DORA’s 2025 report on AI-assisted software development found that roughly 90 percent of surveyed technology professionals were using AI at work. In DORA’s model, higher AI adoption was associated with higher software delivery throughput. Teams were able to move more changes through the delivery system.

DORA also found higher software delivery instability: the kind of failed deployments, emergency fixes, and unplanned rework that appear after software reaches users. More output did not automatically mean more reliable software.

The delivery-instability tension matters because software delivery is not only the act of writing code. AI can help a developer produce new code, but the software team still has to review it, test it, merge it, secure it, deploy it, monitor it, roll it back when needed, and maintain it after the original author has moved on.

The organizational setting around the tool matters. Teams with clear AI policies, usable internal data, strong version control, testing discipline, and good internal platforms are better placed to turn AI assistance into real gains. Teams without those routines may simply produce more changes than their review, testing, and deployment systems can handle.

This is the uneven acceleration problem.

The visible symptom is faster output. The deeper organizational issue is the work required to judge and use that output. The current wave of generative AI makes the pattern unusually visible because it reduces the cost of producing code, documents, summaries, reports, tickets, and recommendations. But it does not remove the need to decide whether the output is good enough, who is allowed to rely on it, and what happens if it is wrong.

The burden does not disappear when output becomes cheaper. It moves to the people and organizations asked to act on the output.

Yukti argued, in earlier work, that visible interfaces often sit on deeper stacks, and that real capability is built over time. Uneven acceleration adds a temporal dimension to those arguments. The issue is no longer only where capability sits in the system: it is how fast each layer moves, where the work of judgment accumulates, and whether the slower layer is learning or only becoming busier.

The Burden Moves Downward

Software makes the burden easy to see because generated output has to pass through review, integration, testing, and maintenance before it becomes part of a working system.

Large open-source repositories are a demanding test of AI-assisted productivity because they are real systems with existing users, maintainers, architecture, review standards, and accumulated context. In the METR randomized trial, experienced developers completed real issues from mature repositories; before work began, each issue was randomly assigned to either allow or disallow AI tools. The developers expected AI to make them faster and often felt that it had done so. Measured task completion showed the opposite: in this setting, AI-allowed tasks took about 19 percent more time.

A first attempt can appear quickly while the whole task still takes longer. Mature-codebase work includes reading old code, understanding design constraints, testing the change, checking side effects, and taking responsibility for the result.

Ordinary office work shows a similar burden shift. Researchers at BetterUp Labs and the Stanford Social Media Lab described AI-generated workplace output that looked polished but lacked the substance needed for the recipient to use it. In survey results reported in Harvard Business Review, 41 percent of respondents said they had received such work, with each instance costing roughly two hours of rework.

The problem is not that every AI-generated memo is bad. It is that cheap production can hide expensive reception. The sender saves time by generating something plausible. The recipient spends time reconstructing intent, checking claims, asking for missing context, and correcting work that should have arrived clearer in the first place.

AI adoption often begins with the tool arriving before the surrounding work has been redesigned. Employees are encouraged to use AI before managers know what good AI-assisted work looks like. Teams produce more candidate output before review systems, standards, and escalation routines adapt.

The generation layer is easy to see. The absorption layer is easy to leave under-resourced.

Figure 1: The Generation Layer and the Absorption Layer

Verification Becomes Institutional

Verifying an AI output is not just a technical step. It is an organizational act. Someone has to know the standard being applied, the evidence needed to accept the answer, the authority to approve it, the cost of error, and the record that will exist if the decision is challenged later.

AI can assist parts of verification. It can help write tests, summarize documents, search logs, detect anomalies, and compare claims. But final responsibility rarely sits inside the model. It sits in the organization that uses the output.

Operational Capacity Is AI Capability made the responsibility chain the main object: AI outputs become consequential through hospitals, courts, schools, welfare agencies, vendors, integrators, and frontline officials. Uneven acceleration adds the temporal pressure. When outputs multiply faster than organizations can verify, contest, escalate, and learn from them, verification becomes institutional rather than merely technical.

Arvind Narayanan’s ICML 2026 keynote makes the same point in economic language. Model capability becomes value only when it is joined to downstream reliability, integration, tacit knowledge, regulation, evaluation, and organizational adaptation. The model is upstream; value appears only when the surrounding system can use it well.

In public administration, those surrounding capacities are tied to legitimacy. If an AI-assisted eligibility system wrongly denies a benefit, the agency still has to preserve the record, explain the basis of the decision, route an appeal, assign responsibility, correct the harm, and learn from the failure. The output may be fast. Accountability is procedural.

Two common shortcuts fail here. The first is “human in the loop.” A reviewer without time, authority, domain knowledge, technical documentation, or escalation rights is not a meaningful control. The second is one-off training. OECD’s work on skills in the AI age treats AI not as a single shock to jobs, but as an uneven change in tasks across occupations, sectors, and regions. That makes adaptive learning institutions more important than one-time skilling.

Uneven acceleration describes what happens when demand for operational capacity rises faster than operational capacity itself. Operational capacity is the broader institutional ability to procure, evaluate, govern, and operationalize complex systems. The absorption layer is that capacity under pressure: the work of turning new outputs into reliable, accountable, and learnable practice.

AI is not the first technology to make one part of work easier before the surrounding routines catch up. Word processors made drafting easier before editing improved. Email made sending easier before filtering improved. Spreadsheets made modeling easier before model auditing improved. AI is a faster, more opaque, and more general version of that familiar pattern.

Organizations therefore need a simple way to ask whether they are absorbing AI or only accumulating more output.

A useful absorption test has four parts: what AI made cheaper to generate; who verifies the output and with what authority; where responsibility moves if the output fails; and what record improves the next deployment.

The test is deliberately modest. It does not require a new metric, a policy blueprint, or a formal audit regime before any AI use can begin. It asks whether an organization can see the work that moved. If nobody can name the verifier, the escalation path, the record of failure, or the learning loop, the institution is not absorbing AI. It is only receiving more output.

Figure 2: The Absorption Test

Governments Move On A Slower Clock

Public institutions face uneven acceleration under higher-stakes conditions. A July 2026 preliminary report of the UN’s Independent International Scientific Panel on AI is useful here because it separates access from value. Infrastructure, cloud services, and APIs may determine who can use AI, but the report stresses that adoption creates value only when it is joined to skills, data quality, workflow change, experimentation capacity, and institutional context.

That distinction matters for government. A ministry can gain access to AI tools faster than it can build evaluation routines, train staff, improve data quality, redesign workflows, or preserve accountability when errors occur. The clock that governs access is not the clock that governs public learning.

Public agencies cannot treat AI adoption as a simple tool rollout. The UN panel also notes that real-world effectiveness is rarely measured well, and that many countries lack the technical expertise to assess the most capable frontier models. The ability to evaluate AI is itself unevenly distributed.

The first slow clock is measurement. The OECD’s 2026 Digital Government Outlook found that only 10 of 36 OECD countries measure the impact of government AI use cases in any form. Some measure across sectors, some across government, and many do not measure systematically at all. That matters because deployment is easier to announce than learning is to prove.

The second slow clock is operational capacity. The UK’s National Audit Office found that many public bodies were piloting or planning AI while skills gaps, recruitment difficulty, data limitations, and unclear benefit measurement constrained responsible uptake. The problem was not a lack of interest. Agencies were trying to use fast-moving systems while their own evaluation, staffing, and governance routines were still catching up.

Technology-law scholars have long described a related pacing problem: emerging technologies often move faster than legal and regulatory systems can respond. Uneven acceleration is the broader institutional version of that problem. Regulation is one absorbing layer. Review systems, procurement routines, audit institutions, evaluation bodies, training systems, and professional judgment are others.

Some institutional lag is temporary. Workflows adapt, training improves, procurement offices learn, and evaluation routines become more sophisticated. But public institutions are not slower only because they are inefficient. They are slower because they carry duties that should not update at software speed.

Public agencies cannot revise rights, obligations, appeal routes, or accountability structures with every product update. Courts, auditors, regulators, ministries, hospitals, and schools carry inherited responsibilities. Their procedures encode legitimacy, not just friction.

The mismatch becomes visible when an AI tool enters a public workflow before the surrounding institution is ready for it. A tool can be procured in a year, piloted in a quarter, and updated continuously. The appeal process that must handle errors may still be paper-based, legally constrained, under-staffed, or spread across offices that do not share data well.

The system moves at software speed; the institution absorbs at procedural speed.

Figure 3: Two Clocks Inside AI Adoption

Institutional time is one reason AI governance cannot be reduced to model policy. A model may become safer, cheaper, or more capable with each release. The public institution using it may still lack the routines to integrate the model’s use into its procedures. Capability has to accumulate on both clocks: in the system being purchased, and in the institution expected to use it.

Access Travels Faster Than Absorption

Uneven acceleration has a national version. Countries can obtain AI tools faster than they can build the institutions that turn those tools into capability.

For middle powers, this gap is especially sharp. Models, APIs, cloud services, developer tools, and enterprise platforms can be imported or rented. Data governance, procurement routines, audit capacity, evaluation standards, appeal routes, and public-sector technical judgment have to be built locally.

A country can rent compute before it has mature data-quality standards. It can expand AI access before it has teams experienced in procurement, public-sector technical units, evaluation bodies, and audit routines that convert access into institutional learning. Access can move quickly because it is often purchased. Absorption moves slowly because it has to be practiced.

Access creates a second question for middle powers: who can change the terms later? Countries that depend on foreign models, cloud infrastructure, and data pipelines may gain AI access while retaining limited influence over standards, safeguards, pricing, model behavior, and local fit. The supplier may control the model, the cloud service, the data pipeline, or the interface through which the system is used.

The strategic question is whether a country can turn imported access into local learning fast enough.

The dynamic trilemma frames this setting. Middle powers want frontier access, low-cost efficiency, and enough sovereignty to avoid strategic exposure. Uneven acceleration makes that tradeoff operational. Even when access is available, the institutions that evaluate and govern the technology may not move at the same speed. The faster a country imports AI tools, the more important it becomes to build the capabilities around access.

The AI race is not one race. Frontier labs may race to improve benchmark capability, while states may race to diffuse good-enough AI through industry, create open-model ecosystems, strengthen domestic chip demand, or build evaluation capacity. China’s AI Plus strategy treats the strategic object as the speed at which AI becomes usable across production, services, governance, and infrastructure, not just the frontier model.

Benchmark races, diffusion races, and institution-building races do not move together. Benchmark capability can shift in months. Industrial absorption, public-sector evaluation, and domestic participation take longer.

The capabilities that matter are easy to miss because they do not look like frontier technology. They look like audit manuals, data validation standards, procurement templates, evaluation benchmarks, incident logs, staff training, and institutional memory. India matters here because it is trying to widen access while building parts of this absorption layer at the same time.

India’s Absorption Test

India is a live test of whether faster access can be converted into public-sector learning: whether compute, data, procurement, evaluation, and audit capacity are beginning to compound together.

Recent initiatives all point in the same direction: India is trying to widen access while building data, evaluation, and procurement capacity around it. The IndiaAI Mission and Compute Portal and AIKosh launch address access; the IndiaAI-Karya MoU points toward data and evaluation; the NeGD AI/ML empanelment creates a procurement channel. The test is whether these pieces become a learning system, not only adoption channels.

Audit gives the sharpest institutional test. In April 2025, the Comptroller and Auditor General issued an Artificial Intelligence Strategy Framework. The CAG is preparing to use AI inside its own audit work, including institutional knowledge systems and AI-assisted report drafting. It is also preparing to audit AI systems used across government. The same institution is therefore accelerating part of its own work while taking on a harder audit object.

The CAG case places both speeds inside one institution. The generation layer accelerates: AI can help draft, analyze, and search institutional knowledge. The absorption layer becomes more burdened: the audit institution must also learn how to inspect AI systems used across government.

Audit institutions matter here because they are memory machines. At their best, they convert scattered errors, weak documentation, procurement failures, and implementation gaps into reusable standards for the next round of public action.

Auditing government AI systems is slower because it is not the same as checking whether software was purchased or deployed. An auditor may need to understand the data pipeline feeding the system, the procurement record that specified performance claims, and the model documentation available to the agency. The audit may also need to follow decision logs, responsible-AI controls promised by the vendor, and the route through which an affected citizen can challenge an error.

The boundary between technical performance and administrative accountability becomes part of the audit object. A model may produce statistically acceptable results while still failing a public purpose if its errors are not visible, appealable, or properly assigned. A procurement file may show that an AI system was delivered, while the harder audit question is whether the agency can evaluate, contest, and improve the system after delivery.

Middle-power capability is often built through this kind of institutional accumulation. Not one model release, not one national champion, and not one compute target, but the slower growth of public institutions that can evaluate, audit, procure, and learn from the systems the country adopts.

India’s AI strategy will be judged partly by its models, compute, startups, and applications. It should also be judged by whether its public institutions become harder to fool, better at specifying what they need, and more capable of saying no to systems they cannot evaluate.

What Generation Leaves Behind

The language of AI strategy often favors speed. Countries are told to move quickly, firms to adopt, workers to use new tools, and public agencies not to fall behind.

Speed is not the same as capability. Faster generation is valuable when it leaves stronger systems behind. It is dangerous when it leaves only more outputs for the same weak review, evaluation, procurement, and accountability routines to absorb.

AI-assisted coding is progress when review, testing, security, and maintenance improve with it. Public-sector pilots build capability when agencies measure impact, retain learning, and improve the next deployment. Compute access deepens capability when it is joined to evaluation, procurement, audit, and data-governance capacity.

Uneven acceleration is therefore not an argument for slowing everything down. Some layers should move quickly. Others should move deliberately because they carry responsibility, legitimacy, safety, and the risk controls that let learning compound. AI lowers the cost of producing candidate knowledge: code, analyses, reports, designs, recommendations, and explanations. The strategic task is to build institutions capable of absorbing those outputs at the pace they are created.

For middle powers, this may become one of the central tests of AI strategy. Frontier access matters. Low-cost efficiency matters. Domestic capability matters. But the everyday machinery that turns access into judgment matters just as much: procurement teams, evaluation bodies, audit institutions, data stewards, public-sector technical units, and professional routines that learn over time.

A strategic mistake follows from measuring adoption more easily than absorption. A government can count AI pilots, compute credits, users, and deployments before it can show that evaluation routines, procurement memory, audit capacity, data stewardship, and professional judgment are compounding.

