continuously. The arithmetic closes quickly and badly.
The power grid compounds this in ways that are easy to underestimate. India's grid runs approximately 70 percent on coal, and while the country has made genuine strides on renewable capacity, the intermittency and reliability profile of its current grid creates real friction for the kind of always-on, power-hungry AI training workloads that define frontier model development.[4] Indian operators pay between nine and eleven cents per kilowatt-hour for grid power — competitive by global standards — but the carbon intensity and reliability questions create both technical and reputational friction for attracting global AI workloads at scale.
The IndiaAI Mission has recognized this clearly. The government initiative subsidizes GPU compute access at roughly $0.67 to $0.95 per GPU-hour — approximately one-third of what global hyperscalers charge — and aims to build a shared AI compute facility that smaller Indian AI companies can rent by the hour.[5] It is a serious effort from a government that is thinking seriously about the problem. It is also, measured against the scale of the asymmetry, a pilot program trying to fill the Ganges with a garden hose.
The Architecture Nobody Planned
There is a wrinkle that makes India's position stranger and more interesting than the raw numbers suggest. India did not just miss the PC era. It skipped it deliberately, elegantly, and at a scale no technology adoption curve had ever seen. Today, approximately 85 percent of Indian AI users are mobile-only.[1] They never owned a desktop. They jumped from a feature phone directly to a large language model running in a browser tab. This is not a poverty story. It is an architecture story.
Mobile-first adoption shapes everything about how AI gets used and, more importantly, what kind of data it generates. Mobile users tend toward shorter interactions, voice-heavy queries, and consumption-oriented rather than creation-oriented AI tasks. A farmer in Telangana asking an AI chatbot about soil pH via voice in Telugu is doing something categorically different from a software engineer in San Francisco asking a frontier model to debug a TypeScript function — and not just culturally. The inference workload, the training signal, and the economic value each user represents to the model provider are structured very differently.
This matters because the 620 million daily prompts from India are not 620 million uniform tokens flowing into a neutral global dataset. They are language data — Hindi, Telugu, Tamil, Bengali, Marathi, Kannada, Gujarati, Malayalam. They are domain data: agriculture, healthcare, education, micro-commerce, religious practice. They are vernacular data of extraordinary richness, currently underweighted in the English-dominant models that the vast majority of those 620 million users are querying.
India generates six times as much AI fuel as it has capacity to refine. The data leaves. The intelligence comes back. The value created stays abroad.
The Colony That Feeds the Empire
Here is the part of this story that keeps me up at night. Not with despair — I have been kept up at night by worse and come through it — but with a specific, sober alertness that I associate with recognizing a pattern I have seen before.
Every one of those 620 million daily prompts flows outward — to servers in Oregon, Virginia, Singapore, and Ireland. The data leaves India. The intelligence comes back. The value created by Indian users accumulates in model weights, training sets, and balance sheets of American and, to a lesser degree, Chinese technology companies. India does not receive a royalty. There is no transfer pricing mechanism for training data. The token flows one direction; the profit flows another.
I have been calling this algorithmic colonization, and I want to be precise about what I mean by that phrase, because precision matters and careless analogies do real damage. I do not mean that the AI companies are malicious actors. The free models, the consumer applications, the API access — they are genuinely useful in India, often transformatively so. I mean something structural: the default architecture of the current AI ecosystem extracts data from countries that generate it and processes that data in countries that own the compute, and this arrangement tends to compound over time because the companies that train on your data get smarter and more competitive, while you remain a consumer of the intelligence your own data helped create.
Access to the most capable frontier models — the ones that can actually reason, plan, and generate reliably at the level businesses require — carries a dollar cost that scales differently in Mumbai than in Manhattan. At a developer median salary of $10,000–$15,000 annually in India, a $20 monthly subscription to a frontier model represents a meaningful tax on building AI-native products. The token limits, latency tolerances, and regional pricing tiers built into most frontier model APIs are not designed to be discriminatory. But they operate, in practice, as a soft gate on who gets to build with the best tools — and who remains a user of what others built with them.[6]
What This Looks Like Through a Singularity Lens
I am writing a book about the singularity — not as a science fiction concept but as a live, ongoing process of compression and acceleration in which human history is folding back on itself at a pace that makes the usual categories of progress and lag dissolve. The thesis, which I am becoming more certain of with each passing month, is that the singularity is not coming. It is here.
But here, it turns out, is not everywhere at once.
The structural AI asymmetry between India and the compute-rich world is a singularity problem. The technology is accelerating, the gains from acceleration are compounding, and the compounding is not neutral — it favors the countries that own the refineries. India can close this gap; the IndiaAI Mission, the domestic AI startup ecosystem, the extraordinary talent pipeline from the IITs and IIMs, the sheer scale of India's domestic market — these are real assets that no other developing country can match. But the gap is not static. It is growing. While India is building 1,000-GPU clusters, the United States is deploying 100,000-GPU clusters. The ratio does not improve automatically. It requires deliberate policy, deliberate investment, and deliberate choices about which AI workloads to localize and which language and domain models to build domestically before the window for doing so narrows further.
The boy at the bus stop eventually gets a motorcycle. But if everyone else in the race just got a helicopter, the metaphor starts to work against you.
Where This Ends — Or Doesn't
I left India a long time ago. I found my library. I built the data company. I learned what infrastructure does to talent when the two are matched — and what happens to talent when infrastructure never arrives. The gap between the two is not a measure of ambition. It is a measure of what came before, who built it, and who it was built to serve.
What I see in the India AI story is not a failure of ambition. The ambition is there — electric, tangible, the kind that fills rooms at Mumbai startup events and IIT convocations and the corridors of MeitY in New Delhi. What I see is an infrastructure debt that is compounding at roughly the same rate as the technology it cannot yet run.
The 620 million prompts will keep flowing regardless. The question is whether, ten years from now, they will be flowing to Indian data centers running Indian-trained models generating intelligence that stays in India — or whether the Ganges will still be draining into someone else's ocean.
That is not a rhetorical question. It is a policy question, an investment question, and an engineering question, and the window for answering it well is measurable in years, not decades. The singularity does not wait for infrastructure plans to mature. It compounds around them.