The India AI summit signalled how India sees itself in the AI race: a stand-in for the Global South. The pitch drew confidence from the previous technology cycle of smartphones and digital payments and assumed the same formula would hold. As technology matured, products would become cheaper, scale would follow, and India could eventually export the model to other Global South countries. In a race dominated by the US and China, India cast itself as the country seeking a middle path. The summit’s slogan, “Sarvajan Hitaya, Sarvajan Sukhaya” (Welfare for all, happiness for all), reflected that positioning and framed AI as a technology to distribute widely rather than gatekeep.

The previous era taught Indian policymakers a particular lesson: the physical layer did not appear to matter very much. Cheap Chinese phones flooded the market, hand-me-down devices filled the gaps, and India already served as the IT back office of the world. The hardware question appeared to solve itself. What remained were software and standardised protocols, areas where India already had expertise. The formula worked because another country absorbed the industrial costs while India consumed that industrialisation on an enormous scale and built efficient digital systems on top of it.

So why does the same formula struggle in the AI era? What changed, or what is finally becoming visible?

India is one of the few countries that have leapfrogged technologically in a single generation. Before Reliance Jio entered the market in 2016, fixed broadband penetration remained extremely low and internet access was limited. Then data prices collapsed, and monthly usage surged. The shift showed that Indian capitalism could scale rapidly, but it also showed that the country did not gradually adapt to the internet so much as plunge into it at extraordinary speed.

Step back further, and the transition after 1991 looks unusual. India did not follow the textbook sequence from an agrarian economy to an industrial one and then to a services economy. Liberalisation and the Y2K boom together built an IT industry that reshaped middle-class aspirations. From the Y2K surge to the Global Capability Centre (GCC) boom, India created a services industry worth hundreds of billions of dollars and employing millions.

Across those same decades, however, manufacturing’s share of GDP fell steadily. World Bank-linked estimates place it around 12.5 to 12.6 per cent in 2024, among the lowest levels in decades. Services now produce most of India’s gross value added while employing a much smaller share of the workforce. The IT story never translated into broad industrialisation. It produced a prosperous high-skill enclave, visible in places like Whitefield and Gurugram, but it could not absorb labour leaving agriculture.

The arrangement worked for decades because the underlying infrastructure came from elsewhere. Cheap labour and large numbers of engineering graduates allowed the IT industry to scale at the right moment in globalisation. Meanwhile, the physical infrastructure behind the mobile internet boom depended heavily on Chinese manufacturing. Xiaomi, Oppo, Vivo, and Realme dominated the handset market, while Huawei and ZTE supplied networking equipment before the 2020 restrictions. India supplied users and market scale. Much of the hardware came from abroad.

The payments story deserves separate credit. By mid-2025, UPI handled more than 640 million daily transactions, slightly ahead of Visa’s global daily average. The National Payments Corporation of India built a system that many Western countries still struggle to replicate, and UPI’s overseas expansion has become a diplomatic success story. Yet the most widely used UPI apps still rely on Apple’s App Store or Google Play. India can regulate those platforms to a point, but it does not control them.

The government has recognised some of these vulnerabilities. That partly explains attempts to support indigenous operating systems and social media platforms such as Koo, whose momentum faded within four years. Nationalist branding alone could not overcome the reality that Indian consumers already inhabit a globalised culture. A services economy produces a different kind of consumer, one whose habits and desires are shaped internationally. The state can fund alternatives. It cannot compel cultural relevance.

Unequal access

Parts of the old formula still work in AI. Frontier Labs continue to build smaller and cheaper models because they want wider adoption. Costs per unit of capability will likely keep falling. But the deeper assumptions no longer hold because the physical base has changed. Chinese phones became cheap because Shenzhen absorbed the industrial cost. Cloud computing expanded because companies like Amazon Web Services rented infrastructure by the hour. AI hardware now sits inside a far tighter geopolitical and industrial system.

In a recent interview with writer and podcaster Dwarkesh Patel, Nvidia CEO Jensen Huang described AI as a “five-layer cake” made up of energy, chips, infrastructure, models, and applications. India’s presence weakens as one moves down the stack.