The generation layer can make an organization or country look more capable before it actually is. The absorption layer decides whether the appearance becomes real.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Further Reading

• Arvind Narayanan, “What will be left for us to work on?” and the companion essay - ICML 2026 keynote and essay arguing that AI’s economic impact depends on downstream reliability, integration, tacit knowledge, regulation, evaluation, and organizational adaptation, not only model capability.

• Gary Marchant, Braden Allenby, and Joseph Herkert, The Growing Gap Between Emerging Technologies and Legal-Ethical Oversight, and Lyria Bennett Moses, “Recurring Dilemmas” - technology-law scholarship on the pacing problem and the recurring difficulty of adapting legal systems to technological change.

Sources and Case Materials

• DORA, State of AI-assisted Software Development and METR on early-2025 AI tools - software evidence on throughput, stability, and the gap between perceived and measured speed.

• Harvard Business Review on AI-generated workslop - workplace evidence on burden shifting to recipients.

• UN Independent International Scientific Panel on AI, Preliminary Report - July 2026 preliminary evidence on uneven adoption, access versus capacity, and underdeveloped evaluation.

• OECD, Skills in the AI Age and Digital Government Outlook 2026 - workforce and public-sector evidence on uneven absorption.

• NAO, Use of Artificial Intelligence in Government - UK public-sector capacity and skills constraints.

• CAG AI Strategy Framework - Indian audit-capacity evidence for AI in government.

• PIB on IndiaAI-Karya MoU - IndiaAI, AIKosh, model evaluation, and dataset standards.

• CSET translation of China’s AI Plus policy - Chinese policy evidence on industrial diffusion, open ecosystems, and AI adoption targets.

In June 2026, Sarvam announced that it had raised $234 million in the first close of a $300 million Series B, at a post-money valuation of $1.5 billion, with HCLTech investing $150 million as lead strategic investor. The sovereignty signal was explicit rather than inferred. In the same announcement, Sarvam described itself as “India’s full-stack sovereign AI company,” and HCLTech framed the investment as a step toward a trusted, globally competitive Indian AI ecosystem.

The framing is understandable. Sarvam sits at the point where several Indian ambitions meet: domestic foundation models, Indian-language capability, enterprise and government applications, and the broader IndiaAI push toward indigenous AI capacity.

A national AI company is a signal. National AI capability is the harder test.

The phrase sovereign AI can make several different claims at once: domestic models, local-language systems, data control, compute access, enterprise deployment, public-sector assurance, and some degree of national agency over a technology stack. That is why it has to be tested layer by layer.

The harder question is what the company, the mission, and the surrounding ecosystem leave behind. The useful signs would be trained engineers, stronger firms, better datasets, local benchmarks, stronger evaluation routines, deeper procurement competence, and public institutions that can absorb AI responsibly. The weaker outcome would be symbolic sovereignty: the announcement standing in for the capability, while the deeper learning, infrastructure, market depth, and operational judgment remain thin.

Sovereign AI now names a strategic problem as much as a technology ambition. For a country like India, the question is how to pursue it while moving through a trade-off between technological sovereignty, frontier access, and low-cost efficiency.

Most countries want all three. They want control: enough domestic capacity and leverage that critical systems cannot be withdrawn or dictated from outside. They want frontier access: the ability to use the best models, chips, cloud systems, and research networks. And they want low-cost efficiency: AI cheap enough to diffuse through firms, universities, public agencies, hospitals, and small enterprises.

AI makes the tension difficult to hide. Control often raises cost. Frontier access often requires dependence. Low-cost efficiency often rests on global specialization and hyperscale systems that weaken domestic control. Control, frontier access, and low-cost efficiency pull against one another; together these pressures form a trilemma of technology strategy.

Three Objectives, One Constraint

Technological sovereignty is often misread as self-sufficiency. Countries rarely control every layer of the AI stack. Chips depend on design tools, lithography systems, fabrication knowledge, and supplier ecosystems. Models depend on compute, data, engineering teams, and evaluation systems. Deployment depends on institutions that can procure, evaluate, govern, and maintain what they use.

Technological sovereignty means agency inside interdependence: the ability to shape the terms on which critical systems are accessed, governed, and improved. In operational terms, this means critical systems cannot be withdrawn without warning; public institutions can audit what they deploy; local firms can build on reliable infrastructure; and the state has options other than accepting whatever external providers offer.

Frontier access matters for the opposite reason. Countries learn by working with the systems that define the frontier. Firms learn by integrating with leading platforms. Researchers learn by working with powerful tools. Public agencies learn by deploying and evaluating real systems before a fully domestic stack appears.

For middle powers, low-cost efficiency is central. AI that is affordable for a major American platform company may be expensive for an Indian startup, a state government, a public university, or a small hospital. If access is too expensive, diffusion narrows. If diffusion narrows, learning narrows. If learning narrows, capability formation slows. Low-cost access can become one of capability’s preconditions, even though it never guarantees capability by itself.

The difficulty is that the three objectives pull against one another. A country can push toward control and frontier access, but it must absorb the cost of subsidies, redundant infrastructure, industrial policy, and long-term institutional funding. A country can pursue control and low-cost efficiency, but it may have to accept domestic systems below the frontier. A country can pursue frontier access and low cost, but only by leaning heavily on external platforms, chips, model ecosystems, and capital.

The static version of the trilemma describes this constraint: each objective makes the others harder to maximize. The dynamic version asks how countries reshape the constraint over time. Capability changes the trade-off itself. A country that builds evaluation capacity, procurement memory, data infrastructure, technical teams, and supplier depth still faces the trilemma. Capability does not erase the constraint. It changes the terms on which the next trade-off is faced.

Figure 1: Three objectives; every country faces trade-offs.

Trade-Offs Change Over Time

Countries inherit a position before they choose a strategy: a chip industry they lack, a services sector they have, an energy grid under strain. Then they try to alter the choices available later.

The sovereignty debate tends to map positions. The Tony Blair Institute’s Control / Steer / Depend frame is the most useful map, and it gets two things right. It rejects technological isolationism: countries exercise sovereignty inside interdependence. And it forces a layer-by-layer view of the stack, against the habit of treating sovereignty as a single national attribute. The harder temporal question is how a country moves among control, steering, and dependence as its capability changes.

The three postures also reveal the mechanism of movement, because each one has a capability price of admission. Control requires the capacity to build, finance, govern, and maintain. Steering requires market power, regulatory credibility, procurement leverage, or standards influence. Dependence becomes strategic only when the dependency is understood well enough to be governed, negotiated, and used for learning.

Posture and practical work sit at different levels. Inside a layer, the practical work may still be to build, adapt, rent, import, or coordinate.

Movement of this kind has precedents. In 1991, ISRO signed a contract with Glavkosmos, then the Soviet external space agency, for cryogenic stages and transfer of technology. The arrangement soon ran into American objections under missile-technology-control rules. In a 1995 Lok Sabha exchange, the government said the agreement had been revised after force majeure was invoked for geopolitical reasons, and that the technology-transfer and training commitments for cryogenic stages had not been carried through. India was left to choose between accepting a weaker launch position or building the capability domestically. ISRO chose to build. The indigenous cryogenic upper stage took roughly two decades to mature and flew successfully in 2014. It was slow, expensive, and repeatedly delayed. It also helped strengthen India’s bargaining position in later space cooperation.

Indian pharmaceuticals moved through a different instrument. The 1970 Patents Act’s process-patent regime helped create legal room for domestic firms to accumulate manufacturing and reverse-engineering capability. Firms, chemists, public laboratories, regulators, and export markets turned that legal room into capability over three decades. The accumulated result shifted India’s position, from a market largely organized around foreign patent-holders toward a major supplier in global generic medicines. In both cases the constraint remained. The country’s position inside it moved, because capability had accumulated in the interim.

External access becomes strategic only when it is organized as learning. Rented compute has to train engineers, improve procurement, and strengthen local data systems; otherwise it remains consumption. Imported tools have to leave behind standards, benchmarks, documentation, supplier knowledge, and institutional memory. Access buys time. Institutions decide whether time becomes capability.

The dynamic trilemma draws on older development literature on late industrialization and technological capability. Amsden and Lall both describe how late developers improve their position when external access, pressure, and institutional discipline become repeatable learning. Technology transfer rarely happens automatically. It has to be absorbed, practiced, adapted, and institutionalized. AI extends the same problem through different instruments: compute investments, data infrastructures, standards bodies, testbeds, procurement teams, and evaluation routines. Each matters only if it leaves behind capability that can be reused in the next round.

Figure 2: Capability changes tomorrow’s trade-off; announcements leave it mostly where it is.

The earlier Yukti essays were building toward this synthesis. The Stack Beneath the Interface showed why AI capability decomposes by layer. Operational Capacity Is AI Capability showed why evaluation, procurement, and institutional judgment are capabilities in their own right. Capability Formation, Not Technology Adoption showed why access and adoption matter only if learning compounds. The dynamic trilemma puts those pieces together: strategy concerns both where a country sits today and what today’s choices make possible tomorrow. Nowhere is that discipline harder to hold than in India’s current moment.

India’s Selective Route Through the Trilemma

India’s route, in AI and other emerging technologies, is likely to be selective sovereignty: depend in some layers, steer in others, and deepen domestic capability where institutional control matters most. The discipline is to distinguish dependencies that preserve access and learning from dependencies that quietly narrow future choices if they remain unmanaged.

Full-stack control sits beyond any rational near-term ambition. The deepest layers of frontier AI are too concentrated, too capital-intensive, and too exposed to supply-chain and geopolitical constraints. Advanced chips, frontier compute, and hyperscale cloud infrastructure will remain partly external for some time.

External dependence reaches beyond procurement. AI is becoming too important to public administration, health care, courts, and everyday service delivery for all reliance to be treated as interchangeable sourcing. Some dependencies can be managed because they are visible, contestable, and paired with domestic learning. Others become strategic exposure because they shape public systems without leaving enough local capacity to question or repair them.

The IndiaAI Mission reflects this mixed posture. Its official architecture includes compute capacity, datasets, application development, startup financing, skilling, indigenous AI capability, and Safe and Trusted AI. The mission is addressing real access constraints. The strategic test is what forms around the access.

Consider the compute pillar in operation. The mission aggregated demand and empanelled private data-center firms through competitive bidding: by March 2026, 38,231 GPUs had been onboarded through 14 empanelled providers, offered to startups, researchers, and other approved users at a subsidized average rate of ₹65 per GPU-hour, a price the government describes as roughly a third of the global average. Earlier portal materials described eligible users receiving up to 40 percent subsidy on AI compute services. IndiaAI did not build a national GPU cluster. As industrial policy, this is a coordination move rather than a construction project: the state used its position as a demand aggregator to change the price of learning.

Follow the subsidy through a funding cycle, and the distinction between consumption and capability stops being abstract. A startup that trains a model on subsidized compute has, at minimum, consumed cheaper GPU-hours. Capability compounds only if something reusable survives the run: evaluation harnesses that outlast the model, benchmark results for Indian languages published where other teams can reuse them, or data work that becomes shared infrastructure beyond one company’s pipeline.

For a public agency, the same question runs through procurement. The first AI contract is often signed with little internal knowledge of what to specify. The second is better only if evaluation results, vendor records, and failure reports were preserved as institutional memory rather than leaving with the official who managed them. Whether routines like these are required deliverables of public support or incidental by-products is much of what separates capability investment from consumption support.

The subsidy’s real test arrives when it ends. If trained teams, published benchmarks, and procurement templates persist, the subsidy was capability investment. A ministry should be able to write a sharper tender in 2029 because of what it learned running subsidized workloads in 2026. If demand collapses back to the pre-subsidy baseline, the subsidy was consumption support with a fiscal wrapper. Reported under-utilization of the portal’s GPUs in early 2026 shows the question is live. The same test applies to the mission’s other pillars: AIKosha can become usable data infrastructure or remain a portal, and safety tools can enter procurement and certification practice or stay at the level of research outputs.

India’s targeted-leverage strategy extends to 2026 AI diplomacy: the M.A.N.A.V. frame asserts agency over how AI enters Indian society, while the Pax Silica initiative seeks more reliable access, through trusted supply-chain cooperation, to the hardware layers India cannot yet control alone.

Digital public infrastructure supplies a different kind of leverage: steering. Aadhaar, UPI, DigiLocker, and related rails give the state a way to shape participation, interoperability, and data flows across the digital stack. As Digital Public Infrastructure and Its Limits argued, DPI can help organize AI adoption, while chips, compute, frontier models, and institutional judgment remain separate capability problems.

Operational capacity is the layer where selective sovereignty becomes visible inside institutions. A ministry that procures AI systems without audit routines can become dependent even when the vendor is domestic. A hospital that uses prediction without local validation can lose judgment even when the model is affordable. A school board that digitizes assessment without credible recourse can modernize the interface while weakening legitimacy. Generative systems make this harder because plausible outputs still need verification; fluency can hide error, bias, and missing context. Why Technology Is an Institutional Problem made this point through the operating layer: procurement, workflow, oversight, and recourse are where technology becomes institutional.

The matrix makes the selective route visible as a capability test. It asks where durable capability has accumulated enough to change posture, where early build efforts are underway, and where managed access still remains the realistic near-term position. Selective sovereignty works as a layer-by-layer judgment about where India can depend, where it can steer, and where domestic capability has to deepen because the public consequences are too important to outsource casually.

Figure 3: Middle powers choose posture by layer.

The matrix shows that managed dependence in one layer can coexist with aggressive capability-building in another. India can accept managed dependence in frontier chips while building and steering public-sector operational capacity: evaluation routines, deployment competence, data systems, and procurement memory that make future dependence more governable. Selective sovereignty puts ambition where capability can actually compound.

Other Paths Show the Cost

China shows one path through the trilemma: pushing toward control by absorbing very high cost. Its strategy reduces exposure in strategic layers by building manufacturing depth, supplier networks, state-directed industrial capacity, technical universities, and redundant domestic alternatives where foreign dependence looks dangerous.

The control China seeks is especially visible in the industrial base around chips, advanced manufacturing, platforms, data governance, and supply chains. Barry Naughton’s work on China’s industrial policy and Chris Miller’s history of the semiconductor contest describe the scale of that effort and its constraints. China’s path shows that the trilemma shifts through decades of cumulative capability investment; isolated technological wins cannot do the same work. More control requires capital, state coordination, redundancy, and a willingness to tolerate inefficiency for resilience. Even then, control remains incomplete. Advanced semiconductors still expose bottlenecks. Export controls still matter. Equipment ecosystems and process knowledge cannot be replicated instantly. Scale changes how much of the cost of sovereignty a country can absorb. The cost remains.

