India’s rise as a technology economy was built on a remarkably powerful economic arbitrage: the world needed software, and India could supply the people required to build and maintain it at a fraction of the cost.
For three decades, this worked extraordinarily well.
The economic significance of Indian IT was never simply about producing brilliant computer scientists. A large part of the industry was built around something more mundane and, economically, perhaps more important: India became exceptionally good at supplying technical labour at scale.
In this sense, much of software engineering occupied a position surprisingly similar to copywriting, accounting or business-process outsourcing. The valuable skill was not necessarily reinventing the underlying system. It was understanding somebody else’s requirements and changing the system accordingly: implementing a workflow, integrating an API, configuring an ERP, migrating a database, testing another release, maintaining legacy software, or making thousands of small modifications for thousands of different customers.
The improvements were often incremental. But India became very good at making them.
That capability created one of the country’s most important export industries. Software and business services generated foreign exchange without requiring India to first become a manufacturing superpower.1 They created a large urban professional class and connected Indian labour directly to global demand.
Generative AI now threatens to alter the economics underneath that model.
When Software Becomes Abundant
For most of the software era, code was expensive because producing code required skilled human time. This constraint created an enormous global division of labour. A company in New York or London could design a system and employ thousands of engineers elsewhere to build, test, integrate and maintain it. Differences in wages mattered enormously because labour represented such a large proportion of the cost of producing software.
AI coding systems begin to weaken that relationship.
If one engineer can eventually accomplish what previously required five, ten or twenty engineers, the economic question is not whether programmers disappear. It is what happens to the value of labour arbitrage when the quantity of labour required to produce a unit of software collapses.
Consider the logic of outsourcing:
- If a project requires 1,000 engineers, reducing the cost of each engineer by 50 per cent produces an enormous saving.
- If the same project eventually requires 50 people working alongside powerful AI systems, wage differences matter considerably less.
The comparative advantage begins migrating elsewhere. Speed matters. Proprietary data matters. Compute matters. Distribution matters. Domain knowledge matters. And, increasingly, access to intelligence itself matters.
This is potentially a profound macroeconomic transition for India.
From Labour Arbitrage to Intelligence Arbitrage
Every technological era has a scarce input around which economic power accumulates. Industrialisation rewarded access to machinery, energy and capital. The internet rewarded distribution, software and networks. India’s services boom exploited the global difference in the price of skilled human labour.
The emerging AI economy may instead revolve around the price of machine intelligence. Today we experience this rather literally through tokens. A company can purchase reasoning from increasingly capable models almost as it purchases cloud computing. Give the model more context, allow it to reason for longer, invoke more agents, run more evaluations and process more information, and generally the bill rises.
That creates a peculiar economic situation.
AI can democratise capability because a small company suddenly has access to programming, analysis, translation, design and reasoning capabilities that previously required substantial teams.
But if intelligence remains predominantly metered, the democratisation is incomplete.
A well-capitalised American company might be able to spend millions of dollars allowing AI systems to continuously examine its codebase, customers, operations and scientific problems. A small Indian company may technically have access to exactly the same model while being economically unable to use it with the same intensity.
The model is democratised. The quantity of intelligence that can be consumed is not.
A company capable of spending one hundred times more on inference can explore more possibilities, run more simulations, generate more software, analyse more documents and retry failed approaches more frequently.
The risk is that the old capital divide simply reappears as a compute divide.
The Price of Intelligence
This makes the declining cost of inference one of the most consequential economic variables of the next decade. We normally discuss AI models in terms of capability: which model scores higher, reasons better or writes better code. For developing economies, price may ultimately matter just as much.
Imagine intelligence becoming 100 times cheaper:
- Procurement: Every small manufacturer can continuously optimise procurement.
- Legal: Every lawyer can interrogate millions of pages of precedent.
- Governance: Every municipality can analyse its infrastructure.
- Education: Every school can provide individualised instruction.
- Enterprise software: Every small business can build software specifically around its workflow rather than purchasing generic software and adapting itself to it.
