Get in touch

14 min readBengaluru

The Rope Sellers Buy a Rope Machine

Indian IT slept through the AI revolution, woke up to a crashing stock price, and is signing partnerships like gym memberships after a heart attack.

  • #indian-it
  • #agentic-ai
  • #anthropic
  • #tcs
  • #infosys
  • #outsourcing
  • #professional-services
  • #china-ai
Sepia Dalí-style drawing of galloping horses melting into liquid as frayed ropes snap, with tiny figures clinging on, a melting clock and bell, and a smoky factory in the distance
Sepia Dalí-style drawing of galloping horses melting into liquid as frayed ropes snap, with tiny figures clinging on, a melting clock and bell, and a smoky factory in the distance

A follow-up to “The Rope Sellers.” Quick recap of that one: professional services survive on accountability moats. Indian IT is the purest bodies-for-hours pyramid ever built, and it has no moat. This is the part where the pyramid meets the machine that automates pyramids.

Two quotes, eleven months apart

July 2025. TCS announces its largest layoff ever, about 12,000 people. CEO K. Krithivasan tells Moneycontrol:

This is not because of AI giving some 20% productivity gains. We are not doing that… It is not because that we need less people.

June 2026. Tata Sons chairman N. Chandrasekaran, at the TCS AGM:

If the company has half a million employees, the day is not far when the company will have half a million AI agents.

He adds that TCS is “unlikely to hire the same number of people” because portions of the work will go to the agents.

Same company. Same leadership. Eleven months.

So which is it? AI has nothing to do with headcount, or AI is about to become half your workforce? The answer, obviously, is that it depends who's asking. Tell the press AI isn't taking jobs. Tell shareholders AI is the entire future. The gap between those two statements is the gap between transformation and panic, and everything in this post lives inside that gap.

The market already voted, and it voted with a chainsaw

Let's start with the part the press releases don't mention: the stock prices, which are on fire, and not in the good way.

The Nifty IT index is down roughly 29% in 2026, making it the worst-performing sectoral index of the year, sitting nearly 30% below its December 2024 peak. Individually: TCS down about 33% for the year, trading at its lowest level since August 2020, with its market cap slipping below ₹10 lakh crore. Wipro down ~31%, flirting with ₹170. Infosys down ~27%, hitting a 52-week low of ₹1,030 against a ₹1,727 high it touched in February. HCL Tech down ~30%. LTIMindtree down ~34%. Brokerages have started capitulating too — Nirmal Bang downgraded TCS to an outright Sell with a target of ₹1,693, roughly half its old target.

Two triggers keep repeating in the selloff coverage, and both are hilarious in a grim way.

Trigger one: Accenture, the industry's bellwether, cut its FY26 revenue guidance and said the quiet part loud — client budgets are not expanding despite all the excitement about AI. Its stock dropped ~18% overnight and dragged the entire Indian IT complex down with it. Clients are interested in AI. They are just not interested in paying IT vendors more for it. They're interested in paying them less. Remember the “AI discount” from the last post? The market finally did the math.

Trigger two, and this is the one I'd frame and hang on a wall: in June 2026, Indian IT stocks fell for five straight sessions, with the Nifty IT index plunging over 5% in a single day, and one of the cited triggers was reporting that Anthropic's Claude Code can sharply reduce the cost and complexity of modernizing legacy software systems.

Read that again. Legacy modernization is one of Indian IT's bread-and-butter revenue lines. Anthropic is the company TCS, Infosys, and Cognizant all signed glossy “strategic partnerships” with. Their partner shipped a product update, and their stocks fell off a cliff, because the market understood instantly what the partnership announcements were designed to obscure: the thing they're reselling is the thing that eats them.

Which brings us to the actual thesis of this post.

The partnerships are share-price defense, not strategy

Look at the timing. Infosys hit its 52-week high on February 3, 2026. It announced its Anthropic partnership that same month, as the AI-disruption selloff was gathering. TCS announced its “Global Premier Partner” status in the Claude Partner Network in June 2026 — the exact month its stock was hitting multi-year lows. Cognizant upgraded its Anthropic partnership in July 2026. Accenture launched an entire “Accenture Anthropic Business Group” in December 2025, right before its guidance cut.