The European Union shows a different path: steering through market size and regulation. The EU is not at the center of frontier model production, advanced-chip supply, or hyperscale AI cloud. Its assets are market size, regulatory authority, institutional credibility, and the ability to make rules travel through compliance. Anu Bradford’s account of the Brussels Effect gives the strongest version of that logic, though its reach is contested and its domestic costs are increasingly visible. The EU AI Act applies the steering ambition to AI, making regulation a strategic instrument as well as a legal framework. Europe changes its position inside the trilemma primarily through steering rather than control.

Europe pays for steering in compliance burden and diffusion speed. Obligations can weigh heavily on smaller firms and deployers. Standards take time to bargain over. Rule-making authority does not by itself create frontier industrial depth. The Union has already paid part of that price in public: in 2026 it agreed to delay the Act’s high-risk obligations and extend simplifications for smaller firms, after sustained pressure over compliance costs. Governance leadership is real technological power: it shapes how firms document systems, classify risks, and prepare for market access. Regulatory influence can shape deployment while domestic frontier leadership still lags.

The United States sits differently because much of the frontier lies inside its firms, capital markets, cloud infrastructure, and legal jurisdiction. Its problem is not how to access the frontier from outside, but how to protect, power, finance, and govern a frontier it substantially hosts. Even there, sovereignty and access carry costs. Export controls, industrial subsidies, power constraints, and supply-chain reconfiguration are all signs that frontier access has acquired a strategic cost.

The comparison clarifies the cost structure. Control absorbs capital, redundancy, and inefficiency. Steering depends on institutional credibility and market leverage. Dependence buys speed but creates exposure. Middle powers face the harder version: choosing which layers to deepen, which to steer, and which to access. A further question then becomes unavoidable: whether any of this still matters once the great powers start forcing choices.

Alignment Is Not the End of Strategy

One response says the middle-power problem has become simpler. If technology stacks harden, if access becomes conditional, and if hedging is punished, then middle powers should choose the larger system that offers the best shelter from their gravest threat.

The alignment-first view sees something real. Great powers are no longer merely offering access to markets, finance, chips, and security. They are using the systems they command to sort other states into more disciplined relationships. A middle power may control a useful node: a resource, a market, a dataset, an institutional niche. A node that can disrupt a system still lacks command over the system itself.

A harder hierarchy changes the terms of strategy. It does not end capability accumulation. Alignment takes several forms. A country can align as a buyer of protection, as a useful node in a larger system, or as a capable partner whose institutions, firms, and operational competence make it harder to ignore. Accumulated capability, more than diplomatic language, separates those postures. Aligned access can also accelerate capability, when partnership transfers knowledge, industrial participation, and institutional experience beyond finished systems.

AI makes capability-based alignment especially important because the technology enters society through ministries, hospitals, courts, regulators, and procurement systems as well as frontier labs. Evaluation capacity, procurement competence, and operational judgment give middle powers a way to turn access into bargaining position and avoid passive integration.

Middle powers face the trilemma most sharply because all three objectives remain necessary. They need frontier access for learning, enough sovereignty to avoid strategic exposure, and low-cost efficiency for broad diffusion. Pursuing any one carelessly weakens the others. That is why the useful question about any single move is what it does to the moves that come after.

The Political Economy of Capability Formation

Control, steering, and dependence are outcomes of capability, but they are also inputs into future capability. A country can depend in ways that teach or in ways that hollow out local capacity. It can steer in ways that build institutional authority or produce symbolic standards nobody follows. It can seek control in ways that create durable industrial depth or waste public money on isolated assets. The useful question about a national model, a cloud partnership, a compute subsidy, or a regulatory regime is what it does to the country’s next set of choices.

Ordinary institutional routines determine whether access teaches or hollows out. Ministries have to fund maintenance after announcements. Universities have to train for judgment alongside tool use. Regulators need technical credibility. Public agencies need to preserve evaluation results, vendor lessons, and failure reports across official postings. Firms have to invest beyond import-and-integrate business models if domestic capability is to compound around access.

Capability-building and dependence organize around different coalitions. Capability-building routines are slow, cumulative, and institutionally dispersed. Dependence is transactional, contractable, and often paid on short cycles. In many procurement systems, a cloud reseller can earn its margin even when the client’s engineers learn little from the workload. A consulting firm can be paid to integrate a system, not to train its client out of needing integration. An official may earn credit for an inauguration within a posting cycle of a few years; a capability program pays off in ten, under someone else’s name. The political economy is asymmetric: dependence has organized beneficiaries today; institutional depth has dispersed beneficiaries tomorrow.

The coalition for dependence often speaks in the language of the trilemma’s third pillar. Imported integration is faster, cheaper, and easier to justify inside a short policy cycle than slow institutional depth. The coalition for capability asks for slower investments whose returns arrive later: knowledge transfer, open standards, reusable benchmarks, institutional learning, and procurement memory. Low-cost efficiency is a real objective, but it also becomes the argument incumbents reach for when capability formation threatens their contracts. A country can remain dependent with ample talent and ambition when the coalition that benefits from dependence is stronger than the coalition that would build its way out.

India can attempt selective and temporal strategy because the country has some negotiating surface: scale, technical labor, public digital rails, strategic geography, and relationships across more than one technology bloc. These assets will continue to matter only if capability compounds around them.

For middle powers, dependence on global systems will remain necessary where access matters more than immediate control. Domestic depth should be built where dependence would block learning, weaken bargaining power, or expose essential public functions. Public infrastructure and regulation can steer markets where the state has leverage. Operational capacity matters where public institutions must evaluate, procure, audit, and contest AI systems. And capability formation has to compound across repeated procurements, deployments, and institutional routines.

Sarvam will matter less as proof that India possesses sovereign AI than as evidence of whether India is becoming more capable of exercising sovereignty over time. The question is whether one well-funded company becomes a site of training, evaluation, deployment learning, and ecosystem spillover, or a proof point in a story about arrival. The same test applies to compute subsidies, public rails, safety institutes, and cloud partnerships, and it will keep applying, case by case, as AI settles into hospitals, courts, classrooms, and ministries.

No country escapes the trilemma. The objective is to invest so that tomorrow’s trade-offs arrive less severe than today’s. The frontier can be leased for a while. Capability cannot.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Earlier Essays

• The Stack Beneath the Interface on why AI capability has to be read layer by layer.

• Operational Capacity Is AI Capability on why evaluation, procurement, and institutional judgment are forms of AI capability.

• Digital Public Infrastructure and Its Limits on where public rails can steer AI adoption, and where they cannot.

• Capability Formation, Not Technology Adoption on why access and adoption become strategic only when learning compounds.

• Why Technology Is an Institutional Problem on why procurement, workflow, oversight, and recourse shape what technology becomes.

Further Reading

• Tony Blair Institute, Sovereignty in the Age of AI — the Control / Steer / Depend typology this essay extends in time.

• Collective Intelligence Project, Solving the Transformative Technology Trilemma through Governance R&D — a different trilemma, focused on progress, participation, and safety at the societal level; adjacent to the national-strategy trilemma here.

• Alexander Gerschenkron’s Economic Backwardness in Historical Perspective, Alice Amsden’s Asia’s Next Giant, Sanjaya Lall’s Learning to Industrialize, and Robert Wade’s Governing the Market — classic development literature on late industrialization, state discipline, technological capability, and learning under constraint.

• William Baumol’s “Entrepreneurship: Productive, Unproductive, and Destructive” — background for the political-economy point that incentives shape whether talent and enterprise become productive capability or rent-seeking activity.

• Barry Naughton’s The Rise of China’s Industrial Policy, 1978 to 2020, Chris Miller’s Chip War, and Anu Bradford’s The Brussels Effect — background on China’s industrial-policy path, semiconductor chokepoints, and EU regulatory power.

• September 2025 GITAM-KSPP lecture slides on the AI trilemma — an earlier AI-strategy version of the trilemma; this essay generalizes the frame across emerging technology strategy.

Sources and Case Materials

• HCLTech press release on Sarvam’s Series B first close

• IndiaAI Mission cabinet approval

• IndiaAI Compute Portal, AIKosha, and public-sector AI competency initiatives

• PIB release on AI compute capacity and data-center growth, March 2026

• PMO on the India AI Impact Summit 2026 and M.A.N.A.V.

• PIB release on India joining Pax Silica

• Economic Times on reported under-utilization of IndiaAI portal GPUs

• ISRO archived page on the GSLV-D5 indigenous cryogenic stage mission

• USITC Journal of International Commerce and Economics paper on Indian patent law and pharmaceutical FDI

• EU AI Act regulatory framework

• Council of the EU on the 2026 AI Act simplification agreement

• Federal Register rule on 2022 U.S. advanced-computing and semiconductor export controls

• State Department statement on 1992 ISRO-Glavkosmos sanctions, Lok Sabha Debates, May 10, 1995, and Federal Register Vol. 57, No. 97, p. 21319 (May 19, 1992) — primary trail for the cryogenic-engine transfer episode. See also Gopal Raj’s Reach for the Stars.

On February 9, 2026, the Central Board of Secondary Education told schools that Class XII answer books would be evaluated through an on-screen marking system.

The change was easy to describe as digitization. Physical answer books would be scanned. Examiners would mark them on screen. CBSE described the system as a way to reduce logistical delay, widen the pool of evaluators, reduce transport time and cost, and avoid common errors in totaling and transfer of marks.

The scale was large even by Indian administrative standards. According to reporting after the evaluation cycle, the system involved 98.66 lakh answer books, about 70,000 evaluators, more than 6,000 evaluation centers, and roughly 88,000 computers. This was not a small pilot that could be repaired through informal supervision. It was a national digital workflow for a high-stakes public examination.

No model was reading the answers. No algorithm was deciding a student’s marks. Human examiners remained formally in charge of evaluation.

The case matters because it shows that the institutional problem begins before AI enters the picture. Technology does not merely enter an institution. It reorganizes where institutions exercise judgment, what evidence can be seen, who can contest error, and how responsibility is distributed.

A paper answer book becomes a scan. An examiner now marks through a portal. A scan-related grievance becomes an online request inside a narrow window of time.

A student’s confidence in evaluation then depends on more than the examiner’s judgment. The examiner can judge only what the system shows. The student can contest only what the system reveals. The board can defend the process only if the digital workflow preserves evidence, authority, and credible recourse.

CBSE did not present the shift as an improvised change. Its circular referred to familiarization access, dry runs, training programs, a call center, and instructional videos. Later reporting also described dry runs, demonstrations, a webinar, and practice access before evaluation began. Those details matter because they show that the institutional question is not whether preparation existed at all. The harder question is what kind of preparation a digitized public workflow requires.

After the results, students and observers raised concerns about missing pages, missing supplementary sheets, blurred images, incorrectly mapped answer books, and the ability of students to verify what had actually been evaluated. CBSE later opened an online window from June 2 to June 6, 2026 for students who had obtained scanned answer books to apply for verification of scan-related issues and/or re-evaluation. The reportable scan issues included missing pages, missing maps or graphs, blurred pages, incorrect answer books, and evaluation against a different set.

The CBSE case also moved upstream into procurement. Sarthak Sidhant, a Class XII student, published a detailed comparison of tender documents and alleged that successive tender revisions softened a blacklisting clause, removed or diluted past-performance disqualifications, and lowered the required software-quality standard in ways that could have favored Coempt EduTeck. The Week reported that he presented these concerns before a Parliamentary standing committee, while The Indian Express reported that CBSE submitted its own response and said portal glitches had been rectified with additional time for re-evaluation applications.

The vendor and the board also pushed back. The Economic Times later reported that Coempt EduTeck defended the scanners used in the system as industry-grade and said records were open to scrutiny. The Times of India reported that CBSE described first-phase re-evaluation results as having been processed through a robust system, even as the board was consulting stakeholders before deciding whether to continue on-screen marking in 2027. A later Times of India report said CBSE stated it had cleared more than 99.7 percent of Class XII verification and re-evaluation cases while disputing a separate claim by Vedant Shrivastava about his own re-evaluation.

The reported error counts are therefore analytically awkward in the right way. About 20 answer-book mix-ups, 68,000 rescans, and more than 13,000 manually evaluated copies are small relative to nearly 10 million answer books. They are also not trivial for the students whose records were affected. High-stakes public systems are not judged only by aggregate error rates. They are judged by whether the institution can identify error, preserve evidence, explain authority, and correct harm without making the affected person carry the whole burden.

The CBSE case should therefore not be treated as an adjudicated failure or a settled case of wrongdoing. The factual record is still mixed: official releases, student allegations, vendor defense, committee attention, news reporting, and official response are all part of the story. The purpose here is not to decide legal responsibility. It is to see what the case reveals about digitized public systems under stress.

The relevant questions were not only whether scanned evaluation was faster or cheaper. They were whether the board had specified the system well, whether the vendor relationship preserved public control, whether students could see and contest the relevant record, whether error categories were anticipated, and whether redress had been designed before failure became public.

The system was digital. The stress points were institutional.

Figure 1: A digital workflow does not only change the medium of evaluation. It changes where judgment, evidence, authority, and recourse sit.

Human Judgment Still Needs Institutional Conditions

Digitization often presents itself as a way to remove friction from public systems. The older process is slow, dispersed, paper-based, and difficult to monitor. The newer process is faster, centralized, digital, and easier to standardize.

Often this is true. But efficiency is not the same as operational capacity.

Digitization is not only the conversion of paper into pixels. It requires complementary assets: reliable records, stable identifiers, clean metadata, trained users, role permissions, audit logs, exception categories, procurement standards, and grievance pathways.

Firms discover this when they try to add AI to messy internal workflows and realize that their institutional knowledge is scattered across emails, spreadsheets, local databases, and memory. Public agencies discover it when a digital workflow has to carry legal authority, citizen trust, and public accountability at scale.

Digital-government standards increasingly make the same point in institutional language. The UK’s Service Standard, the OECD’s Digital Government Policy Framework, and the World Bank’s GovTech Maturity Index all treat public digital systems as more than portals. They ask whether services are usable, secure, interoperable, operated well, and supported by operational capacity.

An on-screen marking system may reduce some errors while making other errors harder for ordinary users to see. A missing page in a physical answer book is one kind of problem. A missing page in a scanned bundle is another. A blurred image on a portal may still carry the authority of an official record. A digital interface can make a process look orderly even when the underlying chain of custody is contested.