This is where the analogy with electricity becomes useful. Economic transformation did not occur merely because electricity was invented. It occurred when electricity became cheap, reliable and sufficiently ubiquitous that firms could reorganise themselves around its availability.
AI may follow a similar trajectory. The most important question may therefore cease to be who has access to AI? Almost everyone eventually will. The more important question becomes: how much intelligence can an organisation afford to consume?
Why On-Premise AI Matters Economically
Local and on-premise deployment changes the marginal economics of intelligence. Privacy and sovereignty matter, particularly for governments, banks, healthcare systems and defence. But ownership converts an operational token expense into installed productive capital.
Cloud computing initially transformed technology because companies no longer needed to own infrastructure. AI may create situations where the economics partially reverse.
Once sufficiently capable models can be deployed on hardware controlled by an organisation, intelligence begins to resemble installed productive capital. There is an upfront investment in hardware, models, integration and energy. But after that investment, the organisation can potentially use that intelligence repeatedly.
- A government department does not have to wonder whether another million documents are worth another million API calls.
- A factory can continuously run models against operational data.
- A university can provide inference to thousands of students.
- A legal system can process enormous archives without treating every additional query as a separate purchase of foreign computational services.
The analogy moves from hiring labour toward owning machinery. And this distinction is particularly important for countries such as India.
The Capital Deepening of Knowledge Work
Economists describe capital deepening as increasing the amount of capital available to each worker. Industrialisation did this physically.2 A farmer with a tractor became vastly more productive than one relying entirely on human labour. A factory worker operating sophisticated machinery could produce dramatically more output per hour.
AI could produce a similar phenomenon in cognitive work:
- A software engineer equipped with persistent machine intelligence becomes a different economic unit from a software engineer without it.
- So does a doctor.
- So does a civil servant.
- So does a lawyer.
The question for India is therefore not simply whether it can train more AI engineers. It is whether it can dramatically increase the amount of machine intelligence available per Indian worker.
That is a capital-formation problem. And it leads to a different interpretation of AI infrastructure: GPUs, inference hardware, models, data centres and local AI systems are not merely technology expenditure. Increasingly, they constitute the productive machinery of a knowledge economy.
India’s Services Problem Could Become Its Advantage
The very characteristics sometimes criticised in India’s technology industry may become useful in the transition. Indian engineers have spent decades dealing with heterogeneous systems, poorly documented processes, legacy databases, idiosyncratic customers and endless exceptions. That is precisely the environment in which autonomous software systems eventually need to operate.
The great asset accumulated by the Indian IT industry may therefore not be code. It may be encoded knowledge about how organisations actually work.
For decades, much of this knowledge remained tribal. An engineer knew how a particular banking implementation behaved. A consulting team understood how a government department actually processed a file rather than how its process manual claimed it did. Someone else knew which integration would fail, which exception mattered and which requirement the customer had forgotten to specify.
Historically, that knowledge travelled with people. AI creates the possibility of converting it into systems.
The next generation of Indian technology companies could therefore sell something fundamentally different from the traditional outsourcing proposition:
- The old proposition: Here are 500 engineers who can operate your process.
- The new proposition: Here is a system that has learned how those 500 engineers operate your process.
That is a radically different economic product. One scales with headcount; the other scales with compute.
The Productivity Question
India’s long-term economic challenge is ultimately productivity. The country cannot become rich simply by moving progressively larger numbers of people into moderately productive service jobs. At some point, output per worker has to rise substantially.
AI offers an unusual route toward doing this because it can increase productivity not only in factories but throughout the service economy:
- A small architecture firm could possess capabilities previously available to a multinational consultancy.
- A district administration could analyse information at a scale that previously required dozens of analysts.
- A ten-person software company could maintain a product surface that once required a hundred engineers.
This is potentially enormous. But there is also an uncomfortable implication. If AI raises the productivity of India’s existing skilled workforce without creating corresponding new demand, the country may need fewer workers to produce the same quantity of exported software services.
