By BarathVector Editorial — 2026-03-07

The GPU Arithmetic: India's AI Ambition Meets Hardware Reality

Part 6 of "The Delivery Deficit" series

By BarathVector Editorial | March 7, 2026

At the India AI Impact Summit in February 2026, IT Minister Ashwini Vaishnaw announced that India currently has 38,000 GPUs under the IndiaAI Mission, with 20,000 more to be added "in the coming weeks." The government's target is 200,000 GPUs by 2028, with a broader aspiration of attracting over USD 200 billion in AI infrastructure investment. The phrase "Manhattan Project for AI for public good" was invoked. Ninety-two countries endorsed the India AI Impact Summit Declaration. Google announced a USD 15 billion AI hub in Visakhapatnam. India set a Guinness World Record for AI responsibility pledges.

The rhetoric was, by any measure, superlative. The numbers beneath it are less so.

The Scale Problem

India's 38,000 GPUs exist in a world where:

• Microsoft committed USD 80 billion to AI data centres in a single fiscal year
• Meta's single Louisiana facility reportedly houses over 100,000 H100-equivalent GPUs
• Google operates an estimated 1 million+ TPUs and GPUs across its global infrastructure
• China, despite US export controls, has stockpiled an estimated 500,000 to one million high-end GPUs through grey-market channels and domestic alternatives like Huawei's Ascend 910B
• Saudi Arabia is building AI compute capacity measured in hundreds of thousands of GPUs, funded by sovereign wealth

India's 38,000 GPUs -- offered to startups at the subsidised rate of Rs 65 per hour -- represent a commendable democratisation of access. They do not represent sovereign AI capability at scale. They represent approximately what a single well-funded American AI startup would consider minimum viable infrastructure.

The Manhattan Project That Wasn't

The Manhattan Project cost USD 2 billion in 1945 dollars -- approximately USD 35 billion in today's money. It employed 125,000 people. It operated under a unified command structure with virtually unlimited political authority. It produced a deliverable within three years.

India's AI "Manhattan Project" has 38,000 GPUs, a Rs 1,000 crore budget allocation, and a timeline that stretches to "by 2028." These are not comparable enterprises. Using the phrase debases the original and inflates the current.

A genuine AI Manhattan Project for India would require:

• Compute at scale: Not 200,000 but 1-2 million GPUs, with a sovereign manufacturing or guaranteed supply chain for replacement and expansion
• Unified command: A single authority with cross-ministerial power, not a mission nested within the IT ministry competing with other priorities
• Funding at scale: Not Rs 1,000 crore but Rs 1 lakh crore over the decade -- still modest by global standards
• Energy infrastructure: AI data centres consume enormous power. India's grid, while improving, cannot reliably power large-scale GPU clusters without dedicated generation capacity
• Talent pipeline: Not just IIT graduates but a deep bench of ML engineers, chip designers, and AI safety researchers retained in India rather than exported to Silicon Valley

The Summit That Stumbled

The India AI Impact Summit itself became an inadvertent exhibit of the gap between announcement and execution:

Organisation: Delegates and journalists reported significant logistical difficulties -- long queues, unclear security instructions, conflicting entry guidelines. A summit positioning India as the global convener of AI governance could not manage its own entry gates.

Misrepresentation: Reports emerged of Chinese AI products being presented as Indian at the exhibition. The irony -- India's AI showcase featuring relabelled Chinese hardware -- was not lost on observers.

Accountability: Amnesty International stated that the summit "failed to rein in destructive practices by governments and technology companies." Civil society and human rights organisations were excluded from high-level plenaries while CEO Roundtables granted multinational corporations parity with sovereign governments.

Binding commitments: The 92-nation declaration, while symbolically significant, contains no binding obligations on AI safety, governance, or investment. It is a statement of aspiration, not a framework for action.

The Sovereign AI Paradox

India unveiled several sovereign AI models at the summit -- foundation models built by Indian teams, tested against global benchmarks, and in some cases outperforming larger international systems. This is genuinely significant. India is proving it can compete at the model layer.

But a sovereign AI model running on imported GPUs, in a data centre cooled by imported compressors, powered by a grid that runs on imported gas, processing data through imported networking equipment -- this is sovereign in the narrowest possible sense. It is intellectual property sovereignty without hardware sovereignty. It is a recipe without a kitchen.

Full-stack AI sovereignty requires: the model, the data, the compute, the chips, the energy, and the rare earths. India has the first two. It is building the third (slowly). It does not have the fourth. It has barely begun thinking about the fifth and sixth.

What Would Real AI Infrastructure Look Like?

India's natural advantages in AI are genuine: a massive domestic market for AI applications, a large engineering talent pool, a government willing to deploy AI in governance (Aadhaar, UPI, CoWIN), and a cost structure that makes AI development cheaper than in the West.

But advantages are not infrastructure. Infrastructure would look like:

• Dedicated AI zones with guaranteed power, cooling, and connectivity -- not repurposed IT parks but purpose-built compute campuses
• A sovereign GPU supply agreement with NVIDIA, AMD, or domestic alternatives that guarantees allocation regardless of geopolitical disruption
• An AI energy strategy that pairs every major compute facility with dedicated generation -- solar, nuclear, or hybrid -- rather than drawing from the national grid
• A retention strategy for AI talent that competes with Silicon Valley on compensation, research freedom, and career trajectory

None of these require authoritarian governance. All of them require money, planning, and the kind of institutional commitment that outlasts election cycles.

The Honest Assessment

India's AI ambitions are correct in direction. The IndiaAI Mission, the compute subsidies, the sovereign model development, the summit diplomacy -- all of this signals intent. And intent matters.

But 38,000 GPUs is not a Manhattan Project. Rs 1,000 crore is not a moonshot. A summit that cannot manage its own entry logistics is not positioned to manage global AI governance. And sovereign models on imported hardware are sovereign in name only.

The arithmetic does not lie. India needs to multiply its AI infrastructure investment by a factor of ten -- at minimum -- to be a serious player in the compute race. The alternative is to remain what it is today: a nation with excellent AI talent, world-class digital governance, and a hardware dependency that renders all of it contingent on someone else's factory.

This is Part 6 of "The Delivery Deficit" series examining India's announcement-execution gap.

Sources: India AI Impact Summit, NVIDIA, Amnesty International, Policy Circle, CIOL, BusinessToday