Human judgment remains important, but the conditions under which judgment is exercised have changed. The public agency can defend the process only if it has retained enough control over scanning, indexing, logging, vendor performance, evidence preservation, and redress.

The bottleneck moves from marking to verification. Digital workflows can scale execution quickly, but recourse remains slower and more labor-intensive. The institution still has to prove that the record is complete, legible, traceable, and open to correction.

Digitization therefore does not remove institutional work. It redistributes it. Responsibilities that once sat with physical records, clerical checks, moderation routines, and inspection now sit with procurement specifications, workflow design, audit logs, portal architecture, and digital evidence.

Institutional judgment is relocated into the design of the workflow: what counts as a valid scan, what gets flagged before evaluation, who can see the original record, how quickly a student can object, what evidence is preserved, and who is accountable when the digital record differs from the material reality it was supposed to represent.

Software did not create this problem. James Scott showed how states govern by making society legible: through categories, records, and simplifications. Digital systems do not abolish those conditions. They formalize them, accelerate them, and sometimes make them harder to contest.

The institutional problem is not downstream of the technology. It is inside the technology’s use. The same software can produce different outcomes depending on the surrounding institution. In one setting, it may become a disciplined evaluation workflow with strong auditability and credible recourse. In another, it may become a fast and opaque system whose errors are visible only to those with the time, literacy, and confidence to challenge it.

Algorithms Formalize Institutional Assumptions

The CBSE case did not involve AI. Algorithmic systems reveal the same pattern more sharply because the institutional assumption is no longer only inside a workflow. It is also inside the model.

Institutional assumptions are the categories, proxies, and background judgments an organization treats as good enough for action: what counts as need, risk, eligibility, evidence, or success.

In the United States, a widely used health-care algorithm was found to underestimate the needs of Black patients. The problem was not that the model had been instructed to discriminate. It had used health-care spending as a proxy for medical need. Because Black patients had historically received less care for the same level of illness, lower spending made them appear less sick to the system.

The technical problem mattered. But the deeper issue sat in the proxy and the setting that made it plausible.

The model converted a social and institutional history into a decision rule. Unequal access to care became lower recorded expenditure. Lower recorded expenditure became lower predicted risk. Lower predicted risk meant fewer Black patients were identified for additional care-management support. For an affected patient, the problem was not experienced as an abstract proxy error. It meant being less likely to be routed toward extra help.

The problem was detected because researchers compared the model’s risk scores with direct measures of illness rather than accepting spending as the ground truth. When the objective shifted toward predicting health needs instead of future costs, the racial disparity fell sharply. The repair was not only technical. It required asking what institutional assumption the proxy had carried into the model.

In both examples, technology becomes consequential through the categories, proxies, workflows, and accountability structures around it. While the CBSE case asks whether an institution can specify, monitor, audit, and correct a high-stakes digital workflow, the health-care algorithm adds a sharper lesson: AI can turn an institutional assumption into an operational routine. A proxy choice becomes an allocation pathway, and the assumption becomes harder to contest once it is embedded in the model.

Categories Decide What Counts as Governance

Whether a system is described as digitization, automation, artificial intelligence, or public infrastructure is not just a semantic issue. The category determines what kind of scrutiny the system receives. It shapes which experts are invited, which rules are applied, which failures are anticipated, and which forms of accountability become visible.

Institutions often govern categories before they govern technologies.

The CBSE dispute shows how much follows from the category. CBSE introduced on-screen marking as a logistics and efficiency upgrade. Students, observers, and parliamentary attention made it legible as something larger: a high-stakes digital evaluation and procurement system. Under the first category, the main questions are speed, cost, and administrative convenience. Under the second, different questions appear: scan integrity, data custody, vendor accountability, examiner workflow, appeal design, and the evidentiary rights of students.

A logistics upgrade is judged by efficiency and cost. A governance mechanism also has to be judged by legitimacy, equity, and recourse.

The system need not be AI to require digital governance.

Similarly, if a health-care algorithm is treated only as an efficiency tool, its proxy choices may appear technical. If it is treated as an institutional allocation mechanism, those same choices become questions of fairness, access, and responsibility.

Public debate often begins after a system has already been categorized too narrowly. By then, the institutional questions have been framed as implementation details.

A system’s category is already a policy decision. It allocates authority, defines the problem, and determines whether a failure is treated as a software issue, a procurement issue, a rights issue, a capacity issue, or a governance issue. In practice, it is often all of these.

The Operating Layer Has Four Parts

The institutional life of technology usually becomes visible in four places: procurement, workflow, oversight, and recourse.

Procurement decides what the public agency has actually bought. It determines the vendor’s obligations, the standards to which the system is held, the audit rights retained by the state, the security expectations, the ownership of data, and the remedies available when performance breaks down.

Workflow decides how the technology enters routine practice. It determines who uses the system, what they see, when they can intervene in real time, what gets logged, and which exceptions are escalated. A workflow can make human judgment more disciplined. It can also make human judgment more dependent on what the system chooses to display.

Oversight decides whether the institution can look back at what the system did after deployment. It determines whether logs are useful for review, whether failures are sampled or investigated, whether audit rights can be exercised, and whether performance claims can be tested against real use.

Recourse decides whether affected people can challenge the system in a meaningful way. It determines whether a student can see the relevant record, whether a patient can understand the basis of a decision, whether a citizen can correct an error, and whether the institution can repair damage without forcing the burden entirely onto the user.

Procurement, workflow, oversight, and recourse are not external safeguards added after deployment. They are part of the technology’s institutional form. Many policy debates become too thin because they ask only whether a technology should be adopted, regulated, banned, subsidized, or scaled. Those are important questions, but they are not enough. The same technology can produce different outcomes depending on the contract, the workflow, the audit process, and the grievance system around it.

The operating layer is where digital governance becomes real. Accountability, learning, and the distribution of authority are decided here, often before the public debate has named them. A public agency may retain formal authority while operational knowledge shifts to a vendor. A citizen may retain a formal right of appeal while the evidence needed to exercise that right sits inside a portal. An organization may announce a feedback process while the system lacks a routine for turning failures into revised standards.

The institutional problem stops being abstract at this operating layer.

Operational Capacity Is AI Capability gives this ability its more precise name: operational capacity. It is the capacity to perform this institutional work repeatedly. A capable institution does not merely buy a system. It knows how to specify what it needs and evaluate what it receives. It can test behavior, monitor failures, preserve evidence, revise workflows, and build credible recourse without surrendering judgment to the vendor, the interface, or the claim of efficiency.

Without that capacity, an institution may adopt the same system and still lose control of the function it was trying to improve.

The difference is not the technology alone. It is the surrounding operational capacity to make the technology answerable to public purpose.

IndiaAI Makes Operational Capacity Strategic

Public systems have always relied on categories, records, procedures, and delegated judgment. What is new is the pressure. AI systems can classify, rank, predict, and recommend at scales that make ordinary administrative review difficult. They can also make dependence harder to see, because the system presents itself as a usable interface rather than as a chain of assumptions, vendors, data, incentives, and constraints.

India’s AI policy ecosystem is trying to operate across this chain rather than only at the level of adoption. The IndiaAI Mission is a serious attempt to build multiple layers at once: compute capacity, datasets, application development, skilling, startup financing, and Safe and Trusted AI are all part of the same mission architecture. Later official releases on the IndiaAI Compute Portal and AIKosha and an AI Safety Institute add operational texture. The Mission is addressing real access constraints around compute, datasets, models, and safety tools. These are necessary foundations. But access expands the opportunity set; operational capacity determines whether those assets become durable capability inside real institutions. The institutional question is whether those assets are surrounded by standards, testing routines, procurement templates, incident reporting, and public-sector practice.

The eight Responsible AI projects selected under the Safe and Trusted AI pillar should be read inside this larger architecture, not as the whole Mission. Policy direction becomes operational capacity only when tools for audit, privacy, bias mitigation, explainability, certification, and governance testing enter the routines of hospitals, school boards, welfare agencies, regulators, and public procurement teams. That is the difference between passive procurement and operational capacity. A safety tool that remains a research output is not the same as an assurance system that institutions can use repeatedly under pressure.

The bridge between the CBSE case and IndiaAI is not that every digital workflow is secretly an AI system. It is that AI will make the same institutional questions harder, more urgent, and more consequential.

A ministry can procure AI tools without developing the capacity to audit models, evaluate proxies, protect data, or design recourse. A university can deploy automated systems without understanding how they reshape teaching, assessment, and student trust. A hospital can use prediction without asking how past inequality enters the data and how present judgment should respond.

The issue is not whether institutions use technology. They already do. The question is whether the work around that use becomes explicit, repeatable, and accountable.

For middle powers and developing societies, this is especially important. They are often encouraged to adopt frontier tools quickly, build digital platforms, and demonstrate technological modernity. But the harder work is less visible. It lies in procurement competence, technical state capacity, public-sector learning, standards, auditability, domain expertise, and credible channels of correction.

The strategic question is whether institutions can turn technological opportunity into durable capability. Operational capacity, sustained over time, is one way capability formation becomes real.

Figure 2: Technology becomes durable capability only when institutions can specify, oversee, contest, and repair its use.

Technology Is Only as Capable as the Institution Around It

The lesson from the CBSE case is not that digital evaluation should be rejected. The lesson from the health-care algorithm is not that prediction should never be used. The lesson from IndiaAI is not that policy missions are insufficient because they do not solve everything immediately.

The lesson is that technology reorganizes the institutional problem: where judgment is exercised, who can see error, how responsibility is distributed, and what counts as evidence.

When operational capacity is weak, technology can make a system look more modern while making accountability more fragile. When operational capacity is strong, technology can make a system faster, more legible, and more reliable.

The central question is not simply what the technology can do. It is what the institution is capable of doing with it.

Efficiency can be purchased. Technology can be procured. Judgment has to be built.

A note of Caution on the CBSE case: the OSM controversy is still evolving. This essay does not adjudicate legal responsibility, vendor liability, or the full factual record. It uses the case to examine what digitized public systems reveal under stress.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Sources and Case Materials

• CBSE circular introducing On-Screen Marking for Class XII answer books

• CBSE press release on verification of scanned answer-book issues and re-evaluation

• Sarthak Sidhant’s tender-document analysis of the CBSE OSM system

• The Week on Sarthak Sidhant’s Parliamentary-panel appearance

• Times of India on CBSE’s response to Vedant Shrivastava’s re-evaluation claim

• Ziad Obermeyer et al. on racial bias in a health-care algorithm

• IndiaAI official releases on the Mission, AIKosha and Compute Portal, AI Safety Institute, and Responsible AI projects

Further Reading

• UK Service Standard and Technology Code of Practice

• OECD Digital Government Policy Framework

• World Bank GovTech Maturity Index

In April 2026, India’s National e-Governance Division listed an empanelment for AI/ML manpower augmentation and AI-driven project delivery under the Digital India programme. The underlying request for empanelment set out the more administrative part: a mechanism through which pre-qualified agencies could supply defined AI/ML roles at standardized rates to ministries, state governments, and public-sector undertakings.

A procurement pathway like this rarely receives the attention given to frontier models or national AI missions. But it points to a deeper distinction. A single AI pilot puts a technology into use. A reusable pathway can also help institutions retain what they learn from repeated deployment.

Access comes first. A country can have access to a technology without being capable in it. It can buy cloud credits, import machines, license software, subsidize compute, or empanel vendors. None of this is trivial. Without access, learning cannot begin.

Adoption is the next step. Technology enters firms, agencies, hospitals, schools, laboratories, and public systems, and begins changing workflows and routines rather than remaining a procurement line or a demonstration project.

Capability formation is slower: the accumulation of the human, institutional, infrastructural, and organizational base that lets a society keep using, adapting, repairing, governing, and improving a technology as the technology itself changes.

Access can be rented. Adoption can be subsidized. Capability has to compound.

Capability forms when people acquire judgment, institutions preserve memory, and the surrounding assets needed to adapt, govern, and improve a technology assemble over time. This accumulation runs through firms, universities, public agencies, laboratories, and entrepreneurial ventures, not through the state alone.

This slow accumulation is easy to mistake for access or adoption because it becomes visible later. The NeGD pathway will build capability only if repeated procurement also produces better judgment, stronger teams, reusable standards, and institutional memory.

Adoption Is Not Depth

The easiest way to misread technological progress is to count visible use. How many firms are using AI? How many students have access to coding tools? How many startups received GPU credits? How many hospitals tested diagnostic systems? How many ministries launched pilots? Adoption data tells us whether a technology is spreading.

Adoption data does not tell us whether capability is accumulating. A firm can use AI through a vendor system without learning how to evaluate it. A hospital can deploy a diagnostic assistant without building the data infrastructure that would let it improve over time. A ministry can buy a platform without developing procurement, audit, or escalation capacity. A university can run AI workshops without changing how students acquire judgment. A country can subsidize access to compute while most of the learning, ownership, and infrastructure depth remain elsewhere.

The adoption metric sees the technology in use. It often misses the institutional surround.

Economists studying general-purpose technologies have long made a related point. Erik Brynjolfsson, Daniel Rock, and Chad Syverson argue that transformative technologies often require complementary investments before their productivity effects appear. The technology arrives before organizations know how to reorganize around it. In the interim, visible adoption can run ahead of measurable capability.

AI sharpens this pattern because the visible layer moves quickly. A model can be adopted in a week. A procurement contract can be signed in a month. A workflow can be changed in a quarter. But the deeper capabilities that make adoption compound often form over years: clean data systems, managerial competence, domain expertise, professional training, evaluation routines, infrastructure reliability, and institutional memory.

Visible adoption moves quickly. Capability formation does not.

Figure 1: Access, adoption, and capability formation unfold on different timescales.

Capability compounds only when human learning survives automation, institutions retain knowledge across project cycles, and complementary assets turn both into repeated use. The next three sections take these pressures in turn: skill formation, institutional retention, and the surrounding assets that let learning compound.

The Skill Ladder Problem

The first mechanism is skill-ladder compression. Many professions produce senior judgment through junior work. Junior lawyers review documents, draft memos, and observe litigation strategy. Medical residents perform supervised diagnostic work and gradually build clinical intuition. Junior analysts clean data, prepare models, write summaries, and learn to structure analysis and communicate results through repeated practice. Software engineers often learn by writing routine code, debugging mistakes, reading existing systems, and absorbing architecture through repetition. These tasks are not glamorous, but they are developmental infrastructure.

