The Demand Paradox

The crash that defines June 2026 began on a Friday. On June 19, Accenture released its Q3 FY2026 earnings after the US market opened — and the number that landed was not the revenue miss. It was the guidance. Accenture did not merely guide flat. It guided a reduction in future revenue growth: full-year growth revised down to 3–4%, below the prior 4–7% range, with management commentary signalling that enterprise clients were explicitly deferring discretionary spending as they assessed what AI meant for their IT headcount. By the close, Accenture had fallen 18% in a single session — its stock touching $128, a nine-year low.

The Indian IT segment did not wait long to draw its conclusions. When markets in Mumbai opened the following Monday, the sector was already re-pricing. HCL Technologies fell. Infosys fell. LTIMindtree and Wipro fell. This was not sympathy selling. It was recognition: if Accenture — the sector's most sophisticated operator, with 750,000 employees and $70B in revenue — was guiding its own revenues down because of AI disruption, the model itself was under structural pressure.

HCL Technologies — the one major IT services firm growing faster than its peers at 7.5% — still saw its stock fall 39% over the trailing twelve months. Infosys, which has been the loudest proponent of its "AI-First" strategy, trades at nearly half its 2025 high. The market is not picking winners and losers within the sector. It is discounting the model entirely.

A sector under repricing

The numbers make the pattern clear: revenue is still growing (for four of five firms), but the market has stopped caring. The divergence between revenue performance and stock performance is the starting point of this analysis.

The market is not extrapolating from current revenue. It is pricing in a future it believes is structurally different — one where the relationship between software demand and IT services revenue has been broken. This article examines whether that belief is correct, and — critically — whether it is permanent.

The GitHub counter

While the sector repricing was playing out, GitHub published another quarter of its Innovation Graph data. India's developer base stands at 24.6 million — up 5.9× from 4.2 million in 2020. Pushes — actual code commits, not registrations — grew 7.2×. Each Indian developer is 23% more active on the platform than five years ago. Indonesia: 5.6×. Brazil: 5.5×.

We analysed this data in detail in The Open Source World Is Being Rebuilt From the Global South. The short version: the developer surge in the exact geographies where IT services firms source most of their talent is not slowing. It is accelerating.

The surface reading is contradiction. Talent is booming. The companies that deploy that talent are being repriced. The deeper reading is that these two data series are measuring different things — and the difference is the key to understanding what is actually happening.

What IT services actually sells

The IT services business model is a spread trade. You identify the gap between what a Western client will pay ($150–200 per hour) and what a developer in Bengaluru will accept ($25–35 per hour), hire at scale, and pocket the margin. Revenue scales with headcount times utilisation times billing rate. The model has been extraordinarily successful: TCS grew from near-zero to one of India's five most valuable companies on this thesis. Accenture went from $44 billion in 2019 to $72 billion in 2025.

The model has a single structural vulnerability: it requires the labour-arbitrage gap to persist. AI is compressing the gap — not from the supply side, but from the demand side. A product team in Seattle, equipped with Claude Code or Cursor, can now produce output that previously required a ten-person offshore engagement. The client's need for 10,000 person-hours falls to 2,000. The billing rate is unchanged. The hours are not.

The upskilling counter-narrative

Every major IT services firm has responded with an AI capability programme. Accenture committed $3 billion in AI investment in 2023, mandating training for all 750,000 employees through its LearnVantage platform. Infosys launched Topaz — its AI enterprise platform — under an explicit "AI-First" mandate from CEO Salil Parekh. LTIMindtree has Canvas.AI. TCS partnered with Anthropic, deploying Claude across 50,000 employees. Wipro has its own AI platform and retrained 225,000 employees on generative AI tools in FY2025.

The market has discounted all of them. Not because the programmes are not real — they are — but because upskilling solves a capability problem when the underlying issue is a business model problem.

If you train 750,000 employees to use AI copilots and each becomes twice as productive, you now need 375,000 people to deliver the same work. The training investment becomes self-defeating unless new demand materialises fast enough to absorb the freed capacity. Which brings us to the data point almost no one is discussing in this context.

The Intercom proof

Intercom, the customer communications platform, published its internal AI deployment data in June 2026. It is the most rigorous primary-source evidence available for what AI actually does to software teams at scale.

The number that matters for this analysis is buried in the detail: the product defect backlog shrank 54%. All critical and high-severity defects — the P1s and P2s that had accumulated for years — were cleared.

This is not a productivity story. It is a demand story.

The ATL problem: work that never gets done

Every engineering organisation runs a hidden dual budget. The visible backlog contains what is formally prioritised, estimated, and sprint-ready. Sitting above it is the ATL pile: tech debt deferred sprint after sprint, P2 and P3 tickets that never reach triage, "above the line" feature requests from business units that are acknowledged and perpetually deprioritised. This work is not unimportant. It is uneconomical at the prevailing cost per unit of output.

When the effective cost per unit of output falls — from $150–200 per hour to $30–50 per hour with AI augmentation — the economic threshold drops with it. Work that was previously unaffordable becomes viable. The ATL pile does not shrink because the business stopped caring about it. It shrinks because the return finally clears the cost. This is what happened at Intercom: the defect backlog cleared not because defects became less important but because resolving them stopped being more expensive than it was worth.

