The Indian AI Paradox

In December 2026, Sundar Pichai, Satya Nadella, and Sriram Krishnan all gave speeches in India within the same week. Pichai was announcing Google’s expansion of its Indian AI research program. Nadella was announcing Microsoft’s cloud infrastructure investment in data centers across multiple Indian cities. Krishnan, the AI policy advisor to the US National Security Council and a former Andreessen Horowitz partner, was representing the American government’s interest in deepening AI cooperation with India.

All three men are Indian-born. All three of them built their careers primarily in the United States. The fact that three of the most consequential AI figures in the world are of Indian origin, spent their formative professional years in California, and are visiting India to announce American investment is a perfect encapsulation of the Indian AI paradox: the country exports its best technology talent with remarkable efficiency and then has to import the capital and tools that talent helps build.

The scale of Indian talent contribution to American AI is extraordinary and documented. At any of the major American AI labs, a count of researchers by national origin would find India disproportionately represented — not just at the individual contributor level but in leadership. The founding teams of Inflection AI, Cohere, and several other significant AI companies include Indian researchers. The Google Brain team, in its most productive years, had Indians in several of its most senior research roles. OpenAI’s technical leadership has had significant Indian representation.

This is partly the legacy of IIT — the Indian Institutes of Technology, which have produced a disproportionate share of technology leadership globally for fifty years. It is partly the consequence of American universities that actively recruit IIT graduates, finding in them a combination of mathematical preparation, English fluency, and competitive drive that makes them attractive PhD candidates. It is partly the consequence of the American technology labor market, which rewards the specific skill set that IIT produces with compensation levels that no Indian company has historically matched.

The consequence is a brain drain that India’s policymakers have been discussing for decades without meaningfully reversing. The IIT alumnus who goes to Stanford for a PhD and joins Google does not represent a failure of the Indian education system — it represents a success, measured by the metric that the individual had every reason to optimize for. It represents a failure, measured by the metric of whether India is building the institutional capacity it needs to develop a domestic AI industry.

UPI — the Unified Payments Interface — is the thing that most development economists cite when they want to explain what India has gotten right on digital infrastructure. The system, built on a government-mandated architecture and launched in 2016, processes over 10 billion transactions per month as of 2026. It works on feature phones. It is free. It operates across banks with no unified consumer product. It has produced financial inclusion at a scale that private sector fintech, operating under market incentives, would never have delivered.

UPI succeeded because the Indian government made a specific architectural choice — interoperable, open, government-run infrastructure — that created a platform everyone could build on rather than a proprietary system that one company controlled. The parallel for AI is Indira AI’s proposal for a shared AI infrastructure layer: compute available to Indian researchers, companies, and startups at subsidized prices, with a governance structure that keeps the infrastructure as public infrastructure rather than a private platform.

Whether the AI infrastructure version of UPI is achievable is the central question in Indian AI policy. The two cases have crucial differences. UPI required solving a coordination problem (getting banks to interoperate) but not a capability development problem (the underlying technology was well understood). AI infrastructure requires both: coordinating access to compute, and ensuring that the models and tools built on that compute are competitive with the frontier products that Indian users will compare them to.

India’s government has shown it can solve coordination problems at scale — the Aadhaar biometric identity system, the GSTN tax network, the CoWIN vaccination platform are all examples of large-scale digital coordination that worked. The capability development problem is harder, because it requires not just organizational capacity but scientific research capacity that takes decades to build.

The domestic AI startup ecosystem in India has grown substantially since 2022, producing companies that are interesting in different ways. Sarvam AI, founded in Bangalore, has built multilingual foundation models focused on Indian languages and has made more progress on Indian language quality than any multinational’s localized product. Krutrim, the AI startup founded by Ola’s Bhavish Aggarwal, has received enormous attention (partly due to its founder’s public profile) and made investments in Indian language AI that are at least directionally correct. Niramai and several other medical AI companies are building diagnostic tools specifically designed for the disease burdens and clinical presentation patterns of the Indian population.

These companies are operating at a scale that is modest compared to the frontier labs but meaningful relative to the Indian market’s starting point. Sarvam’s models for Hindi, Tamil, Telugu, and several other Indian languages are genuinely better, for Indian-language tasks, than the localized versions of frontier American models. This is the domain-specific advantage that a focused team with local knowledge can achieve even at limited compute scale, and it is the most plausible near-term form of Indian AI competitiveness.

The challenge for these companies is the commercial path. The Indian consumer market is price-sensitive to a degree that makes charging for AI services difficult. The Indian government is a potentially large procurement customer, but government procurement is slow and subject to political considerations that do not always reward the best technical solution. The international market — selling Indian-language AI services to the Indian diaspora globally, or selling localization expertise to multinationals — is real but limited in scale.

The deeper structural issue is the relationship between the Indian AI ecosystem and the American AI ecosystem. India’s AI industry is not, for the most part, competing with American AI — it is deeply embedded in it. Indian IT services companies (TCS, Infosys, Wipro, HCL) have all made significant investments in AI integration services, helping Western clients deploy and customize American AI products. This is a large and growing business that reflects genuine Indian IT capability, but it is fundamentally a services relationship built on infrastructure that someone else built and owns.

Whether India can move from being a world-class integrator and customizer of AI products to being a world-class developer of AI capabilities is the strategic question. The infrastructure for this exists in partial form. The talent exists, though much of it is abroad. The capital investment is coming, slowly. What is less clear is whether the research culture — the institutional emphasis on basic research, the willingness to publish negative results and failed experiments, the tolerance for uncertainty that frontier research requires — is present in sufficient strength.

The American AI research culture that produced transformers, RLHF, and diffusion models was built over decades in universities with specific norms around academic freedom, graduate student funding, and the relationship between industry and academia. India’s IITs and IISc produce excellent trained researchers. They have not, historically, produced the kind of basic AI research that generates architectural breakthroughs. That is not a permanent condition — Oxford and Cambridge were not always the leading scientific universities they eventually became — but it is the current situation, and changing it is a generational project.

The optimistic case for Indian AI, stated concisely, is this: India has the talent, the market scale, the digital infrastructure precedents (UPI, Aadhaar), and the political motivation to build a significant domestic AI capability. The pessimistic case is that it has been in a similar position relative to software infrastructure for thirty years and has consistently produced the world’s best IT services companies rather than the world’s best software products.

The distinction between services and products is not a matter of intelligence or engineering skill. It is a matter of market structure, risk tolerance, and the specific organizational forms that innovation requires. Building a product company requires different institutions and incentives than building a services company, and India’s technology ecosystem has optimized, successfully, for services.

Changing that optimization is the real challenge in the Indian AI paradox. It is not a technical problem. It is an institutional one, and institutions change slowly.