AI intensifies a familiar automation pattern: it reaches into entry-level cognitive practice. Drafting, screening, summarizing, classification, first-pass coding, document review, translation, and routine analysis are now targets for automation, acceleration, or partial substitution. Compared with doing the same work without AI, these tasks can require less time and labor. They are also the tasks through which junior workers begin to develop senior judgment.

Automating entry-level cognitive tasks can raise productivity while compressing skill ladders. If junior workers are hired less often because routine tasks are automated, fewer people accumulate the tacit knowledge that becomes senior expertise. If junior workers are hired but asked mainly to supervise AI outputs, they may skip the messy practice through which judgment forms. If senior professionals use AI to produce more output while junior pipelines narrow, the short-term organization may become more efficient while the long-term profession becomes less capable.

In a large customer-service field study, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that generative AI delivered its largest performance gains to novice workers, partly by transmitting practices from stronger peers. But a recent study of AI-assisted coding found measurable productivity gains for experienced developers, not early-career developers, raising the possibility of a wider skill gap. The institutional question is whether AI becomes a learning scaffold or a substitute for the practice through which judgment forms.

Further, the costs and benefits are not evenly distributed. A firm may benefit from reducing junior hiring costs today while the profession, university system, or wider economy bears the long-term cost of a thinner training pipeline. Senior workers may gain leverage if AI amplifies their productivity while entry-level pathways narrow. Educational institutions may be asked to produce practice-ready graduates without the workplace apprenticeships that used to complete the training. Skill formation is therefore not only a design problem. It is a distributional conflict over who pays to preserve the learning curve.

Historically, automation often shifted work rather than simply eliminating it. ATMs reduced routine cash handling, but branch work moved toward advisory and relationship roles. Generative AI creates a different training risk because it can remove parts of the cognitive practice layer through which senior judgment forms. If junior workers move too quickly from producing first drafts to supervising machine outputs, they may lose the baseline craft required to know when the output is wrong.

The harder skill-ladder problem is that higher-order judgment has to come from somewhere. Deming and Noray’s work on STEM careers shows that fast-changing technical fields can make skills obsolete more quickly, changing the returns to experience. AI adds a related pressure from the other direction. If the entry-level tasks through which experience accumulates are automated before new training systems emerge, the pipeline of expertise can weaken.

A country should therefore treat skill-ladder compression as more than a labor-market issue. A society cannot build AI capability only by giving people AI tools. It has to redesign the pathways through which people learn to use, question, improve, and govern those tools. That may mean new apprenticeship models, supervised AI use in professional education, simulation environments, evaluation-heavy training, and institutional roles that deliberately preserve learning opportunities even when AI reduces the time and labor required to complete some entry-level tasks.

The policy world is beginning to notice this, though the institutional response is still young. The India AI Impact Summit 2026 outcomes included an Equitable AI Transition Playbook developed with the International Labour Organization and voluntary guiding principles for reskilling in the age of AI. These are signals, not solutions. They show that workforce transition is moving from generic upskilling rhetoric toward a harder question: how to protect human learning systems while firms adopt tools that bypass parts of the entry-level practice on which those systems relied.

Without redesigned learning pathways, AI adoption may hollow out the very skill base it depends on.

Capability Dispersal

Even when skill ladders survive, capability can disperse when the people and routines holding their knowledge leave. Capability dispersal is the loss of accumulated judgment, routines, and institutional memory even when the visible system remains in place. Capability does not sit only in equipment, software, or individual talent. It sits in teams, tacit knowledge, relationships, documentation, procurement habits, maintenance practices, and the accumulated memory of what has failed before. Project-cycle funding often produces fragile capability because it builds the visible object faster than the institution around it.

A grant funds a data platform. Consultants build a system. A pilot hires technical staff. A lab acquires equipment. A ministry creates a short-term innovation unit. For a few years, the system appears capable. Then the project ends. Staff move. Contracts lapse. Documentation goes stale. Software is not updated. Procurement knowledge dissipates. The equipment remains, but the institutional memory evaporates. The capability was never the equipment alone.

Many systems compensate for this fragility through exceptional individuals. A capable public official, founder, lab head, engineer, procurement officer, or project director holds the system together through judgment, relationships, urgency, and memory. The system works while that person is present. It weakens when the person leaves.

Exceptional people matter, but they are not a capability strategy. The harder institutional question is what remains when initiative has to survive succession, budget cycles, political turnover, vendor changes, and technical drift.

An International Science Council technology profile on data infrastructure in Global South science systems makes this pattern visible across science systems. Durable data infrastructures depend less on the initial procurement of storage and more on five quieter conditions: a mandate or policy anchor, governance architecture before technical design, institutional embedding in national systems, lightweight open standards, and funding continuity. The lesson travels beyond data repositories.

Funding continuity matters more than funding spectacle. A smaller system embedded in a national agency, with stable staff, clear authority, common standards, and recurring resources, may accumulate more capability than a larger system funded as a short-lived project. The difference is not visible at launch. It becomes visible when the first staff leave, when the first integration breaks, when standards need revision, when data volumes grow, or when a new technology has to be absorbed.

DIRISA, South Africa’s Data Intensive Research Initiative is a useful example of the quieter model. DIRISA is not interesting because it offers a spectacular technology object. It is interesting because it is embedded in the country’s national cyberinfrastructure system, alongside high-performance computing and research-network capacity, rather than standing alone as a short project grant. That kind of embedding does not guarantee success, but it reduces the most common failure mode: infrastructure built as a project, then abandoned as the people and funding that made it work disperse.

Capability dispersal is especially important for middle powers and Global South institutions. Where technical labor markets are deep and well-funded, departing staff can often be replaced. Where institutions are thin, specialized knowledge is harder to rebuild. A trained engineer, data steward, procurement specialist, lab manager, or clinical informatics lead may carry a disproportionate share of capability. When that person leaves, the system may still exist on paper while its effective capacity has already weakened.

Visible infrastructure can therefore be misleading. A data center, AI lab, robotics testbed, digital platform, or research facility may signal capability. But the durable capability lies in whether the institution can staff it, update it, govern it, connect it to users, finance it through cycles, and learn from its failures.

The launch moment reveals what was purchased. Maintenance reveals what was actually built.

Figure 2: Capability compounds when learning and institutional memory persist across time.

Institutional Complementarity

Retaining skill and institutional memory is still not enough if the surrounding assets never assemble. David Teece used the phrase “complementary assets” to explain why firms do not profit from innovation through invention alone. They need manufacturing, distribution, marketing, service networks, standards, and other assets that allow an invention to become valuable. Douglass North made the broader institutional point: formal rules and informal norms shape whether actors have incentives to invest, coordinate, trust, and adapt.

AI follows the same logic. A model is more useful when the organization around it has clean data, domain expertise, cloud access, trained staff, workflow redesign, evaluation routines, legal clarity, and managerial capacity. Without those complements, AI adoption may produce demos, scattered pilots, or superficial productivity gains.

Consider a hospital adopting an AI diagnostic system. The visible moment is procurement: the hospital buys or licenses the tool. The public story is that AI is entering healthcare. But whether the system becomes useful depends on older and slower capabilities. Are patient records digitized and coded consistently? Can IT staff integrate the tool with existing systems and maintain the data pipeline? Do clinicians have the training and time to use it? Can the hospital validate performance in its own patient population? Does the procurement budget cover integration, updates, and maintenance after purchase?

The AI system arrives late in the story. The capability formation began years earlier, in electronic health records, coding standards, staff training, procurement choices, legal rules, hospital management, and clinical governance. The model may be new. The capability that makes the model useful is accumulated.

Complementary assets explain why adoption gaps persist across firms and sectors. Some organizations are ready to absorb AI because they have already made complementary investments. Others can access the same tool and get far less value from it. A multinational firm with structured data and in-house engineering teams can integrate AI differently from a small firm relying on off-the-shelf tools. A well-managed hospital with digital records can use diagnostic support differently from a clinic still dependent on paper files. A ministry with procurement and audit capacity can deploy AI differently from an agency buying vendor systems it cannot inspect.

Technology spreads unevenly because capability is unevenly distributed. The development problem is therefore not only access. It is absorptive capacity: the ability to recognize, adapt, use, and improve a technology inside a real organization.

Across these settings, the test is the same. Firms have to absorb technology into production; public agencies into accountable workflows; universities into training systems; science systems into research infrastructure; regulators into enforcement capacity; and labor markets without destroying the pathways through which expertise forms.

Questions like these unfold on a slower timescale than adoption questions. They are also more strategic.

Complementarity becomes clearer when AI leaves the screen. Robotics, sensors, drones, laboratory automation, and other embodied technologies require more than models and code. Simulation helps, but it does not eliminate the reality gap: physical systems encounter noise, latency, weather, surfaces, power limits, human proximity, and failure modes that virtual environments cannot fully absorb. A robot control model can be downloaded quickly. Safe test environments, maintenance routines, and operational safeguards cannot.

What Middle Powers Must Build

Capability formation is a general problem, but it is especially visible in middle powers. Frontier economies enter new technology cycles with accumulated institutional depth that can look like a natural technological advantage. Dense research universities, venture networks, supplier ecosystems, legal services, standards bodies, procurement markets, deep capital pools, and specialized labor markets were assembled over decades. When a new technology arrives, many complementary capabilities are already nearby.

Middle powers cannot assume that inheritance. Cloud computing, open-weight models, modular hardware, global APIs, and international research networks lower the cost of entering and experimenting in a technological field without reproducing the whole stack domestically. But the ability to enter a technological field is not the same as the ability to keep learning within it.

The strategic task is threefold: use external systems where access is necessary, build local depth where dependence would block learning, and turn deployment into a mechanism for accumulating judgment, talent, procurement capacity, evaluation routines, and institutional memory. An open-weight or externally developed model can therefore become a capability-building input when local institutions adapt it, evaluate it against local conditions, and retain what they learn through deployment.

The NeGD empanelment is one Indian test of this approach. Pre-qualified agencies, defined roles, and standardized rates can reduce the friction of acquiring AI expertise. The capability effect depends on what assignments leave behind: reusable documentation, stronger public teams, common evaluation routines, and knowledge that survives the contract.

The alternative is not autarky. No serious middle-power strategy can build everything domestically. The question is which capabilities must be built at home because outsourcing them would prevent learning, bargaining, evaluation, or future adaptation.

Stack-building can begin without capability automatically compounding. Data centers, model partnerships, semiconductor facilities, compute portals, and procurement pathways matter because of what forms around them: skills, suppliers, standards, evaluation routines, maintenance practices, and institutional memory. The test is whether the asset creates conditions for repeated learning.

For AI, that might mean domestic evaluation capacity that includes red-teaming and safety testing, local-language data systems, public procurement expertise, sectoral testbeds, low-cost deployment engineering, clinical and legal audit routines, compute access strategies, and training systems that preserve skill formation under automation. It may also mean procurement architectures that prevent every public agency from relearning the same deployment problem alone.

For data infrastructure, capability may mean standards, mandates, repositories, data stewards, and funding continuity. For robotics, it may mean safety protocols, maintenance technicians, local integration firms, controlled test environments, and domain-specific deployment knowledge.

The details vary by technology. Some technologies travel more easily than others, create learning faster, or require fewer complementary assets. Durable capability still depends on whether local institutions can keep learning after first adoption.

The three mechanisms now in view reinforce one another: skill-ladder compression, capability dispersal, and institutional complementarity. If skill ladders compress, fewer people accumulate the tacit judgment that becomes institutional memory. If institutional memory disperses, complementary assets are harder to operate and improve. If complementary assets are missing, adoption produces less learning, which weakens the next generation of skill formation. Capability formation is therefore not a checklist of inputs. It is a compounding system.

Capability forms where institutions create conditions for repeated learning. Systems have to be used, evaluated, repaired, adapted, and improved across time. People have to stay, institutions have to remember, failures have to be recorded, and local actors have to move from using a technology to modifying the conditions under which it is used.

Entrepreneurs often act as translators between capability layers: science and application, public missions and private products, infrastructure and use cases, local problems and global tools. But their learning compounds only when finance, procurement, standards, testing environments, research institutions, patient customers, and credible demand can sustain it.

Startups can accelerate learning. They cannot substitute for the system that lets learning accumulate.

What Compounds

The fundamental question in capability formation is what compounds. Does a deployment create learning that stays inside the institution? Does a training program produce judgment or only tool familiarity? Does infrastructure become embedded in governance or remain a project asset? Does adoption deepen local bargaining power or increase dependence on external providers?

Policy can widen access and shape diffusion through public infrastructure. Durable technological agency depends on whether the human, institutional, infrastructural, and organizational base keeps deepening after the first wave of use.

Capability formation is what turns access into agency, deployment into learning, and exceptional initiative into institutional memory.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Disclosure: The linked International Science Council technology profiles on data infrastructure and robotics were commissioned by the Council and co-authored by Venkat Nadella and Chandan Nagarajappa.

Earlier Essays

• The Stack Beneath the Interface — why AI access rests on a deeper and unevenly controlled stack.

• Operational Capacity Is AI Capability — why AI capability depends on the institutions that procure, evaluate, and maintain it.

• Digital Public Infrastructure and Its Limits — where public rails can coordinate AI adoption, and where they cannot.

Further Reading

• NeGD request for empanelment for AI/ML manpower augmentation and AI-driven project delivery

• Brynjolfsson, Rock, and Syverson on AI and the modern productivity paradox

• Brynjolfsson, Li, and Raymond on generative AI at work

• Deming and Noray on STEM careers and changing skill requirements

• Daniotti, Wachs, Feng, and Neffke on generative AI and software-development careers

At the 2026 India AI Impact Summit, the government launched the Global AI Impact Commons, a voluntary initiative for sharing, replicating, and scaling AI use cases across countries. The launch drew on a tempting idea from India’s digital public infrastructure success: perhaps AI, too, can be governed through public rails. If shared digital infrastructure helped identity, payments, documents, and data flows operate at population scale, perhaps a similar public architecture could help AI diffuse more widely.

The intuition is useful, but it can mislead. DPI can organize how AI adoption happens. It cannot by itself create the industrial depth, compute, or institutional capacity required for AI deployment and frontier competition. Coordination and capability are different problems.

Public rails are most useful when AI enters public services. They can coordinate identity, consent, records, audit, redress, and workflow across agencies and vendors. They cannot resolve the deeper constraints of models and the infrastructure beneath them: compute, cloud, chips, energy, and data centers. Nor can they replace the institutional judgment required to evaluate AI systems. Public rails can still support downstream learning, but they are not a substitute for the deeper stack.

