The Last
Arbitrage

The labour arbitrage that underwrites Indian IT is elegant in its simplicity. A senior software engineer in the United States commands a fully-loaded cost — salary, benefits, real estate, management overhead — of $140,000 to $180,000 per annum. An equivalent engineer in Bengaluru, Hyderabad, or Pune costs $25,000 to $38,000. This is not a niche differential; it is a 4.5-to-1 ratio applied across 5.4 million workers and compounded across three decades of compound annual growth.

The durability of this arbitrage was not a given. It required sustained investment in engineering education, English-language proficiency at scale, telecommunications infrastructure, and — critically — the trust infrastructure of delivery: the CMM certifications, the ISO audits, the CSAT frameworks, the account management layer that convinced Fortune 500 chief information officers to route mission-critical workloads to the other side of the planet.

These were not cheap investments, and they are not trivially replicable. Vietnam, Poland, and the Philippines have each absorbed portions of the arbitrage at the margin. But none has replicated the ecosystem. India's IT sector is not merely a collection of skilled engineers; it is a confidence machine — and confidence, in enterprise technology procurement, is the highest-value asset.

The question is what maintains that confidence when the underlying price signal — the 4.5-to-1 cost ratio — begins to erode. If an AI-assisted engineer in Toronto can perform the same function as a team of junior engineers in Chennai, at one-tenth the cost and with zero latency in communication, the confidence premium that India has built does not disappear. But its ability to command a price premium — on top of an already diminished arbitrage — does narrow considerably.

This is the first, most important point: the threat is not replacement. The threat is compression. And compression, in an industry whose profitability is predicated on volume and margin management at scale, is an existential financial event even at levels that would be invisible in any other sector.

Consider the arithmetic. If AI tools reduce the junior headcount required for a given engagement by 30 percent — a conservative estimate, based on published productivity data from GitHub Copilot and Cursor — and junior headcount represents 60 percent of billable hours in a typical IT services delivery model, the revenue impact on a renewals-based engagement is a 18 percent decline in billing at constant scope. This is before any renegotiation by the client. It is the arithmetic of productivity, flowing directly into the top line.

In the annual reports of India's ten largest IT exporters, there is a word that appears with remarkable frequency: opportunity. AI is an opportunity. Generative AI is an opportunity. The global digital transformation wave is an opportunity. Automation, cloud, and platform economics are — invariably — opportunities.

What is notably absent is a quantification of risk. Specifically: what percentage of current billable hours are performing tasks that AI can now replicate, or will be able to replicate within the tenure of a three-year client contract?

Our analysis, drawing on published capability benchmarks from OpenAI, Google DeepMind, and Anthropic, cross-referenced with task taxonomy data from McKinsey Global Institute and NASSCOM, suggests the answer is between 28 and 45 percent of current IT services revenue. Not immediately. Not next quarter. But within the contract cycle — within the horizon that determines whether a client renews at the same staffing level, or whether the conversation shifts to outcomes-based pricing. Anthropic's Q1 2026 announcement of Claude's commercial-grade legacy modernisation capability — a category previously excluded from at-risk analysis — places that estimate under active upward revision, with our working range now trending towards 38–55 percent.

The phrase "outcomes-based pricing" deserves careful unpacking. In practice, it is a client's polite way of saying: we want the same output with fewer of your people. The commercial implication — a reduction in billing at constant scope — is a revenue contraction event. The firms that understand this call it a "value-based model transition." The firms that don't will experience it as unexpected churn.

This is the number nobody mentions. It does not appear in risk disclosures. It does not feature in analyst day presentations. And it is, we argue, the most important single number in Indian corporate finance for the next 48 months.

The popular discourse on AI coding capability has oscillated between two equally unhelpful extremes: that these models will replace all programmers, and that they are merely autocomplete toys. Both miss what is actually happening in enterprise software procurement.

