By Ramachandran Rajeev Kumar — 2026-02-28

For five days in February, Bharat Mandapam glittered. Twenty heads of government. A hundred-plus CEOs. Representatives from 118 countries. Five lakh participants. The India AI Impact Summit 2026 was, by every diplomatic measure, a triumph -- the first global AI summit hosted in the Global South, capped by the New Delhi Declaration and pledges exceeding two hundred billion dollars. Prime Minister Modi spoke of Vasudhaiva Kutumbakam. President Macron invoked Franco-Indian ambition. The UN Secretary-General nodded approvingly.

The noise was welcome. Summits are necessary. But let us ask a sharper question -- the one that no panel at Bharat Mandapam could comfortably answer: Can a twenty-three-year-old developer in Patna, or Coimbatore, or Guwahati, actually access affordable GPU compute and train a model? Not consume one. Train one.

Because that is the only litmus test that matters.

The Declaration and the Distance

The New Delhi Declaration -- endorsed by 89 countries -- is a well-constructed document. It proposes a Global AI Impact Commons for sharing use cases, a Trusted AI Commons for benchmarks and best practices, and a Charter for the Democratic Diffusion of AI. It speaks of equity, safety, and inclusion. The seven thematic "sutras" -- People, Planet, Progress, and four more covering economic growth, democratisation, human capital, and resilience -- read like the architectural blueprint for a just AI order.

None of this is objectionable. All of it is necessary at the level of international signalling. India positioned itself as the voice of the developing world, and it did so with genuine diplomatic weight.

But declarations are not data centres. Charters are not chips. And the distance between signing a communique in an air-conditioned hall and deploying inference at the edge in a Tier-3 city is -- to borrow from the summit's own language -- a matter of "democratic diffusion" that remains largely undiffused.

The Numbers: Progress With Caveats

Credit where it is due. India has moved faster on AI compute than most observers expected.

The IndiaAI Mission has deployed over 38,000 GPUs -- three times the original 10,000-unit target. At the summit, IT Minister Ashwini Vaishnaw announced 20,000 more GPUs would be added in the coming weeks, with plans to cross 100,000 by end of 2026. The hardware mix now includes NVIDIA H100 and H200 chips, plus Google Trillium TPUs from a third procurement round. Subsidised access is offered at roughly 65 rupees per hour -- between five and nine times cheaper than commercial cloud rates on AWS or Azure.

These are real numbers. The subsidy is real. And the pipeline of $200 billion in private AI investment, across all five layers of the stack, suggests serious capital is arriving.

But zoom into the structure, and the picture develops fracture lines.

Yotta Data Services -- a single Mumbai-based company -- controls an estimated 60 to 70 percent of India's GPU capacity. It is now deploying over 20,700 NVIDIA Blackwell Ultra GPUs as part of a $2 billion AI hub. L&T is building gigawatt-scale AI factory infrastructure in Chennai and Mumbai. E2E Networks rounds out the top tier. This is not a distributed compute landscape. It is a concentrated one, with capacity clustered in a handful of corporate hands and a narrow geography.

For a developer in Lucknow -- or a research lab in Bhubaneswar -- the subsidised GPU is not a phone call away. It requires registration through Meri Pehchaan, a project proposal outlining technical approach and estimated compute needs, institutional backing from a DPIIT-registered entity or recognised academic institution, and approval through a government portal. The subsidy is real. The bureaucracy around it is also real.

The Sovereignty Illusion

India now ranks third on Stanford's Global AI Vibrancy Index, leaping from seventh place in 2023. The country has over 159,000 recognised startups. AI market projections touch $126 billion by 2030 and a potential $1.7 trillion contribution to GDP by 2035.

Impressive. But here is the uncomfortable truth that the summit's own Bloomberg coverage surfaced: approximately 90 percent of global AI investment in 2026 comes from just two countries -- the United States and China. India is building rapidly, but it is building on rented foundations.

Every GPU in India -- every H100, every Blackwell Ultra -- is designed and manufactured by NVIDIA, a company that commands somewhere between 70 and 95 percent of the global AI chip market. India designs no competitive AI accelerators. It fabricates no advanced chips domestically. The Tata Semiconductor fab at Dholera, Gujarat, is expected to produce its first commercial chips by late 2026 -- an important milestone, but these are not cutting-edge AI chips. The Semiconductor Mission 2.0, allocated a modest 1,000 crore rupees for 2026-27, focuses on design IP, equipment, and supply chains. These are long-cycle investments. The AI race is a short-cycle contest.

The question the summit did not grapple with -- or rather, could not grapple with in a multilateral setting -- is structural dependency. Is India building sovereign AI capacity, or is it constructing the world's fastest-growing backend for American silicon?

When one foreign company's chips power your entire national compute stack, and one domestic company controls 60-plus percent of the rack space, "sovereignty" becomes a word that must be spoken carefully.

Governance Without Hardware Is Theater

India's AI governance framework, released alongside the summit, takes a deliberately light-touch approach. It opts for principle-based regulation over a standalone AI law -- a conscious rejection of the EU's AI Act model. Seven guiding sutras. Differentiated oversight based on risk. New institutional bodies including an AI Governance Group, a Technology and Policy Expert Committee, and an AI Safety Institute.