India’s per capita electricity consumption remains far below that of China and the United States. The country still faces summer power shortfalls. Most advanced chips are imported. India’s first commercial semiconductor wafer facility in Dholera is expected to lag several generations behind the chips used for frontier AI training. India’s data-centre capacity also trails major global hubs. Northern Virginia alone exceeds India’s operational capacity several times over. At the application layer, however, India remains strong through fintech, edtech, and IT services. Yet those applications depend entirely on the weaker layers beneath them.

Then came the “Mythos moment” in April, which reminded many governments that access to frontier AI models remains selective. Anthropic announced restrictions around a new cybersecurity model, limiting access largely to US-based firms and allied governments. OpenAI followed with a similarly narrow rollout of its own advanced model. Analysts such as Anton Leicht argued that the economics of serving frontier models, combined with security concerns in Washington, create strong incentives to limit access. The episode suggested that countries cannot indefinitely build AI ecosystems on borrowed infrastructure and borrowed models.

The physical world returns

The success of India’s IT industry, the resilience of its services economy, and the recovery after COVID together encouraged a broader assumption that the physical world mattered less than before. AI has challenged that assumption. The wars and shipping disruptions around West Asia sharpened the point further.

India imports most of its crude oil, much of it through the Strait of Hormuz. It also depends heavily on imported fertilisers, lithium, cobalt, and rare-earth materials. Semiconductor fabrication remains concentrated in Taiwan. Frontier AI models remain overwhelmingly American. The country hosting an AI summit and presenting itself as a broker between major powers discovered how dependent it still was on global supply chains.

None of this means India can recreate the industrial path followed by China in the 1980s. That world no longer exists. Labour-intensive manufacturing faces different economic and geopolitical conditions today. Yet, conceding that reality differs from convincing oneself that the physical layer no longer matters.

Private companies appear to understand the urgency. Firms such as Pixxel and Sarvam AI have discussed orbital or alternative compute infrastructure because terrestrial constraints remain severe. The broader lesson is harder to avoid: the digital economy India celebrated for decades rested on infrastructure largely built elsewhere.

The counterargument is obvious. India is building aggressively now. Microsoft, Google, Amazon, and Reliance Industries have announced large-scale AI and cloud infrastructure investments across the country. These projects amount to one of the largest waves of digital infrastructure spending India has seen.

Yet the ownership structure matters. Much of the capital, hardware, orchestration software, and frontier models remain foreign. India supplies land, power, construction, and labour. The comparison is less like South Korea building Samsung into a national champion and closer to Vietnam assembling phones designed and controlled elsewhere.

The decade ahead will not hinge on whether data centres get built. They almost certainly will. The harder question is whether India can use this period to build the systems it still lacks: reliable large-scale power, semiconductor fabrication at competitive nodes, domestic model development, and institutions capable of distinguishing between hosting technology and owning it.

That distinction appears even in smaller details. Economist Shruti Rajagopalan recently wrote that Tata Electronics discovered the soil at its Dholera semiconductor site could not support the near-zero vibration conditions required for chip fabrication. The project had to redesign its foundations, delaying production further.

The anecdote works both literally and metaphorically. India continues trying to place advanced industrial systems on a base that was never fully prepared to carry them. The Production Linked Incentive scheme allocated enormous sums across multiple sectors, but only a small share had been disbursed by late 2025. Large firms complained of delays and compliance burdens.

At the same time, India already employs a substantial share of the world’s semiconductor design engineers. AMD, Qualcomm, NXP Semiconductors, and Intel all run major design centres in the country. This remains one of the few layers where Indians already work at the technological frontier. Rajagopalan argued that a serious AI strategy may need to begin there, by turning engineering talent into domestic companies rather than relying solely on imported systems and foreign-owned infrastructure.

Kalim Ahmed is a writer and an open-source researcher who focuses on tech accountability, disinformation, and foreign information manipulation and interference (FIMI).

Also Read | Don’t panic about Ghibli-style AI art. Know what’s really at stake

Also Read | AI in India is trapped in culture wars