India’s DPI experience shows what coordination infrastructure can do. Aadhaar made identity machine-verifiable. UPI made payments interoperable. DigiLocker, account aggregators, and related data-exchange systems extended the same logic into documents and consent-based flows. These systems created rails that reduce transaction costs, expand participation, and let other actors build services on top.

AI, however, is not only a coordination problem. It is also an industrial, infrastructural, and operational problem. Its bottlenecks include the infrastructure beneath models, from compute and chips to energy and cloud capacity. Others sit in the institutional capacity to evaluate AI-mediated decisions.

DPI is not a template to copy into AI. It is a steering instrument: useful for organizing adoption and participation, but unable to remove deeper infrastructure bottlenecks. Its power lies in that specific role, and so do its limits.

What DPI Actually Solved

Digital public infrastructure is now a global policy term. The World Bank describes DPI as foundational digital systems such as identity, payments, and secure data exchange that support digital services across the public and private sectors.

The term is newer than the problem it names. Information-infrastructure scholars have long studied how shared technical systems become background conditions for social and economic life. Susan Leigh Star and Karen Ruhleder’s canonical work on the ecology of infrastructure is one older anchor for this way of thinking: infrastructure is not just equipment, but a relational system that works only when it is embedded in practices, users, standards, maintenance, and organizational routines. DPI is a newer policy name for a related institutional problem.

India made this problem visible at a scale few countries had attempted. Aadhaar created a digital identity layer. UPI created interoperable instant payments. In May 2026 alone, UPI processed 23.2 billion transactions. Other systems built around document exchange, consent, and service delivery extended the logic. Their strategic significance lies in the kind of problem they solved: coordination.

Before UPI, digital payments were fragmented across banks, wallets, cards, and settlement systems. UPI created a common protocol that let different banks and payment apps interoperate above that fragmentation.

UPI did not appear from nowhere. It assembled and exposed underlying primitives, including bank payment rails such as IMPS, through a public platform layer that developers and applications could build around. That is the distinctive DPI move: not inventing every underlying technology, but turning available primitives into a coordination architecture.

The BIS study of India’s digital financial infrastructure is useful because it does not treat this as only a payments story. It frames India’s approach as digital financial infrastructure provided as a public good: identity, payments, and data-sharing architecture layered so that innovation can happen above shared rails. The achievement was not just a popular app or a transaction volume milestone. It was a design architecture that changed who could connect, transact, verify, and build.

DPI’s coordination logic explains why it became attractive in development circles. If identity, payments, and data exchange become interoperable, then other services can be layered on top: welfare payments, financial inclusion, health and education credentials, agricultural support, and private innovation. The IMF’s working paper on India’s digital journeymakes this argument in developmental terms: India Stack supported inclusion, competition, market expansion, and public-expenditure efficiency by providing foundational digital infrastructure. The mechanism is institutional coordination through digital rails. DPI worked because the binding constraint was coordination, not silicon.

Silicon and global technology systems still mattered to UPI, but they were not the constraint the public rail had to solve.

India did not need to own frontier semiconductor fabrication to scale UPI. It did not need globally dominant cloud companies to make QR-code payments interoperable. The deeper layers needed for digital coordination, including mobile devices, telecom networks, basic cloud capacity, and consumer internet adoption, were becoming accessible enough for the coordination layer to matter. DPI did not eliminate dependence on global technology systems. It used available technology layers to build a domestic coordination architecture on top.

DPI also operated in a favorable cost regime. Once the rail existed, each additional payment message or identity check could move through a standardized digital process at very low marginal cost. The hard part was getting banks, agencies, standards, users, and service providers to coordinate around the same rails.

AI is different because the stack and the cost structure change together. Frontier training requires concentrated, capital-intensive compute. Individual inference requests are cheaper than frontier training, but serving models repeatedly at scale creates continuing demand for compute, memory, electricity, cooling, networking, and cloud infrastructure. The IEA’s analysis of energy and AI emphasizes the data-center electricity constraint, while Microsoft Research’s work on AI inference energy shows how workload design shapes energy demand at inference time.

DPI made digital coordination cheaper over infrastructure that had become broadly accessible. AI, by contrast, makes repeated claims on scarce physical infrastructure. Every time a model is trained or queried at scale, supporting compute has to be powered, networked, and cooled; hosting and updating the system add recurring costs.

Figure 1: DPI coordinates many actors through shared rails; AI repeatedly draws on concentrated infrastructure.

Steering Without Owning

The most interesting thing about DPI is not digitization by itself. E-governance initiatives can digitize forms, portals, and service delivery. DPI does something more structural: it lays public rails and changes the terms on which banks, agencies, firms, and service providers operate.

When a government establishes or commissions an identity rail, it changes how banks verify customers, how welfare agencies authenticate beneficiaries, how telecom firms onboard users, and how private services establish trust. When it establishes a payment rail, it changes market entry for payment providers, merchants, banks, and app developers. When it establishes a data-exchange layer, it changes who can access which records, under what consent conditions, and through which interfaces.

DPI belongs in the broader family of infrastructure-as-governance moves. States do not only fix market failures after the fact. They also shape markets through investment, standards, procurement, information architecture, legal authority, and organizational coordination. DPI combines several of these tools at once.

DPI lets the state steer without owning every service built above the rail. That is its strategic importance for middle powers. A country may not control the deepest global technology stack. It may not own the devices, chips, operating systems, cloud platforms, or software ecosystems on which digital life depends. But it can still shape domestic participation by building shared protocols, registries, standards, and service interfaces. This is not full technological sovereignty. It is not passive dependence either. It is steering.

India’s DPI experience matters because it shows steering at population scale. The state did not build every fintech product or own every consumer-facing payment app. It built a common rail around which banks, fintech firms, merchants, and users could coordinate. The rail changed the market without replacing the market.

Coordination also redistributed leverage. UPI altered the bargaining position of banks, payment apps, merchants, regulators, and users by moving competition away from bilateral relationships and proprietary payment rails toward interfaces, data, customer acquisition, and compliance with a shared protocol. DPI applied to AI will create similar leverage questions. Public workflows, registries, audit interfaces, and procurement channels will not merely connect actors; they will shape where bargaining power accumulates.

Public rails therefore create politics as well as efficiency. They do not automatically produce public value. They create an institutional terrain on which public value has to be produced, defended, measured, and corrected.

Public Rails Need Public Governance

The DPI debate often gets pulled into two familiar stories. One story presents DPI as democratic infrastructure: open, interoperable, inclusive, participatory, and developmental. It gives countries a way to avoid dependence on closed private platforms, lets small firms build on shared rails, and gives the state leverage over service delivery without monopolizing every service itself. The other story presents DPI as digital control: identity systems become surveillance systems, consent becomes coercive, authentication failures become exclusion, and public infrastructure becomes a channel through which private firms harvest public value.

The democratic-infrastructure story can become too flat. DPI is not inherently democratic because it is digital, interoperable, or state-backed. A payment rail can expand access while still concentrating consumer-facing power in a few private apps. A digital identity system can reduce duplication while also creating exclusion when authentication fails. A data-exchange layer can support user control while becoming opaque if consent is reduced to a checkbox.

DPI is not inherently authoritarian either. Shared infrastructure can reduce dependence on proprietary platforms, lower entry barriers, improve portability, and make some services more accountable. The political question is not whether DPI is good or bad in general. It is what constraint DPI is being asked to solve, and what safeguards make the infrastructure genuinely public in practice.

The safeguards literature matters because public rails need enforceable obligations. The Universal DPI Safeguards Framework, developed through a UNDP and UN technology-envoy process, translates that principle into safety, inclusion, accountability, redress, governance standards, and lifecycle assessment. That framing is useful because it refuses both hype and rejection. It asks what must be true for public infrastructure to remain public.

India’s own experience shows why those obligations matter. Reetika Khera’s work on Aadhaar and welfare programmesargued that digital identity could produce exclusion when beneficiaries were denied entitlements because of authentication failures, linking problems, or administrative rigidity. The State of Aadhaar 2019 report found broad coverage and frequent use, but also documented errors, service denial, and unresolved problems for a minority of users. Aadhaar should not define the entire DPI category, but it shows the governance problem clearly: public infrastructure becomes public through the quality of its recourse, correction, and accountability.

DPI’s public-private boundary sharpens the problem. DPI is often described as “public rails for private innovation.” But public rails can still enable private concentration above them. The Bennett School’s review of DPI promises and realitiespoints to the governance, market-structure, and state-capacity questions that follow from this model. DPI is neither simply public provision nor ordinary privatization. It is a hybrid institutional architecture.

Hybrid infrastructure can support inclusion or produce exclusion. It can enable competition or concentration. It can strengthen public capacity or outsource public functions into systems that few people can contest.

DPI politics is about institutional design and maintenance. Standards need revision, authentication failures need correction, grievance channels need staffing, and users need recourse. A public rail becomes politically visible not when it works, but when it denies a benefit, misroutes a record, fails to authenticate a person, or leaves no one clearly responsible for repair. Public value has to be maintained after the infrastructure scales.

What DPI Can Do For AI

AI in public services creates a new coordination problem. Common rules for identity, consent, records, audit, escalation, and service workflow have to work across agencies and vendors. These are the institutional components DPI can help coordinate.

Several proposals now describe DPI applied to AI. The Global AI Impact Commons frames AI diffusion around shared use cases and reusable public-service deployments. The CDPI DPI-AI framework offers one explicit version through AI Blocks, DPI Workflows, and Public Agents: modular capabilities and auditable service recipes that connect AI systems to safeguards and human oversight. Neither is a universal model. Together, they show that the policy category is becoming real.

Three claims need to be kept distinct. First, AI can be used inside public-service workflows. Second, public rails can make those technology use-cases more reusable, auditable, and substitutable across agencies and vendors. Third, neither achievement is the same as frontier AI capability.

Fragmented pilots are the immediate problem public rails can address. A chatbot, recommender system, and eligibility tool can each arrive with a separate data pipeline, vendor contract, consent flow, and error-handling process.

Shared registries can stop AI systems from inventing their own versions of institutional truth. Consent and data-exchange rails can prevent every project from improvising data access separately. Reusable workflow templates can specify human review, confidence thresholds, appeals, and logs. Local-language services can connect private or public models to accountable public-service workflows instead of leaving integration to vendor-specific apps alone. Together, these are public rails for AI: shared scaffolding around public use.

In public-service workflows, DPI shapes the conditions under which AI touches society.

Repeated use of shared rails can build limited but real capability. Shared infrastructure can aggregate demand, create common data standards, make providers easier to substitute, and give agencies repeated experience with evaluation and repair. This downstream effect strengthens adoption and institutional learning without changing where frontier models, compute, cloud systems, and energy infrastructure come from.

The essay on operational capacity argued that procurement, evaluation, audit, contestability, escalation, and responsibility are scarce capacities in AI deployment. DPI can support those capacities if it is designed as accountable service infrastructure. A welfare agency using an AI eligibility recommender should not merely receive a score. The agency should know which registry was queried, whether the decision was automated or reviewed, how the citizen can contest it, and how repeated errors will be detected.

AI changes the character of the workflow. Classic DPI systems route bounded transactions: verify an identity, move a payment, exchange a document, authenticate a request. These systems can fail badly, but the operation itself is usually defined in clear terms. AI systems introduce probabilistic judgment into those rails. They classify, recommend, rank, and predict. Probabilistic judgment is not a faster version of a bounded transaction. It is a different kind of operation, and a DPI rail carrying it is doing something the rail was not originally designed for.

Samagra Vedika shows the risk in compressed form. The AIAAIC repository includes the Samagra Vedika incident, where database consolidation and algorithmic decision processes in Telangana were reported to have denied food rations to eligible citizens. The operational-capacity essay treats the case as a failure of evaluation, audit, and recourse. Here the lesson is narrower: once algorithmic matching enters a public workflow, a flawed record can become difficult for the affected person to contest.

A public-facing service may be digital, standardized, and formally rule-bound, while the matching logic inside the workflow remains difficult to inspect. If a system associates the wrong person, household, asset, or entitlement record, the person affected does not experience a technical classification. They experience a benefit denial.

The risk in DPI applied to AI is that public rails can make services interoperable while probabilistic systems make the judgment inside those services harder to see. The risk does not cancel the diffusion value. In a country like India, AI adoption will not happen only through frontier labs. It will happen through banks, schools, hospitals, municipal offices, and state portals. Shared rails can reduce adoption costs and give the state a way to set participation rules without building every application itself.

The DPI category travels better than the India Stack template. The transferable lesson is not “build India Stack for AI.” It is that shared infrastructure can become a steering instrument where a technology ecosystem is fragmented and public coordination can reduce participation costs.

Where DPI Reaches Its Limits

DPI reaches its limit when the binding constraint is no longer coordination. AI has coordination problems, but it is not only a coordination problem. The essay on the stack beneath the interface looked at GPUs, fabs, CUDA, cloud infrastructure, energy, data centers, and model ecosystems. Better public digital rails do not make those constraints disappear.

Coordination and the deeper AI stack are two different problems. Where the immediate constraint is fragmented public-service adoption, improvised workflows, weak contestability, or disconnected data exchange, DPI can do real strategic work. Where the constraint is access to compute, advanced chips, cloud infrastructure, energy, model ecosystems, or institutional evaluation capacity, the instrument has changed. A coordination strategy becomes a category error when it is mistaken for a capability strategy.

The Global AI Impact Commons is a clean stress test of the distinction. A repository of reusable deployments can reduce duplication, help agencies compare use cases, and spread common safeguards. But the model still has to be hosted, the inference bill paid, and updates supplied. A public API can localize access while leaving model behavior, pricing, cloud exposure, and export-control vulnerability governed elsewhere.

Figure 2: DPI can shape participation rules and deployment routines. It cannot by itself produce the deeper infrastructure.

The IndiaAI compute-access effort and DPI therefore belong to different strategic categories. Compute access lowers the near-term cost of using scarce infrastructure whose deepest layers remain externally controlled. DPI builds coordination infrastructure where the state can shape the rules of participation. Both can be useful. They are not the same move.

The strategic principle is simple: instruments work when they match the constraint. Fragmented systems need coordination. Exclusionary markets may need public rails that lower entry barriers. Accountability problems may need logs, redress pathways, and standardized workflows. Compute concentration, chip fabrication, and energy availability require other instruments. DPI may organize access, aggregate demand, and improve bargaining terms. It does not produce the chips, compute, cloud, data centers, or energy capacity beneath the rail.

