When something scarce and consequential is at stake, watch who invokes the language of sharing. It is rarely the party that controls the asset. The one with the weak hand asks to “open the cards”. The one holding leverage speaks instead of prudence, sequencing and risk. That simple asymmetry is the right way to read today’s fashionable demand to ‘democratise’ access to artificial intelligence and its ‘diffusion’.
The claim is morally attractive, especially when framed for the Global South. But it is analytically incomplete. Asking for democratisation is not an alternative to building strategic capability. It is, at best, a complement to it. At worst, it is a substitute for hard choices that must be made at home.
AI is not a monolithic technology that can be diffused like a textbook. It is a layered production system with tight complementarities and real bottlenecks. At the frontier, it rests on advanced semiconductors, dense compute clusters, high-bandwidth interconnects, reliable energy, cooling systems, engineering depth, evaluation tooling, data pipelines and deployment infrastructure. High fixed costs, steep learning curves and scale effects create natural concentration. Marginal access to one layer without the others yields sharply diminishing returns.
This is why ‘open-access’ rhetoric sits uneasily with industrial reality. You may be offered model weights, developer tools or research collaborations. But the binding constraints remain compute, energy and the tacit know-how that converts prototypes into reliable systems.
One is reminded of Michael Polanyi’s words that we know more than we can say. Capability is embedded in routines, tooling and organisational learning. It cannot be transferred wholesale by declaration.
Once AI becomes dual-use and general-purpose—embedded in defence analytics, cyber operations, surveillance, logistics optimisation and financial risk management—it moves from commerce to strategy. In such domains, export controls and technology-denial regimes are standard instruments of statecraft. The ability to deny access is itself a bargaining chip.
The best example would be nuclear technology. The nuclear age began with the language of peaceful diffusion. It consolidated into safeguards, export controls and a hierarchy of legitimacy. The universalist vocabulary was compatible with restrictive architecture because the technology altered the distribution of coercive capacity. AI is not nuclear in risk profile. But it is strategic in consequence.
So who asks for democratisation? Typically those outside the frontier. Their argument draws on fairness and global public goods language. The claim is not frivolous. But justice claims do not change the incentive compatibility of the supply side. When the supplier internalises the benefits of control and can externalise many costs of exclusion, the equilibrium is conditional diffusion.
India is well aware of this. The latest Economic Survey’s chapter on moving from import substitution to strategic resilience and then to strategic indispensability captures the hierarchy correctly. Growth must build durable, productive capability and reduce external vulnerability. Resilience is necessary.
Translate that frame to AI and the implications are immediate. If the 21st century runs on ‘compute’ and the minerals and energy that feed it, then AI capability is inseparable from energy policy, grid stability, logistics and regulatory credibility. Data centres are energy-intensive physical infrastructure, not abstract digital assets. Frontier model development requires capital depth, risk tolerance and outcome-oriented institutions. Without a domestically built stack, democratisation remains rhetorical.
The Survey’s tiered approach to indigenisation offers a disciplined template. First, identify non-negotiable vulnerabilities where denial would impose asymmetric national costs. In AI terms, that means assured compute capacity for critical missions, secure cloud and edge infrastructure for core sectors, and resilience buffers for energy and connectivity. The objective is continuity under stress, not the lowest possible short-run cost.
Second, build economically feasible capabilities where imports persist due to coordination failures or early scale disadvantages. For AI, this includes servers, storage, networking equipment, power electronics, specialised cooling, testing and certification infrastructure, model evaluation systems and domain-specific datasets. Support here must be time-bound and performance-linked. The central government has already been doing this by supporting Sarvam AI.
Third, avoid symbolic localisation in areas where substitution would raise economy-wide input costs without meaningful strategic gain. Restraint can be strategic. An input-cost reduction strategy is foundational for building resilience. Resilience that taxes downstream industry is fragility in disguise.
Advanced manufacturing matters in this story not as nostalgia but as a disciplining system. Manufacturing forces reliability, predictable rules, enforceable contracts and institutional follow-through. It exposes weaknesses that sheltered sectors can conceal. AI capability will require similar discipline. If approvals are slow, power unreliable, logistics brittle and regulatory friction high, frontier systems will not scale regardless of how many conferences celebrate democratisation.
There is also a macro-financial dimension that cannot be ignored. India’s external stability, though prudently managed, still relies significantly on capital inflows. Capital is conditional and reversible; export earnings are earned repeatedly through competitiveness. Currencies backed by persistent export capability behave differently under stress than those backed primarily by portfolio flows. If AI is to be a source of strength rather than dependency, it must feed into manufacturing exports and global value chain integration. Strategic indispensability means becoming a node others cannot bypass. It does not mean self-sufficiency for its own sake.
None of this argues against international cooperation on AI governance. Standards, interoperability, safety norms and research partnerships are necessary. But they are complements, not substitutes.
The temptation to substitute rhetoric for construction is understandable. Democratisation sounds virtuous and immediate. Capability-building is slower, messier and politically inconvenient. It requires institutional reform: outcome-oriented bureaucracy, tolerance for honest failure, credible withdrawal of support for non-performers and coordination across ministries and states. It requires aligning procurement with learning and learning with export discipline. It requires patience in a period when decades seem to happen in weeks.
But the alternative is dependence dressed as solidarity. The reader must therefore confront the uncomfortable hierarchy. In a world where trade, technology and finance are increasingly strategic, democratisation is meaningful only when underwritten by strength. The choice is not between cooperation and capability. It is between capability with cooperation, and cooperation without capability.
Aditya Sinha | Public policy professional
(Views are personal)
(On X @adityasinha004)