The transition could therefore be simultaneously productivity-enhancing and employment-displacing. That is precisely why protecting the existing outsourcing model would be the wrong strategic objective.
The goal should be to move upward in the economic stack before the old advantage erodes.
From Selling Hours to Owning Systems
India’s technology industry has historically monetised people exceptionally well. The next challenge is monetising intellectual property, models, infrastructure and autonomous systems. That means moving from selling hours to selling outcomes; from implementation to products; from manpower augmentation to machine augmentation; and eventually from providing technical labour to providing technical capacity.
The distinction matters for national income:
- A services business grows largely by hiring more people.
- A software platform can grow revenue much faster than employment.
- An autonomous system can potentially go further still: continuously performing work that previously required recurring human expenditure.
For an economy with India’s population, this creates legitimate questions about employment. But refusing productivity improvements is not a sustainable answer. Countries become wealthy by producing more value with fewer inputs, then allowing labour and capital to migrate toward new forms of demand.
The strategic question is whether India owns enough of the new productive assets to capture that additional value.
A New Form of Digital Public Infrastructure
India’s digital public infrastructure experience can extend from transactions to cognition. India’s DPI strategy has largely focused on reducing the marginal cost of transactions: identity, payments, document exchange and other common rails.
Imagine shared or locally deployable intelligence infrastructure that allows universities, municipalities, courts, hospitals and small businesses to run models cheaply against their own information.
The public good would no longer merely be an API that verifies identity or moves money. It could be a baseline quantity of machine intelligence available throughout the economy.
This need not mean one giant government model. It could mean standards, open models, shared compute, interoperable data layers, evaluation infrastructure and deployable systems that prevent every institution from having to recreate the entire AI stack independently.
India would then be attempting something much more ambitious than building domestic alternatives to foreign AI companies. It would be trying to reduce the domestic price of intelligence.
That could have enormous spillover effects.
The Next Indian Arbitrage
India’s first technology arbitrage relied on cheap human intelligence; the next relies on cheap machine intelligence. The first was straightforward: human intelligence was expensive in rich countries and cheaper in India. The next one could be almost the inverse: make machine intelligence so inexpensive that Indian workers and institutions can consume vastly more of it.
That requires thinking beyond models. It requires cheap inference, energy, chips and data centres; deployable smaller models; better foundational datasets; systems capable of operating against messy Indian information; evaluation infrastructure; and an ecosystem capable of turning general intelligence into domain-specific productive capacity.
Most importantly, it requires abandoning the assumption that India’s objective should be to preserve the economics of the services economy.
The services revolution succeeded because India exploited the dominant technological architecture of its time better than almost anyone else. That architecture is changing. Software is becoming cheaper to produce. Intelligence is becoming purchasable. Human execution is increasingly being converted into computational execution.
India can treat this as a threat to an industry that employs millions of people. Or it can recognise what may be the larger opportunity.
The country spent thirty years making human technical intelligence inexpensive enough for the world to consume at scale. Its next economic project may be to make machine intelligence inexpensive enough for India itself to consume at scale.
Footnotes
-
Software exports, back-office services, finance and, more recently, global capability centres for multinational firms have carried a large share of India's growth story, while the industrial base that historically absorbed millions of semi-skilled workers has grown only modestly. Tatvita Analysts, The Services Leapfrog and Missing Manufacturing Multiplier: India's Structural Transformation, 1951–2025 (2026-08-17). This datum shows that software and service exports drove India's economic growth and export earnings while its industrial manufacturing base remained modest, supporting the claim that India generated foreign revenue without first establishing manufacturing dominance. ↩
-
Capital deepening refers to an increase in the proportion of the capital stock to the number of labor hours worked. Federal Reserve Bank of St. Louis, How Capital Deepening Affects Labor Productivity (2018-04-19). This definition establishes that capital deepening is a technical measure of capital relative to labor, supplying the factual premise for the author's assertion that industrialisation physically increased the capital available to each worker. ↩