AI is not UPI: Why going by the UPI model risks stalling progress on artificial intelligence

This reasoning is flawed. The proposed analogy usually unfolds along familiar lines. AI development demands significant compute infrastructure—most notably high-end graphics processing units (GPUs), dominated today by Nvidia.

These are scarce and largely unaffordable for Indian startups and research labs. Capital, therefore, is framed as the binding constraint. The solution, it is argued, is for the state to step in: procure GPUs at scale, create a shared compute pool and rent access at nominal cost. Remove capital as a barrier and innovation will follow—just as it did with UPI.

But AI is not a payment system and cannot be governed through a compulsory licensing regime. It is a complex, evolving global supply chain that has been in the making for decades. Unlike UPI, AI did not emerge through a linear process of state-led design followed by private adoption. AI evolved through intersecting layers of code, data, labour, frameworks and compute power—each progressing at different speeds. Capital alone did not determine outcomes, nor did scale or state backing.

Until recently, it was widely believed that training competitive models required access to tens of thousands of GPUs, effectively restricting serious AI development to a handful of capital-rich firms. That belief shaped global policy thinking, including sweeping semiconductor export controls imposed by the US on China. These controls were premised on the belief that AI could be placed under a form of compulsory licensing through chip access.

That assumption proved misplaced. While sanctions raised costs and altered pathways, they did not confer durable control. Chinese firms responded by re-configuring the supply chain—they sought sovereignty not merely over models, but over data, frameworks and silicon itself. Frameworks like PyTorch remained central, even as efforts accelerated to reduce dependence on Nvidia hardware through adapters and alternative chips such as Huawei’s Ascend series. The supply chain adapted because it was neither shallow nor centralized.

This matters for India’s AI policy debate. It demonstrates that even the most powerful state, acting with allies, cannot place AI under a compulsory regime. The reason is structural: AI’s supply chain is both deep and broad. Compute is one input among many. Optimizing—or subsidizing—only that variable does not produce breakthroughs; it merely distorts incentives.

UPI, by contrast, worked precisely because the problem was narrow. Payments could be standardized. Interoperability could be mandated. Participation could be licensed. The institutional architecture—spanning the regulator, operating entities and the banking system—made this possible.

India’s policy record shows a recurring tendency to identify ‘national champions,’ channel subsidies towards them and align regulatory and administrative resources to ensure their success. Over time, this creates a symbiotic relationship in which the boundary between state and firm blurs, even if formal separation remains intact.

Such an approach may be defensible in sectors characterized by stable technologies and scale-driven economics. In AI, it is far more problematic. Models succeed not because they are favoured, but because they are useful. When utility surges and commercial success follows, protected national champions face an uncomfortable reality: they are preferred at home, but struggle to compete abroad—the arena where AI competitiveness is ultimately revealed.

The response historically has been predictable. Firms either petition the state for mandates, restrictions or bans to sustain relevance or pivot away from frontier innovation into adjacent niches. The result is familiar: national champions become local heroes but globally irrelevant.

None of this is an argument for state withdrawal. Public investment in research, skills, data and access is essential. But AI sovereignty cannot be manufactured through administrative tools or central allocations. It emerges from participation in global supply chains, not insulation; from diffusion of capability, not concentration of control.

Compute availability is the most visible input, but also the easiest to misprice, misallocate and overemphasize. When access is centrally provisioned, incentives shift from problem selection to resource consumption, privileging entities adept at navigating committees over those with technical know-how.

In a field where progress often comes from small teams pursuing unfashionable ideas with limited resources, such an approach biases the system against the kind of exploratory work that moves AI forward. AI is not UPI. Designing policy as though it were would be an expensive misreading of both.

The authors are, respectively, a corporate advisor and author of ‘Family and Dhanda’; and a strategic security and digital policy researcher.