IIndia's AI Sovereignty Problem

In 1973, the oil crisis taught India a generation-defining lesson: depending on foreign suppliers for the substance that powers your economy is not a commercial decision but a strategic one. Every policy from the strategic petroleum reserve to the LNG long-term contracts to the renewable transition flowed, in part, from that single shock. Half a century later, India is on the edge of an identical lesson with an entirely new substance, intelligence, and almost nobody in the public conversation is treating it with the gravity it deserves.

Intelligence is the new oil, except worse. Oil powers physical work; intelligence powers cognitive work. Oil flows through pipes you can see and police; intelligence flows through APIs you cannot. Oil dependence is a problem that ends at the gas pump; intelligence dependence is a problem that follows you into every email, every decision, every diagnosis, every contract. A country that imports its intelligence at the foundation-model layer is, slowly and quietly, importing the worldview embedded in those models, in every cognitive task they touch.

¶What sovereignty actually means here

Sovereignty in AI is not the question of whether India can train a frontier model that beats GPT on an English benchmark. That is a vanity project that would cost a hundred billion rupees and would not solve sovereignty in any meaningful sense. The frontier model is the wrong unit of analysis, because frontier capability is converging fast across labs and the relative gap between the best closed model and the best open model has shrunk every year for three years running.

The right unit of analysis is something more boring and more important: who controls the inference layer, the evaluation layer, the data pipelines, and the deployment infrastructure for AI used inside India. The frontier model is a fashion. The infrastructure beneath it is the strategic asset. Sovereignty is whether, in 2035, an Indian doctor, lawyer, civil servant, or farmer can do their work without depending on a foreign API call that could be denied, throttled, priced up, or weaponized.

¶Three layers, three different fights

The AI stack, simplified, has three strategic layers. The model layer is where the foundation weights live. The inference layer is where the model is served, at scale, to users. The evaluation layer is where the model's behaviour is measured, audited, and improved. Each layer has different sovereignty stakes.

The model layer is the most expensive and the least strategic per rupee spent. Open weights, Llama, Mistral, Qwen, and the Indian-trained models that will follow, have already collapsed the cost of having a competitive base model. India should participate in this conversation, but the right level of ambition is "useful Indic models," not "GPT-5 parity." Useful Indic models trained on careful Indian data, available openly, used by the country's professional class, is a 100-million-dollar program. Chasing parity with closed labs at the frontier is a 10-billion-dollar program that produces, at the end, a model that gets surpassed within twelve months.

The inference layer is where the actual sovereignty fight is. Who runs the servers that serve the model to the doctor in Indore? Where are the chips? Whose terms of service govern the API? What happens to the queries logged in the inference path? If the inference is happening in a foreign cloud, then every diagnostic conversation between an Indian doctor and her AI assistant is, at the substrate level, observable by people who do not answer to Indian law. This is not paranoia. It is the unavoidable consequence of the supply chain choice.

The evaluation layer is the rarest and most underestimated. A model is, in operation, the sum of its behaviours, not its weights. Whoever can systematically evaluate a model's behaviour can systematically improve it, certify it for use in critical sectors, and detect drift. India has, by some estimates, fewer than five thousand people who can do this work seriously. The country needs fifty thousand. Building this workforce is the cheapest, fastest, and most strategic move available, and it is barely being talked about.

¶What gets imported when you import a model

Importing a frontier model trained primarily on English internet data means importing, by default, the assumptions of that internet. The cultural references that are normalized. The legal frameworks that are assumed. The histories that are taught. The names that are read as professional and the names that are read as suspicious. The medical guidelines that are presented as default. The agricultural advice calibrated for climates and crops different from yours. None of this is malicious. All of it is structural. And once embedded in the tools your professional class uses every day, it becomes the cognitive substrate of a generation.

This is not an argument for cutting off foreign models. It is an argument for making sure the substrate of professional cognition in India is at least co-Indian, with the foreign models in the mix but not the only thing in the mix. The right state is a thriving ecosystem of Indic models, small, large, sector-specific, multilingual, used alongside global models, evaluated by an Indian-built evaluation discipline, deployed on infrastructure that respects Indian law.

¶Sovereignty is built from below

The temptation is to think that sovereignty has to be built by the state, top-down, through grand projects. Some of it does. The compute infrastructure, the procurement preferences, the public funding for evaluation. But most of it is built from below, by communities of practitioners who make the boring choices: which models to evaluate, which datasets to curate, which deployments to standardize. The community of Indian AI evaluators, doctors using AI, lawyers using AI, teachers using AI, they are the actual sovereignty layer. The government can set the conditions. The community has to do the work.

This is why Bharath.CLUB and AI.Bharath.CLUB exist. The community is the missing layer. Sarasvat.ai works on wisdom-first models. Eval.qa works on evaluation. The community works on adoption, criticism, and trust. None of these alone is sufficient. Together, they are the beginning of a sovereign stack.

¶The window is narrow

The window for getting this right is not a generation. It is the next three to five years. After that, the cognitive substrate of Indian professional life will have settled into whatever shape it settles into, and changing it later will require effort proportional to the entrenchment. Energy sovereignty took India fifty years and is still incomplete. Intelligence sovereignty cannot afford that timeline. The systems being deployed now are training the habits, vocabularies, and defaults of the next century's professional class. We have a short window to be a country that helped shape its own intelligence layer, rather than one that imported it without noticing.