The Jevons mechanism in software

Economists call this the Jevons paradox: as the efficiency of using a resource improves, total consumption of that resource tends to increase rather than decrease. More efficient steam engines in 19th-century Britain did not reduce coal consumption. They made coal-powered industry viable at scales that had not been conceivable before, and total coal use rose sharply.

The same mechanism is operating in software. As the cost per unit of output falls, the threshold for what is worth building falls with it. Every company carries an informal backlog of things it wanted to build but could not justify at $200 per hour. At $30 per hour — the effective rate when AI augments a developer — those projects get funded. The Intercom 54% defect clearance is the Jevons mechanism made visible.

This is only the first-order effect. Intercom also notes that 600+ non-R&D employees — support, operations, product managers — are now using Claude Code for their own work. As the tools become accessible to people who are not engineers, the definition of who can initiate software work expands. The total addressable market for software development is not "engineers." It is everyone who has a problem that software could solve.

The open-source evidence

If the ATL thesis is correct — that AI is making previously-uneconomical work viable — we should see it in large-scale, observable data. We looked at the microsoft/vscode repository on GitHub: one of the most intensively tracked open-source projects in the world, with five continuous years of issue history and hundreds of thousands of filed bugs.

We measured two things: how many new bugs are filed each year, and how many are closed. Both are direct signals of software demand — filing rate shows how intensively the codebase is being used and tested; closing rate shows how much maintenance work is actually getting done.

From 2021 through 2024, both metrics were essentially flat — filing hovering around 3,900 per year, closing around 3,400. The backlog accumulated slowly as teams could not quite keep pace with incoming work.

In 2025, both lines broke upward simultaneously. New bug filings jumped 63% to 6,176 — software is being used more intensively, tested more thoroughly, and held to a higher standard. Bug closures jumped 39% to 4,205 — teams are clearing far more work than before. The Jevons mechanism made visible: lower cost per unit of output drives more total output in both directions.

The disintermediation trap

If demand is expanding, why are IT services revenues not showing it?

Because the expanded demand is being captured in-house. Intercom's 153 internal contributors built 267 Claude Code plugins. The 600+ non-R&D users were not served by a Wipro engagement. The defect backlog clearance, the doubled product output, the deferred ATL features — all of it happened inside the company, not through an external service relationship.

The pattern is consistent across every company doing this seriously. Firms capturing the demand expansion are building the capability directly. The external services layer is being bypassed — not because the work disappeared, but because the technology made it economical to do yourself.

This is the genuine risk the market is pricing in. And it is real. The question is whether it is permanent.

The J-curve

Every efficiency technology follows the same path. Initial deployment reduces the cost per unit of output — and in the short term, that appears to reduce demand for the old delivery model. The market prices in the left side of the curve. It is not pricing in the right.

Below is the condensed thesis of this analysis. The left arm is where Accenture's stock price landed on June 19. The right arm is where the Jevons mechanism points.

Who survives — and on what terms

Not all IT services models face equal exposure. The differentiation is not by company size or geography. It is by delivery model.

LTIMindtree's strongest moat is in manufacturing and BFSI — sectors where domain knowledge (ERP configurations, regulatory workflow, lending models) compounds in ways that commodity delivery hours do not. An AI copilot does not know the idiosyncrasies of a particular client's SAP landscape. A consulting team with fifteen years of implementation history does.

Wipro, declining at minus 1.6%, has the weakest positioning: revenue concentrated in time-and-materials and staff augmentation, with the least differentiated domain portfolio. The path to recovery requires a model transition the company has not yet credibly executed.

The transition, not the destination

Here is where the market may be making a category error.

What the stock prices of Accenture, Infosys, HCL, and LTIMindtree are pricing in is permanent disintermediation of the IT services model. That is probably too pessimistic.

The demand expansion is real and accelerating. The question is the time constant — how long before IT services firms figure out how to capture a share of it, rather than watching it go in-house? Companies are beginning to do internally what they previously outsourced. But most enterprises are early in that journey. The firms that outsourced application development to TCS or Infosys in 2015 are not going to restructure those relationships overnight.

Many — particularly mid-market enterprises without the engineering culture of an Intercom — will continue to rely on service partners who can bring AI-augmented delivery capability to them. They do not want to manage 600 employees using Claude Code. They want a vendor that does.

The IT services firms that navigate the transition period will emerge into a genuinely larger total market: more software work being done, at lower cost per unit, across a broader base of clients who can now afford things they previously could not. Accenture's $3 billion AI investment, if it lands correctly, positions them as the AI-augmented delivery partner for Fortune 500 clients who cannot or do not want to build the capability themselves.

The question is not whether the demand expansion eventually reaches IT services revenues. It will. Demand for software work is larger now than it was two years ago, and it is growing faster. The Jevons mechanism does not reverse. The question is which firms are still standing — and which have pivoted their model early enough — when the distribution of that demand shifts from in-house capture to serviceable contracts.

The market is right about the pain. It may be wrong about the permanence.