What AI is actually good at is the reduction of the discovery-to-delivery cycle for well-specified problems. A requirement that would previously take a junior engineer four days to implement can now be scaffolded by an AI tool in four hours, reviewed by a senior engineer in one, and delivered with documentation attached. The total billable hours collapse, but the client value does not.

This is the structural threat. Not that AI replaces the senior engineer — it does not, and will not for some time. But that it hollows out the pyramid. Indian IT's delivery model depends on a ratio: one senior architect supervising five mid-level engineers supervising fifteen juniors. If AI absorbs the function of the fifteen juniors, the pyramid inverts, and the cost model that made India competitive — which was always a volume model, not a premium model — loses its mathematical basis.

The firms that understand this are beginning to reframe their value proposition around what AI cannot do: managing client relationships under uncertainty, navigating regulatory environments, building institutional trust, and — most critically — understanding what the client needs before the client has articulated it.

These capabilities are real, and they are genuinely scarce. The question is whether they can be priced at a level that sustains a $224 billion export industry.

We believe they cannot — at least not without significant structural adjustment. The premium for human judgment in enterprise software is real, but it is a premium that the market has historically been reluctant to pay explicitly. Indian IT's model has been to include human judgment in the price of headcount, not to price it separately. As headcount declines, the judgment premium must be extracted differently — through advisory relationships, retainers, governance roles, and outcome-based arrangements that require a fundamentally different sales motion.

To be fair to India's IT majors, the competitive response is not absent. It is accelerating. TCS, Infosys, Wipro, HCL, and Tech Mahindra have all announced significant AI investment programmes in the past eighteen months. They are training hundreds of thousands of employees. They are building proprietary AI platforms. They are rewriting their service portfolios.

What they have not done — and this is the observation that the institutional community has been slow to absorb — is make a single, clear disclosure about what percentage of their current revenue base is at structural risk of AI displacement within a five-year horizon.

This is not an oversight. It is a choice. The choice is rational: any such disclosure would immediately invite client renegotiation conversations. But it leaves the market — and, more importantly, the ecosystem of talent, policy, and education that India's IT sector depends upon — operating without the information it needs to adapt.

This is the practice of designing the workflows, interfaces, and governance structures by which human professionals collaborate with AI systems. It is the discipline that will determine whether enterprises extract value from their AI investments or merely produce AI-generated content that no one trusts. It is the highest-margin professional service of the next decade.

And it does not yet exist — not formally, not as a P&L line — in any Indian IT firm. There are teams. There are practices. There are centres of excellence with aspirational names. But there is no client-facing offer, priced at a premium, that positions any Indian firm as the definitive authority on how enterprises should govern their relationship with artificial intelligence.

This absence is significant not because it reflects incompetence. India's IT talent pool contains precisely the combination of technical depth and systems-thinking that this service requires. The absence reflects a deeper institutional hesitation: to name the threat explicitly enough to build the defence around it.

The firms that will survive the arbitrage compression are not the ones that can automate most aggressively. They are the ones that can define and own the layer above automation: the layer where human judgment, client relationship, and institutional knowledge determine the difference between an AI system that works and one that is abandoned after six months.

India's IT sector has the relationships, the delivery infrastructure, and the client trust to own that layer. What it lacks — so far — is the intellectual honesty to acknowledge that the layer below it is closing.

Consider a senior engineer at a large Indian IT services firm. She joined in 2019, at ₹22 lakh per annum. By 2023, through consistent performance appraisals, her salary has reached ₹48 lakh. Her ESOP grant — 3,000 Restricted Stock Units, vesting over four years — completed its final tranche in December 2023. At the prevailing market price, the fully-vested position represents approximately ₹28 lakh.

That ₹28 lakh was not abstract. It was the down payment on a flat in Whitefield. It was her daughter's undergraduate education abroad, partially funded. It was the early retirement buffer that allowed her to take the risk of joining a startup in 2025. None of these plans required the stock to go up. They required it not to go down — and to remain exercisable in a market where her skills still commanded a premium.