Democratic governance follows the same principle. DPI can support consent, transparency, and public accountability, but participation is not produced by infrastructure alone. A consent layer is not democratic if refusing consent means losing access to essential services. An audit log is not accountability if no institution has the power or capacity to act on it. That is slow institutional work.

The Category, Not The Template

DPI’s most important contribution is not that it gives India a universal model for AI. It gives us a sharper way to think about technological statecraft. States do not only regulate technologies after markets form. They also shape the conditions under which markets, public agencies, firms, and citizens interact, including through shared digital infrastructure.

The right lesson from DPI is therefore not confidence or cynicism. It is constraint diagnosis. Before copying the DPI template into AI, a country has to ask what problem it is actually solving: fragmentation, market concentration, data access, institutional legitimacy, evaluation capacity, or physical infrastructure. These are not uniquely AI problems. What changes is which constraint becomes binding, how quickly systems scale, and how opacity or concentrated infrastructure alters the available instruments. Each answer requires a different instrument.

DPI is powerful because infrastructure can govern. Its strategic value lies in showing that states can shape technological systems without owning every layer beneath them.

The limit of DPI is where the harder work of capability formation begins: how institutions turn access and adoption into capability that compounds rather than disperses.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Further Reading

• World Bank, Digital Public Infrastructure and Services

• Susan Leigh Star and Karen Ruhleder, Steps Toward an Ecology of Infrastructure

• BIS, The Design of Digital Financial Infrastructure: Lessons from India

• UNDP/OSET, Universal DPI Safeguards Framework

• CDPI, DPI-AI Framework Paper

Between 2005 and 2019, the Dutch Tax and Customs Administration wrongfully accused approximately 26,000 families of fraudulently claiming childcare benefits. Many were forced to repay benefits in full. Families fell into debt. Some lost homes, jobs, and custody of children. In 2021, the scandal helped bring down the Dutch government.

The case is often remembered as a failure of automated decision-making. That is true, but incomplete.

The Dutch system used risk profiles to flag childcare-benefit claims for possible fraud. According to the OECD, families were flagged because of flawed data, minor administrative errors such as missing signatures, and indicators that disproportionately affected households with dual nationality or migrant backgrounds. Amnesty International described the risk-classification system as discriminatory and warned that nationality was used as a risk factor in fraud detection.

The deeper failure was operational. The risk system produced suspicion, but the tax administration lacked enough capacity to evaluate, contest, and correct how that suspicion moved through the agency.

That surrounding ability is operational capacity: the institutional ability to evaluate, contest, govern, and correct how an AI system works in practice.

The risk score did not remain an isolated technical output. It entered an administrative machine. Officials treated suspicion as administrative fact.

Families struggled to understand why they had been flagged. Appeals and corrections were slow, opaque, or ineffective. The tax administration had a system for generating risk, but not enough capacity for interpreting risk, challenging risk, or reversing harm once risk had hardened into action.

The failure became operational when a technical signal entered an administrative workflow and began shaping consequences. AI governance usually becomes real at this point: not when a model produces an output, but when an institution decides what the output means, who can challenge it, and what happens when it is wrong.

Where AI Actually Touches Society

Most AI debate begins with the model. The model is evaluated, benchmarked, released, criticized, celebrated, or feared. This is understandable: models are visible, named, measured, and easy to discuss as technical progress.

AI rarely touches society as a model alone. It touches society through hospitals, courts, schools, banks, welfare agencies, police departments, tax systems, call centers, farms, insurance firms, and public-service portals. These are not neutral delivery channels. They are institutions with routines, incentives, budgets, procurement rules, expertise, ways of reporting and learning from errors, and accountability structures. Those features shape what an AI output becomes once it enters the world.

The Stack Beneath the Interface looked beneath the interface and showed how AI access can depend on deeper infrastructure. This essay looks beneath the output and asks a different question: which institutions can interpret, challenge, and govern what the compute produces?

The examples that follow move across institutional levels, from agencies to hospitals, courts, and welfare systems. The level changes; the mechanism does not.

In a hospital, clinicians need to know when to trust an AI system, when to override it, when to escalate uncertainty, and how to document responsibility. A transcription assistant that saves time is different from a diagnostic recommender that changes care pathways. The model matters, but so does the workflow into which it is inserted.

A court can use an AI translation system legitimately only if affected people can challenge errors, officials can identify failure modes, and the court can decide when the translation is reliable enough to shape legal understanding. A mistranslation in a tourist app is an inconvenience. A mistranslation in a legal proceeding can alter someone’s rights.

In agriculture or welfare delivery, the model is only one part of the surrounding system. Data quality, field-level feedback, domain expertise, grievance channels, and operational discipline determine whether the tool can be used well. A crop advisory system that cannot hear back from farmers when advice fails is not a learning system. It is a broadcast system that reports confidence without being able to learn from failure.

India offers a parallel example of the same pattern. In Telangana, an algorithmic system called Samagra Vedika consolidated data from several government databases to create digital profiles of residents and help officials determine welfare eligibility. A 2024 investigation by Al Jazeera and the Pulitzer Center’s AI Accountability Network reported that the system wrongly denied access to subsidized food for many poor households. In one case, a widow was denied food security benefits after the system confused her late husband with a car owner. The administrative process accepted the database match as administrative truth.

The failure was a classic database problem of entity resolution: the technical challenge of determining when different records refer to the same real-world person or household. But once that match entered the administrative workflow, it was no longer experienced as a database error. It became an eligibility decision.

Telangana is not the same case as the Dutch childcare benefits scandal, but both reveal the same institutional pattern. Automated classification entered a public decision process. The system produced an eligibility signal. The public agency lacked adequate mechanisms for explanation, verification, correction, and redress. The affected citizen became responsible for disproving a machine-readable error.

The output may be cheaper to generate than to interpret, contest, escalate, and correct. The scarce capacity is the surrounding ability to do those things well.

Figure 1: AI becomes consequential through institutional workflow; procurement, documentation, grievance channels, and incident learning determine whether that workflow can be governed.

The Reallocation of Scarcity

For some tasks, AI lowers the time and marginal cost required to produce an output relative to existing workflows. Here, an output means the artifact or signal a system produces before an institution acts on it: a summary, translation, recommendation, classification, image, first draft, code suggestion, risk score, or customer response. Visible abundance expands. Scarcity does not disappear; it moves.

When prediction based on statistical likelihood becomes cheaper, more decisions become candidates for probabilistic inference. When generating text, code, images, classifications, and recommendations requires less time and labor, institutions can produce more of them. In both cases, human judgment does not disappear. It shifts toward defining objectives, interpreting outputs, managing exceptions, and deciding when an output is good enough to act on. AI can make inference cheaper; it does not make judgment automatic.

In practice, AI-generated outputs can accumulate faster than managers, officials, clinicians, teachers, or caseworkers can verify or integrate them. Without workflow redesign, faster prediction does not translate into faster realized output. It translates into a larger queue of outputs waiting for human interpretation. The bottleneck moves from producing outputs to absorbing them responsibly.

These pressures predate AI. AI can intensify them by increasing output volume, introducing systems whose performance varies across contexts and updates, and lengthening the chain between model provider and affected decision. More outputs compete for institutional attention; plausible outputs still require verification; and responsibility can diffuse across more actors.

The movement of scarcity explains why a hospital can buy a system before it can evaluate it, a court can pilot a translation tool before it knows how to handle contested outputs, or a welfare agency can automate classification before it has built a meaningful appeals process.

The stakes go beyond workflow efficiency. Institutions do not only process information. They produce legitimate judgment. A court does not merely translate speech; it certifies what can be heard in a legal process. A welfare agency does not merely classify eligibility; it determines whether a public entitlement is recognized. A hospital does not merely summarize symptoms; it authorizes clinical attention. When verification fails, the institution’s claim to judgment weakens.

The reallocation of scarcity is the operational version of the stack problem. In The Stack Beneath the Interface, the deeper scarcity sits in chips, compute, cloud infrastructure, and energy. Here, the scarcity sits inside organizations — in their capacity to evaluate, contest, and remain accountable for what AI produces.

For many institutions, the binding constraint will not be whether AI can generate an answer. It will be whether the institution knows what to do with the answer.

Procurement Is Not Capacity

Procurement, in this context, is not just buying software. It is the institutional process through which a hospital, court, school, welfare agency, firm, or ministry specifies a need, evaluates suppliers, receives assurances, negotiates obligations, and decides whether a system can be trusted inside its own workflow.

Institutional dependence becomes sharper when AI systems are bought from outside firms.

The firm that builds, trains, hosts, updates, or maintains the system usually understands the technical architecture better than the institution buying it. The institution understands the public setting better than the supplier, but often lacks meaningful access to the model’s design, data, limitations, or update behavior. Each side sees only part of the problem.

Information asymmetry makes procurement look more capable than it is. A hospital, court, school, or welfare agency can buy an AI system, receive documentation, negotiate a contract, and still lack the ability to evaluate how the system behaves in its own workflow. The technical system may be acquired before the institutional capacity to govern it exists.

AI procurement adds a different assurance burden to the problems agencies already face with complex software. Probabilistic models can be tested and audited, but their performance has to be measured empirically across relevant populations, workflows, and operating conditions. Benchmarks and pre-deployment tests remain useful; they do not eliminate the need to monitor changes in inputs, model updates, and real-world performance after deployment. The challenge is not that these systems are unauditable. It is that procurement often ends before institutions have built the contextual testing and ongoing monitoring needed to audit them well.

Government procurement already struggles with systems whose internal logic agencies cannot fully inspect. Suppliers provide assurances that may not cover real-world use. Agencies rely on performance claims, demos, certifications, or contractual language. When failure occurs, technical responsibility, administrative responsibility, and public accountability can pull apart. The user or citizen sits downstream of that fragmentation.

Procurement documents and compliance rules can name obligations without creating the capacity to meet them. Compliance asks whether a rule has been followed. Operational capacity asks whether an institution can understand, test, govern, and revise the system it is using.

A checklist can be completed while the failure mode remains poorly understood. Fairness, liability, and auditability matter only when institutions have representative test results, logs, incident records, technical documentation, access to system behavior, and the expertise to act on that evidence.

Human-in-the-loop oversight has the same problem. High-stakes AI systems should not be allowed to convert outputs into consequences without accountable human judgment. That does not mean a person must review every output. Institutions have to decide which outputs can be sampled or monitored automatically and which require human review because risk, uncertainty, exception, or appeal crosses a threshold. A reviewer without time, authority, domain knowledge, technical documentation, or a meaningful appeals process is not oversight in the substantive sense. Operational capacity is what makes human oversight more than a label.

Operational Capacity Is Capability

Operational capacity is the institutional ability to evaluate, govern, audit, procure, and operationalize AI systems in real settings. It is the practical capacity to make AI usable without surrendering judgment to the system being used.

The definition has to stay disciplined. Operational capacity is not regulation in general, technical expertise by itself, organizational readiness in the abstract, or institutional quality as a whole. It is the specific ability to evaluate, contest, revise, and remain accountable for AI-mediated decisions.

Operational capacity is therefore a form of technological capability. Technological capability is usually imagined as the ability to build models, design chips, own compute, or produce frontier research. As AI diffuses, a large share of the social value and social risk will appear in institutions that do not build frontier systems. Hospitals, courts, universities, firms, banks, and welfare agencies will be users, integrators, and accountable decision-makers.

Institutional capability will not be measured by whether these users can train a frontier model. It will be measured by whether they can ask the right questions before deployment, detect failure after deployment, and preserve accountability when outputs shape decisions.

For middle powers, operational capacity is a plausible layer of technological capability. Few countries can quickly build full-stack control over advanced semiconductors, frontier compute, model ecosystems, and hyperscale cloud infrastructure. But many can build stronger institutional capacity around evaluation, procurement, auditing, domain adaptation, and responsible deployment. This capacity is slower than announcement-led AI strategy and less visible than model launches. It accumulates through practice.

An agency that learns how to procure AI responsibly becomes harder to mislead. A hospital that builds evaluation routines becomes less dependent on supplier claims. A court that defines contestability standards before deploying translation or summarization tools protects institutional legitimacy. A public system that tracks failures and allows correction becomes more capable over time.

Operational capacity has a concrete institutional form. It looks like procurement rules that require model documentation before purchase, not after deployment. It includes audit logs that record when an AI output shaped a decision, appeal pathways that allow affected people to contest a machine-readable classification, incident registers that let institutions see whether errors are isolated or patterned, and public-sector technical teams that can test supplier claims instead of merely receiving them. In software-engineering language, this is the institutional version of observability, testing, incident learning, and service-level discipline.

AI systems update at software speed; procurement rules, appeals processes, audit cultures, legal review, staff training, and institutional memory move at procedural speed. Failure often accumulates in that gap: systems change faster than institutions learn how to govern them.

India’s AI policy language is beginning to move in this direction. The IndiaAI Mission includes a Safe and Trusted AI pillar, and official summit materials describe the need for safety testing, transparency, auditing tools, and interoperable assurance mechanisms. In 2024, the Mission selected eight Responsible AI projects under that pillar, including work on indigenous tools, frameworks, and guidelines for ethical, transparent, and trustworthy AI technologies. Direction becomes capacity only as it produces routines, evidence standards, and institutional memory. Testing and assurance have to become operating capacities, not slogans.

Operational capacity is not a substitute for physical infrastructure; it is the institutional infrastructure that makes judgment, verification, and contestability possible.

The strategic mistake is to treat AI adoption as evidence that an institution is becoming more capable. Adoption can happen quickly. Capability forms slowly. The difference between the two is often invisible until failure occurs.

The Dutch childcare benefits scandal and Telangana’s Samagra Vedika show the same warning in different settings. The danger is not only that automated systems make mistakes. The danger is that institutions adopt systems faster than they build the capacity to govern the mistakes.

The Responsibility Chain Gets Longer

AI can make many institutional outputs easier to produce. Reports, forms, translations, risk scores, eligibility recommendations, summaries, and service responses can require less time and labor. In resource-constrained settings, that may be very useful.

Institutions need to know who checks outputs, who contests them, who explains them, who records the decision, who notices clustered errors, and who can stop a system when it begins producing harm. The hardest question is who remains responsible when technical design, procurement, frontline use, and administrative authority each hold part of the causal chain.

In many AI deployments, responsibility is distributed across a longer chain: model provider, system integrator, software supplier, public agency, frontline official, and affected citizen. The more distributed the chain becomes, the more operational capacity has to be designed rather than assumed.

Figure 2: As more actors enter an AI deployment, information and authority fragment; accountability must be designed across the chain.