The framework is sensible. In a country where AI deployment is still nascent for most enterprises, a heavy regulatory regime would strangle innovation before it breathes. The pro-innovation posture is correct.

But governance is only meaningful when it governs something that exists. And what exists today is a compute infrastructure that is insufficient, geographically concentrated, dependency-laden, and bureaucratically gated.

Regulating AI without owning the means of AI production is like writing traffic laws for a country with three highways.

India can write the most elegant governance framework in the world. It can host the most photogenic summit. It can sign the most inclusive declaration. None of it changes the brute material reality: if you cannot offer a young Indian engineer affordable, accessible, low-friction compute, you are governing a mirage.

What Would Real Infrastructure Look Like?

The summit announced the right ambitions. The execution blueprint, however, remains incomplete. Here is what real AI infrastructure democratisation would require -- not in 2030, but now.

First -- distributed compute, not concentrated compute. India needs GPU clusters in Tier-2 and Tier-3 cities, not just Mumbai and Chennai. The IndiaAI Mission should partner with state governments and academic institutions to create regional AI compute nodes. A researcher at IIT Guwahati should not be queuing for capacity hosted 2,500 kilometres away.

Second -- zero-friction access. The current registration process through Meri Pehchaan, DigiLocker, and project proposals creates a credentialing barrier that filters out exactly the demographic India claims to empower: independent developers, small teams, and students without institutional affiliation. A subsidised GPU that requires a DPIIT registration number is not democratised compute. It is institutional compute with a discount.

Third -- domestic chip design for AI. The Semiconductor Mission 2.0 is a start, but the 1,000 crore allocation is a fraction of what is needed. India must invest aggressively in AI-specific chip design -- not just fab capacity for legacy nodes. ISRO designs world-class processors for space. DRDO builds sophisticated electronics for defence. The talent exists. The policy ambition, so far, does not match.

Fourth -- break the single-vendor dependency. NVIDIA's dominance is global, not uniquely Indian. But India could diversify towards AMD Instinct GPUs, Intel's Gaudi accelerators, and domestically developed AI chips. The IndiaAI procurement process should mandate vendor diversity, not default to NVIDIA because it is the path of least resistance.

Fifth -- energy infrastructure as AI infrastructure. AI data centres are ravenous consumers of power. India's plans to reduce AI infrastructure energy use by 35 percent through clean energy are promising but unquantified. Every new GPU cluster needs a parallel energy plan. Without it, compute expansion hits a power ceiling.

The Startup Paradox

India has fewer than ten globally notable machine learning models. Its deep-tech startup ecosystem -- companies building core AI systems rather than deploying existing ones -- remains thin. Patent intensity in AI lags both the US and China by wide margins. The overwhelming majority of India's 159,000 startups are application-layer companies: they consume AI, they do not create it.

This is not a criticism. Application-layer innovation is valuable. India's AI boom is pushing companies to trade near-term revenue for users, and consumer AI adoption is accelerating. But a nation that only consumes AI is a nation that rents its technological future.

The summit's promise of 50 deep-tech startups emerging "in the coming years" is aspirational. The question is whether the infrastructure exists to support them. You do not build a foundational model on 65-rupee-an-hour subsidised compute accessed through a government portal. You build it on thousands of GPUs running uninterrupted for weeks, with dedicated networking, massive storage, and engineering teams that can iterate without waiting for approval cycles.

India's sovereign AI models -- announced with fanfare at the summit -- are a step in the right direction. But they need to be stress-tested against global benchmarks, not celebrated in isolation. The hard truth is that India remains early in frontier model-building, and no amount of governance scaffolding will substitute for raw compute availability.

The Budget Signal

The Union Budget 2026-27 offers a revealing signal of priorities. It provides a tax holiday until 2047 for foreign companies using Indian data centres. It allocates $1.1 billion to a venture capital fund for AI startups. Both are welcome.

But notice the asymmetry: the tax holiday incentivises foreign firms to use Indian infrastructure. It does not incentivise Indian firms to build foundational AI. The VC fund is substantial but generic -- it does not specifically target the compute gap or chip design.

India risks building beautiful server farms that become outsourced processing centres for Silicon Valley, while its own developers continue to train models on Google Colab's free tier. That is not sovereignty. That is digital sharecropping with better branding.

The Only Metric That Matters

Strip away the diplomacy, the declarations, and the distinguished guests. The only metric that will determine whether India becomes an AI creator or an AI consumer is this: How many Indians can train a machine learning model on local infrastructure, at a cost they can afford, without navigating a bureaucratic labyrinth?

When that number is in the hundreds of thousands -- when a college student in Ranchi can spin up a GPU cluster as easily as she opens a UPI app -- India will have earned the right to call itself an AI power.

Until then, the summit was the overture. The infrastructure is the opera. And right now, the orchestra is still tuning.

The noise was welcome. The declarations were statesmanlike. But India has built mirages before -- in manufacturing, in defence production, in semiconductor dreams that stayed on paper for a decade. The AI summit will be judged not by how many presidents attended, but by how many processors a student in Nagpur can access. Hardware is not a footnote to governance. It is the prerequisite. Without silicon, the summit is just a mirage shimmering in the Delhi heat.