The signal: India’s technology advantage is being tested by AI. For decades, a large part of the export model rested on delivering skilled work at a lower cost than in client markets. Generative AI and automation put pressure on labour-hour pricing—but they also create new opportunities in implementation, domain expertise, multilingual products and AI operations. The shift is not from labour to tokens overnight. It is from selling capacity alone to proving outcomes.
- The old arbitrage is under pressure—not extinct
- The new economics: models, tokens and completed work
- The Indic tokenisation opportunity is real—but specific
- Sovereign compute can lower barriers, not erase costs
- Where Indian technology companies can build a durable moat
- The CXO test: measure the whole system
- The FINA view: India must move up the value chain
The old arbitrage is under pressure—not extinct
India’s IT and business-services industry built a durable global position through engineering talent, process expertise, scale and cost competitiveness. AI changes the economics of some work: code generation, document processing, support triage and other repeatable tasks can require fewer manual hours. Buyers may therefore ask suppliers to deliver faster, automate more and price contracts around outcomes rather than staffing levels.
That does not mean India’s technology sector has become obsolete. Enterprises still need systems integration, security, data engineering, domain knowledge, change management and accountable operations. AI creates demand for these capabilities while putting pressure on suppliers that compete mainly by adding people to a project. The strategic challenge is to capture more value per engagement, not to assume every task can be automated or every client will immediately abandon time-and-materials contracts.
For context, NASSCOM’s industry reporting places India’s technology sector at hundreds of billions of dollars in annual revenue. Industry totals vary by definition and fiscal year, so they should not be confused with the revenue of IT services alone. The key signal is the scale of the installed base—and the importance of adapting that base to AI-enabled delivery.
See the NASSCOM research and industry publications for the underlying sector estimates.
The new economics: models, tokens and completed work
AI introduces new cost variables: model choice, input and output tokens, context length, latency, tool calls, retries, human review and the cost of running the workflow in production. Comparing an API price with an employee’s hourly rate is therefore misleading. A useful comparison asks how much it costs to complete a defined task at the required quality, speed, security and scale.
A lower token price does not automatically make a workflow cheaper. A model may need more prompts, longer context or human correction. A smaller model may be sufficient for a bounded task—or fail on the edge cases that matter. The right unit of analysis is cost per accepted outcome, including exceptions, monitoring and operational support.
This is where Indian technology firms can differentiate themselves. They can combine model selection with enterprise integration, evaluation, process redesign and ongoing service ownership. The opportunity is not merely to resell access to a model; it is to help a customer safely complete work that was previously slow, expensive or difficult to scale.
The Indic tokenisation opportunity is real—but specific
Multilingual AI creates a further opportunity. Tokenisers split text into units that models process. If a tokenizer represents one language less efficiently than another, equivalent content may consume more tokens, use more context and cost more to process. This can affect multilingual products, especially when they handle long documents or high volumes of user interactions.
Research published in 2026 has measured tokenisation inefficiencies for Indian languages and found that tokenizer design can materially affect token counts. A separate paper introducing MUTANT-Indic reports improved tokenizer fertility and inference throughput against the baselines it tested. These findings support investment in better multilingual tokenisation; they do not justify a universal claim that every Western model costs five times more for every Indic language or task.
Sources: The Tokenizer Tax: Quantifying and Explaining the Cross-Lingual Cost of Subword Tokenization for Indian Languages and the MUTANT multilingual tokenizer research.
For Indian companies, the business case should be tested with representative content: benchmark token counts, task accuracy, latency, model cost and human correction rates across languages and scripts. A native or multilingual model is valuable when it improves the full workflow—not merely when it produces fewer tokens.
Sovereign compute can lower barriers, not erase costs
India is also expanding access to shared AI compute through the IndiaAI Mission. Government information describes a subsidised compute-as-a-service model and a growing pool of accelerators available to eligible startups, researchers and other users. The official portal publishes GPU configurations and prices, which vary by hardware, instance size and reservation period.
Subsidised access can help teams experiment, fine-tune models and test workloads without building a data centre. But a headline GPU-hour rate is not the total cost of an AI product. Storage, networking, engineering, model serving, data preparation, reliability, security and utilisation all affect economics. Eligibility and actual pricing also matter; a public programme’s rate should not be presented as a guaranteed price for every user or workload.
Sources: the IndiaAI Compute Portal pricing calculator and the Government of India’s IndiaAI Mission overview.
Domestic infrastructure can also support data-residency, latency and procurement requirements in particular use cases. It does not automatically make a system sovereign or compliant. Buyers still need to examine who operates the service, where data travels, who can access it, which model licence applies and how the service can be audited or exited. For a related perspective, read our analysis of sovereign cloud and the difference between data residency and control.
Where Indian technology companies can build a durable moat
- Own the workflow: specialise in an industry process, integrate the systems involved and take responsibility for measurable results.
- Make multilingual AI useful: evaluate models and tokenisers on real Indian-language tasks, including accuracy, safety and user experience.
- Build repeatable delivery assets: reusable connectors, evaluations, policy controls and deployment patterns can improve margins across clients.
- Operate AI after launch: monitoring, model changes, incident response, security and quality assurance are continuing needs, not one-off implementation work.
- Price against value carefully: outcome-based contracts can reward productivity, but require agreed baselines, quality measures, exception handling and a clear allocation of risk.
- Use compute strategically: benchmark public, private and hyperscale options on total cost and operational fit before committing to a platform.
The CXO test: measure the whole system
For technology leaders and service providers, the transition should start with a small number of workflows whose baseline cost, quality and completion time are known. Compare a human-led process, an AI-assisted process and—where appropriate—an automated process. Include human review, failed attempts, exception handling and ongoing operations in the calculation.
Then test the result in production-like conditions. Track accepted outcomes, error rates, customer impact, cost per completed task, latency, security incidents and the amount of work that still needs expert intervention. Expand only when the economics and risk controls hold up beyond a demo.
Recent export data provides a useful baseline for this transition. The Reserve Bank of India’s September 2026 survey estimates that India’s software-services exports, excluding sales through overseas commercial presence, rose 8.2% in FY2025–26 to US$221.4 billion. Within that total, IT services accounted for US$147.0 billion, BPO services US$56.0 billion and engineering services US$12.0 billion. The figures show that the export engine remains substantial; they do not by themselves prove that AI has already displaced labour-hour contracts or lifted productivity. For executives, the next test is whether providers can translate AI adoption into higher-value engineering, implementation and managed outcomes while protecting service quality and margins. See the RBI’s 2025–26 software and ITES exports survey.
The FINA view: India must move up the value chain
India’s next technology advantage will not come from pretending that labour arbitrage has vanished or that subsidised GPUs automatically create an AI economy. It will come from combining talent with domain knowledge, reliable AI systems, multilingual capability and the ability to deliver measurable outcomes at scale.
The shift is from selling hours to owning more of the work—and from importing generic intelligence to building products and services that fit local languages, institutions and business realities. That transition will be uneven. Companies that can prove quality, economics and accountability will be better positioned than those relying on dramatic claims about token costs or the imminent collapse of the old model.
Bottom line: India’s technology arbitrage is being redefined, not simply reversed. The enduring advantage will be the capacity to turn compute and AI into dependable business outcomes.