Responsibility is therefore a capability question. A country that cannot build every layer of the AI stack may still build serious capacity at the operational layer. It can develop procurement standards, institutional audit routines, public-sector technical teams, domain-specific evaluation benchmarks, grievance mechanisms, incident registers, and stress tests for high-stakes deployments. In some domains, one form of agency is retaining the ability to refuse systems it cannot evaluate.

AI capability is not defined only by who has the largest models or the most compute. It also depends on which institutions can use AI without losing the ability to judge, contest, and correct what AI produces.

That is why the next question is not only how AI spreads, but how institutions build the capacity to keep learning from it.

When generation becomes cheap, judgment becomes infrastructure.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Further Reading

• OECD on AI in tax administration and the Dutch Toeslagenaffaire

• Amnesty International on the Dutch childcare benefits scandal

• Al Jazeera / Pulitzer Center investigation on Telangana’s Samagra Vedika

• Amnesty International technical explainer on Samagra Vedika

• PIB on IndiaAI Responsible AI projects under the Safe and Trusted AI pillar

In 2026, India’s Ministry of Electronics and Information Technology reported that the IndiaAI Mission’s common compute facility had onboarded 38,231 graphics processing units through 14 empaneled service providers and data centers. The facility was meant to provide startups, researchers, academia, and other eligible users with access to AI compute through domestic cloud providers.

The pricing was striking. According to a parliamentary reply, the facility was offering compute at an average subsidized rate of Rs 65 per GPU hour, except for select high-end GPUs.

The objective was direct: give Indian startups, researchers, and public institutions cheaper access to the computational infrastructure required to build and deploy advanced AI systems. For a country pursuing AI capability at scale, the logic is clear. Compute has become foundational infrastructure, and infrastructure access changes who can build.

The compute subsidy widens access, but the GPUs it rents still sit inside supply chains India does not control. Advanced AI chips depend on design, fabrication, and semiconductor supply networks concentrated outside India, and they are often reached through cloud infrastructure shaped by firms outside India. The Indian state is widening domestic access to AI, but much of the technological frontier it is subsidizing remains controlled elsewhere.

The access-with-dependence pattern is not uniquely Indian. India makes it unusually visible. Many technology middle powers are entering AI through adoption, cloud access, model adaptation, and subsidized compute rather than full-stack control. They can accelerate domestic use while remaining dependent at depths that become more strategically important over time.

The compute subsidy can widen domestic access and deepen strategic dependence at once. Openness at the interface is not openness across the stack. A country can be strong at one level and dependent at another.

The question is what kind of capability India is building when the user-facing layer becomes cheaper while the underlying infrastructure remains concentrated elsewhere. The answer lies in the stack beneath the interface.

The Infrastructure Behind the Interface

Seen from the surface, the public experience of subsidized AI is deceptively lightweight. A user opens a browser, enters a prompt, and receives a fluent answer. A developer calls an API or adapts an open-weight model for a narrower task. A hospital integrates an AI assistant into transcription workflows. A state government tests translation tools across regional languages. From the user’s perspective, intelligence appears as software: accessible through interfaces, downloadable in compressed weights, deployable at low marginal cost.

That surface experience is real. The barriers to experimentation have fallen, sometimes dramatically. Open-source tools have proliferated. Smaller teams can now build products that would have required much larger research organizations a decade ago. For many practical tasks, no one needs to train a large model from scratch anymore.

This is where the stack matters.

Underneath the interface is a layered system. Applications depend on models and inference services; those depend on cloud capacity, chips, fabrication networks, data centers, energy, and the software systems that connect them. The layers connect, but they do not share the same economics or the same concentration of power.

Three broad postures sit beneath these layer-by-layer choices: build domestic capacity, adapt external systems while accumulating local know-how, or depend on external systems where domestic alternatives are not feasible. The finer terms in the figure below name the practical forms those positions take.

Figure 1: A country may build, adapt, rent, or depend at different layers of the same AI system.

AI’s cost structure runs on two questions: what becomes cheaper, and what remains scarce. Two operations often get collapsed. Training a frontier model means building or substantially improving the model itself; it requires large, capital-intensive compute clusters. Inference means running a model for users; each request is cheaper than training a frontier model, but repeated use still consumes compute, electricity, cooling, and network infrastructure. A country can become fluent at inference and model adaptation while remaining dependent for frontier training.

As one moves down the stack, capital intensity rises, and with it concentration, because fewer actors can fund the depth. Surface accessibility is consistent with stack-wide concentration.

Consider the chip layer. Modern AI depends on advanced GPUs and related accelerators whose design requires accumulated architectural expertise and mature software ecosystems. Fabricating these chips requires extreme-ultraviolet lithography systems, advanced manufacturing capacity concentrated in firms such as TSMC and Samsung, and supply chains spanning highly specialized chemical, equipment, and materials ecosystems. The chokepoints are not accidental. They are the outcome of decades of globalization optimizing for efficiency through specialization. AI changed the strategic significance of that specialization; it has not loosened the chokepoints.

The software layer compounds the concentration. NVIDIA’s position in AI does not come only from hardware performance. It comes from CUDA: a deeply embedded software ecosystem around which machine learning tools, optimization libraries, developer workflows, and research pipelines have accumulated for over a decade. These ecosystems also shape training pathways. When engineers learn frontier AI through the libraries, deployment practices, and optimization routines of a dominant toolchain, talent formation reinforces the technical stack it depends on. A rival chip is not just competing against silicon. It is competing against an ecosystem.

Data cuts across these layers rather than sitting neatly inside one of them. Frontier training depends on large corpora; model adaptation depends on domain and language data; deployment depends on operational records, feedback, and governance that make data usable. A country can have broad access to models while remaining dependent on external datasets, benchmarks, labeling systems, or feedback loops.

Cloud infrastructure adds another dependency. Most startups, universities, and public agencies do not build large AI compute clusters themselves; they rent capacity. This lowers upfront costs and speeds adoption, but it also means that domestic AI capability is mediated by a small number of global cloud providers whose pricing, availability, infrastructure footprints, and geopolitical exposure shape what downstream actors can build.

For most firms and public systems, the relevant cost is not frontier training they undertake themselves but recurring inference at scale. Individual requests may be relatively cheap; serving them reliably across many users creates a continuing claim on compute, electricity, cooling, and network infrastructure.

And then there is energy.

Data centers are physical systems. They require land, electricity, cooling, water, and the planning needed to bring them together. India’s Ministry of Electronics and Information Technology has said that the country’s data-center capacity increased from about 375 MW in 2020 to around 1,500 MW by 2025. Industry estimates suggest it could exceed 4,500 MW in the next five to six years.

At that scale, AI is not only a software system. It is a claim on electricity, land, cooling, transmission infrastructure, and administrative coordination. The bottleneck for data-center expansion is rarely the model itself. It is grid capacity, site selection, power procurement, water management, and the state’s ability to assemble large infrastructure projects in compressed timeframes.

Localization does not eliminate dependence. A startup hosting a model on domestic infrastructure may still depend on foreign chips, external model architectures, third-party benchmarks, and global developer ecosystems. A drone running on-board inference may still rely on a retraining pipeline whose updates it cannot survive without. The moves intended to reduce dependence can leave hidden dependencies intact.

The stack decomposes by layer and concentrates by depth. It also changes the kind of binding constraint at each layer. At the chip layer, the constraint is industrial and geopolitical: fabrication knowledge, equipment chokepoints, supplier ecosystems, and export exposure. At the data-center layer, the constraint shifts toward administration: land, grid capacity, power procurement, and the state’s ability to assemble infrastructure quickly. Strategy has to differ by layer because the constraint differs by layer.

Figure 2: Public rails can shape participation and adoption; concentrated infrastructure requires different capabilities and instruments.

India’s Layered Capability

India enters the AI transition with real strengths.

India has a large software-services ecosystem, a deep pool of engineering talent, a large domestic market, and experience deploying digital systems across scale and heterogeneity. These capabilities matter. They will matter increasingly as AI moves into adaptation, deployment, and operational integration rather than only frontier model development.

But India’s strengths sit unevenly across the stack.

At the application layer, India is relatively well-positioned. Startups, enterprises, and public institutions can build on top of existing models and create tools for sectors such as healthcare, education, financial services, and agriculture. The application layer is competitive globally, and India is in that competition.

At the middle of the stack, the picture is mixed. India has growing capability in model adaptation, AI operations, software integration, and workflow automation, among others. But many of these capabilities still depend on external model architectures, foreign cloud infrastructure, imported chips, and global developer ecosystems. India has real middle-layer capability, but not middle-layer independence.

Dependence runs deepest at the base of the stack. Frontier compute, advanced semiconductor fabrication, hyperscale cloud capacity, and the dominant AI software ecosystems remain concentrated outside India. The IndiaAI subsidy reaches into these depths as a buyer, not as a producer.

Layered structure matters because the Indian AI debate often treats capability as flat. The success of Aadhaar, UPI, and related digital public infrastructure created a powerful narrative: India could compensate for late industrialization in some domains through scale, public digital rails, and institutional coordination.

Digital public infrastructure is the name now commonly used for shared, secure, and interoperable digital systems that support access to public and private services. In India, Aadhaar, UPI, DigiLocker, and related data-exchange systems made that idea globally visible. They showed that state-backed digital coordination could change markets at population scale.

The DPI lesson was real. It can also encourage a mistaken analogy. India’s digital public infrastructure emerged in an environment where the deeper layers needed for large-scale digital coordination, including mobile networks, basic cloud services, and consumer devices, were relatively accessible and not yet the central terrain of great-power technological rivalry. India did not need sovereign semiconductor fabrication, frontier compute clusters, or globally dominant cloud infrastructure to scale digital payments. The core problem was coordination: identity, payments, interoperability, governance, and adoption.

AI changes the economics underneath that coordination layer. Its bottlenecks are not only coordinative; they are infrastructural. DPI worked by coordinating accessible digital rails; AI requires negotiating scarce industrial inputs. Inputs such as compute, chips, cloud capacity, and data centers are becoming more concentrated precisely as they become more economically important.

India’s earlier digital successes remain relevant, but they cannot be reused as a template without modification.

The instruments that worked when the binding constraint was coordination are not the same instruments that work when the binding constraint is concentrated infrastructure. The better way to understand India’s position is as a layered capability structure. India can be strong in deployment while dependent in compute. It can be capable in adaptation while reliant on external model architectures. It can shape how AI diffuses through a large society without controlling the deepest frontier systems beneath it.

The layers also move at different speeds. Application-layer adoption can advance in months, while fabs, power systems, cloud regions, and specialized supply chains take years or decades to build.

This is the middle-power condition in AI: real scientific, technical, administrative, and market capacity, but not full-stack control over the frontier systems that define the technology.

Selective Sovereignty by Layer

Policy made the layered structure visible at the India AI Impact Summit in New Delhi.

At the summit, the Prime Minister described India’s AI approach through the M.A.N.A.V. frame, five principles covering ethics, accountability, sovereignty, accessibility, and legitimacy in how AI is embedded in society. The emphasis was on legitimacy, access, accountability, and sovereignty over how AI enters public life.

At the same summit, India formally joined the Pax Silica coalition, as part of strategic technology and supply-chain cooperation with the United States. The initiative is framed around securing the “silicon stack”: critical minerals, semiconductor fabrication, advanced AI systems, and deployment infrastructure.

The same logic is visible in the May 2026 agreement between Tata Electronics and ASML to support India’s first front-end semiconductor fab in Dholera. The partnership can deepen domestic fabrication capability while leaving access to advanced lithography embedded in an international supplier relationship.

The policy contrast matters. In one domain, India is asserting national governance over legitimacy, inclusion, evaluation, and public purpose. In another, it is coordinating internationally on the semiconductor and hardware supply chains that make frontier AI possible. That is selective sovereignty by layer.

Selective sovereignty is not a clean synthesis. It is a managed compromise: autonomy in some domains, dependence in others, and constant judgment about which dependencies are tolerable. It becomes a strategy only if the country can say why a layer is being ceded, what leverage is being preserved in exchange, and what domestic capability is being built around the dependency. Without that test, the same posture can become passive integration with better language.

Layered analysis does not by itself settle whether India should train frontier models or rely on open-weight and external systems. It supplies the criterion for that choice: which option builds learning, leverage, and evaluative capacity at the layers India can realistically deepen.

The test has to be layer-specific because the constraint is layer-specific. A country may rationally depend on external fabrication capacity where the constraint is industrial and geopolitical, while still needing to build domestic capacity around data centers, energy, evaluation, and deployment. Some dependencies cannot be procured around; some capacities require public authority and coordination.

MANAV’s sovereignty language is therefore best read as sovereignty over data and over the governance, legitimacy, accessibility, and social embedding of AI, not sovereignty over every layer of the stack. A country may build domestic depth where durable capability can accumulate, shape outcomes where institutional leverage exists, and integrate strategically where full control is economically unrealistic.

Technological Power Decomposes by Layer

Technological power decomposes by layer. A firm can localize an application and still depend on a foreign retraining pipeline. A government can widen compute access while remaining a buyer at the deepest layers. Each layer has its own concentration logic, scale economics, and institutional requirements for entry.

The strategic question is where a country sits in the stack, on what terms, and with what leverage.

Layered analysis also clarifies what is continuous in AI’s emergence, and what is new.

AI did not invent layered technological structure. Electrification, the integrated circuit, telecom networks, and the internet all had deep infrastructures that were capital-intensive, geographically uneven, and shaped by strategic competition. What is new in AI is the compression of timescales, the depth of opacity in some parts of the system, and the speed at which strategically important bottlenecks have become concentrated.

Stack analysis is only the beginning. It shows where capability has to be built, and why adoption alone cannot answer that question. Once the stack is visible, the harder questions begin. Where in the stack does a country accumulate durable capability? Which institutional pathways allow shaping outcomes without controlling everything? What kind of operational capacity does responsible use require when access is cheap but evaluation is scarce?

The internet era treated software as detached from matter. AI returns matter to the picture.

Visual note: The diagrams in this essay are original Yukti visuals, designed from the author’s briefs and produced with AI-assisted code generation, then reviewed before publication.

Further Reading

• Ministry of Electronics and Information Technology parliamentary reply on IndiaAI compute facility

• PIB on IndiaAI compute capacity and data-center growth

• PMO on the India AI Impact Summit 2026 and M.A.N.A.V.

• PIB on India joining Pax Silica

• PIB on the Tata Electronics-ASML agreement for the Dholera semiconductor fab