Both conditions are now uncertain. The AI-driven compression of demand for IT services is not yet a stock-price event — earnings have not yet declined materially at most majors. But the forward P/E compression that typically precedes an earnings revision is already visible in the spread between Indian IT sector valuations and the broader Nifty 50. A 20% de-rating from peak multiples translates, for our hypothetical engineer, to a ₹5.6 lakh reduction in ESOP value — not catastrophic, but enough to foreclose the Whitefield flat at current prices.

The ESOP trap is compounded by a structural asymmetry: most RSU schemes have exercise windows tied to quarterly earnings blackout periods. An employee who anticipates a de-rating and wishes to exit before it materialises is often legally constrained from doing so. The insider trading framework — designed to protect market integrity — inadvertently locks employees into equity positions they have no mechanism to hedge. Senior engineers at the top-5 IT firms collectively hold an estimated ₹45,000–60,000 crore in unvested or recently-vested RSUs. This is not a rounding error in the context of Indian household wealth.

The fear of job loss is not the same as job loss. It is, in many ways, more economically damaging — because it operates through every household simultaneously, before the first pink slip is issued. When an engineer in Bengaluru hears that his firm's headcount planning for FY2026 involves a 15% reduction in junior roles, he does not wait for a letter. He defers the car upgrade. He pulls the children from the international school. He lets the home-loan pre-payment lapse. He cancels the Maldives holiday. None of these decisions register in any employment statistic. All of them register in GDP.

India's IT sector is uniquely concentrated in five urban centres — Bengaluru, Hyderabad, Pune, Chennai, and the NCR corridor — where it has become the dominant driver of middle-class consumption for roughly two decades. The real estate markets of Whitefield, Gachibowli, Hinjawadi, and Old Mahabalipuram Road were built, literally, on the assumption of continuous IT employment growth. The loan books that financed them were underwritten on the same assumption.

The leverage dimension is equally acute. A dual-income IT household in Bengaluru that took a ₹1.2 crore home loan in 2021 — at 7.2% interest, on a combined income of ₹32 lakh — carries an EMI-to-income ratio of approximately 43%. This is within bank lending guidelines, but it leaves virtually no buffer for income disruption. A single job loss in the household — not retrenchment, simply non-renewal of a project contract — transforms a performing loan into a stressed account within two quarters.

RBI data for FY2025 shows housing loans outstanding in Bengaluru, Hyderabad, and Pune collectively exceed ₹7.8 lakh crore. ADWIZR estimates that approximately 31–36% of this credit book is serviced by IT-sector households. A sustained period of IT employment uncertainty — even without material job losses — is sufficient to generate a detectable increase in early-stage delinquencies in these geographies. The credit institution most exposed, by branch network and loan-book concentration, is not a public-sector lender. It is the private banking sector, whose Bengaluru and Hyderabad branches account for a disproportionate share of premium housing loan portfolios.

India's Grade A commercial office stock exceeds 700 million square feet, of which IT and technology-aligned occupiers represent approximately 65% of annual absorption across the five principal IT hub cities. In Bengaluru alone — where total Grade A stock stands at approximately 225 million square feet — IT sector tenants account for an estimated 68% of annual take-up, a proportion that has held broadly stable since 2016. This concentration is not incidental. The SEZ-based development model, the software parks framework, and two decades of campus-scale lease structuring have built an office market that is, at the asset level, a leveraged exposure to IT employment headcount.

The contractual structure of IT office leases has obscured this concentration risk from visible market data. Enterprise IT campus leases carry 5–9 year tenures, triple-net structures, and headcount-based expansion clauses that created the appearance of durable, predictable cash flows. That durability was contingent on headcount assumptions that are now under revision.

Listed real estate investment trusts have amplified this concentration risk into the public equity market. Embassy Office Parks REIT carries approximately 87% IT-sector tenant exposure by leasable area. Mindspace Business Parks REIT stands at approximately 74%. Brookfield India REIT at approximately 68%. The weighted average across the three listed vehicles is 77% IT-sector exposure — a figure that appears in no stress scenario in any of their published investor presentations.