Every one of these announcements landed while the stock was bleeding or about to bleed. That's not a coincidence and it's not a strategy. It's a press-release IV drip. When your stock is down 30% on fears that AI destroys your business model, you announce an AI partnership. Not because it changes anything — because the announcement is the product. The audience is the shareholder, not the client.

And what are these partnerships, mechanically? TCS will “equip 50,000 associates with Claude.” Infosys will build Claude-based agents for regulated industries. Cognizant will deploy Claude to its developers. Strip the adjectives and every single one is the same arrangement: we will implement someone else's model for our clients, and pay the model company for the privilege.

This is a car dealership. Anthropic builds the engine. TCS sells it, installs it, services it, and takes a margin on the labor. There is nothing shameful about being a dealership — dealerships make money — but let's not stand in the showroom calling ourselves an automotive innovator. The dealership doesn't own the engine, doesn't set the engine's price, and gets exactly zero say when the engine company ships a self-installing engine. Which, per the June selloff, it just did.

Everest Group analysts are already warning these firms need “model portability, abstraction layers, fallback models” as standard practice. Translation: your strategic partnership is a dependency with a logo.

How we got here: twenty years of not even trying

None of this was inevitable. It's the bill for two decades of choosing comfortable margins over building anything.

Indian IT was born in panic — other people's panic. Y2K needed billions of lines of ancient COBOL fixed fast and cheap, India had the English-speaking engineers who still knew mainframe languages, and software exports went from $1 billion in 1997 to $6.2 billion by 2001. Y2K gave the industry its delivery model, its reputation, and its cash. Everything since has been the same trick at bigger scale: bill a body in dollars, pay it in rupees, keep the spread.

Every few years someone announced the model was dead, and every time they were wrong. SaaS was supposed to kill IT services — if Salesforce hosts your CRM, who needs the consultants? Instead, someone had to integrate Salesforce with SAP, sync the data warehouse, customize the workflows, and babysit the whole contraption. SaaS created a decade-long systems-integration boom. “The end of IT jobs” became an inside joke.

I'm deliberately not leaning on that pattern for comfort this time, and here's why in one sentence: SaaS complemented IT labor; agentic AI is built to substitute for it. The integration work, the maintenance, the migration, the ticket queues, the code itself — the new technology automates the exact complementary work that saved the industry last time. Maybe enterprise AI deployment turns out to be so messy that it becomes the new integration tax and the body shops get another twenty years. That's possible. But it's a bet, not a law of nature, and the people making the bet show no sign of understanding the technology well enough to price it.

How would they? Look at the R&D line. TCS, Infosys, and Wipro spend roughly 0.3–0.5% of revenue on R&D. Accenture spends about double that share and it's rising. The companies actually building AI — Google, Microsoft, Meta, Amazon — spend north of $20 billion a year each. TCS's R&D budget is a footnote you need a magnifying glass to find. For twenty years the margins were too good to bother. Why build IP when renting people pays 25%?

And when the industry did briefly produce a technical, innovation-minded CEO, it fired him.

Vishal Sikka, Infosys's first non-founder CEO, backed OpenAI in 2015 — a donation, he confirmed in February 2026, of “$3 million or something like that,” back when OpenAI was a pure nonprofit. Sikka pushed the Nia AI platform, design thinking, moving up the value chain. He was ground down by a public war with founder Narayana Murthy over strategy, governance, and pay, and resigned in August 2017 citing a “continuous drumbeat of distractions.” The stock fell 10% the day he left. He was replaced by Salil Parekh, who returned Infosys to services-first cost discipline and generous buybacks — exactly the finances-over-innovation trade the board wanted.

Nine years after donating to OpenAI, Infosys signed a partnership with OpenAI. To resell OpenAI's models. The company that could have owned a sliver of the AI revolution now runs one of its dealerships. If you wrote this as fiction, an editor would cut it for being too on the nose.