The residential micro-markets of Whitefield, Sarjapur Road, Gachibowli, Hinjawadi, and Old Mahabalipuram Road are not merely influenced by IT employment growth — their price discovery mechanisms are structurally dependent on it. Average residential prices in these corridors grew at 8–12% compound annual rates between 2012 and 2024, underwritten by rising IT salaries and the implicit assumption that IT employment was a permanent economic condition rather than a contingent one. The typical buyer in these markets is a mid-to-senior IT professional with a 20-year housing loan, an EMI-to-income ratio of 35–45%, and no material income diversification. The asset and the income that services it are exposed to the same underlying risk factor.

Pre-sales — the primary financing mechanism for tier-1 residential developers in these corridors — represent 60–75% of project-level funding commitments. Developer feasibility analyses underwrite buyer profiles that are, overwhelmingly, IT-sector employees. A sustained period of IT employment uncertainty does not produce immediate price corrections; it produces booking cancellations, site visit declines, and a progressive deterioration in new project viability that manifests 18–24 months before any observable price movement. Real estate markets lag. The demand-side signal is visible now. The price signal is not — yet.

Residential inquiry data for the 12-month period ending March 2025 shows a statistically significant decline in site visit volumes across all primary IT-corridor micro-markets, in the absence of any material employment event. Whitefield recorded a 22% decline in qualified residential inquiries. Gachibowli fell 21%. Sarjapur Road fell 19%. Hinjawadi registered an 18% decline. These movements preceded by months the AI capability announcements that would have provided an explicit narrative trigger. A fair reading must acknowledge confounding variables: sustained high interest rates through FY24–25, post-pandemic oversupply in select corridors, pricing fatigue after the 2021–2023 residential bull run, and — in Bengaluru specifically — the well-publicised water infrastructure crisis. Each of these contributes to the softening. But the pattern of decline is geographically concentrated in IT-dependent micro-markets, not broadly distributed across metro residential markets. Non-IT-corridor segments in the same cities have not recorded equivalent declines. The most parsimonious explanation is anticipatory demand contraction — potential buyers in the IT sector, aware of uncertainty in their employment environment, deferring the single largest financial commitment of their careers. This is the most important leading indicator in the data set.

Bernstein’s Open Letter to the Prime Minister

On 23 April 2026 — less than two months after the original publication of this analysis — Bernstein’s Venugopal Garre and Nikhil Arela published an open letter to Prime Minister Narendra Modi that reads, in its structural diagnosis, as an institutional echo of the thesis presented in these ten parts. It is the first such letter from Bernstein since 2019.

The letter warns that India risks “under-delivering on its potential” without urgent structural reforms, and names, explicitly, the AI displacement threat to IT services and BPO — sectors employing an estimated 10–15 million people that form the backbone of India’s aspirational middle class. Bernstein flags that generative AI is “increasingly capable of handling coding, customer support, and back-office functions” — precisely the task categories mapped in Part III of this report.

Most critically, Bernstein raises the spectre of India becoming a “permanent consumer” in the AI economy — using AI tools built elsewhere, training global models on Indian data, without building domestic frontier capability. This maps directly to the intelligence arbitrage thesis in Part X: the last arbitrage of labour is ending; the first arbitrage of intelligence is available — but only if India builds rather than merely consumes.

Bernstein’s letter also contained a structural fiscal warning that bears directly on the “permanent consumer” thesis. State-level cash transfer programmes — approximately &rupee;1.7–2.5 lakh crore annually, consuming 2–3% of GSDP in several states — are crowding out capital expenditure on precisely the infrastructure that an AI transition requires: compute capacity, data governance frameworks, research institutions, and innovation ecosystems. Capital locked into broad transfer schemes is capital not being spent on the building blocks of domestic AI capability. The reform window is not merely narrowing because AI is advancing; it is narrowing because the fiscal space to respond is being consumed by other commitments.