The present: record profits, mass layoffs, and a dividend firehose

Here's the current picture, and notice that no two pieces of it are compatible.

TCS closed FY26 with revenue of ₹2.67 lakh crore, net profit of ₹52,820 crore, and 25% operating margins — a four-year high. Not a distressed company. In the same year: net headcount down more than 23,000, largest layoff in company history, and ₹39,571 crore paid out in dividends under a policy of returning 80–100% of free cash flow to shareholders. Record margins, staff cuts, and essentially all the cash shipped out the door — while R&D stays at rounding-error levels. That is the balance sheet of a company harvesting a mature business, narrated in the vocabulary of a company transforming.

The AI theater is everywhere. TCS's “AI revenue” run-rate went $1.5B → $1.8B → $2.3B across three quarters, which sounds great until you notice it's ~7% of revenue and “AI revenue” is a category elastic enough to include any project with a model in the same building. Infosys announced it trained 250,000 employees in AI and GenAI — then cut 8,440 people in a single quarter. Train a quarter million people in the technology, quietly shed the people. Nobody squares those on the same earnings call, because they can't be squared.

Meanwhile the fresher package sits where it's sat since roughly the UPA government: ₹3.5 lakh, its real value halved by inflation, to the point where Indian outlets now run the “a metro plumber out-earns a TCS fresher” comparison unironically. And the trained mid-level talent is walking out the side door to Global Capability Centers — the in-house India tech centers of multinationals — which now number 2,117, employ 2.36 million people on $98.4 billion of revenue, hired over 5 lakh people in 2026 alone, and pay 12–20% more. The service firms built the talent pool. Their clients are now hiring from it directly and skipping the markup.

Meanwhile, China is playing an entirely different sport

Here's the comparison that should keep every Indian IT board awake, and doesn't.

While Indian IT signed dealership agreements, Chinese firms built the cars. Chinese open-weight models went from 1.2% of global LLM usage in late 2024 to roughly 30% by late 2025, per OpenRouter's study of 100 trillion tokens of real traffic. Alibaba's Qwen family passed one billion cumulative downloads on Hugging Face by March 2026 — the fastest any open-source model family has ever hit that mark — captured over half of all global open-source model downloads, and has 180,000+ derivative models, more than Google and Meta combined. In February 2026 alone, Qwen logged 153.6 million downloads: more than the next eight competitors combined. DeepSeek shipped V4 in July 2026. Moonshot's Kimi K3 is a serious agentic coding model. Stanford and Berkeley researchers train top-performing models on Qwen bases for $30–50.

And China turned this into foreign policy. At the 2026 World AI Conference, Xi Jinping launched a 29-nation AI alliance and called open-source AI a “rare, historic opportunity.” Singapore's OCBC bank runs DeepSeek and Qwen internally. Indonesia's Indosat builds on DeepSeek. Malaysia is building sovereign AI on Huawei silicon. The Global South's default AI stack is increasingly Chinese, because it's open, cheap, and good.

Remember when the stereotype was that China copies and India codes? China took US chip sanctions — an actual, deliberate attempt to kneecap its AI industry — and responded by building leaner, cheaper frontier-adjacent models and giving them away as geopolitical strategy. India's IT industry, facing no sanctions, sitting on $280+ billion of annual revenue and the world's largest engineering workforce, responded to the same decade by… increasing dividends.

The national picture is barely better. The IndiaAI Mission's five-year budget is about ₹10,371 crore (~$1.2 billion), of which roughly ₹400 crore had actually been released by early 2026. A single US hyperscaler spends more on R&D in two weeks. Sarvam AI is the honorable exception — genuine from-scratch foundational models built in India, a $1.5 billion valuation, $150 million from HCLTech (credit where due: the only major that bought equity in a model builder rather than a reseller badge). One Sarvam does not close a two-orders-of-magnitude gap.

DeepSeek reportedly trained a frontier-class reasoning model for single-digit millions. That figure is debated, but even the skeptical estimates land well inside what TCS pays out in dividends in a week. The capability was purchasable. The choice not to buy it was a choice.

The BPO wing is on a shorter fuse

The voice side of outsourcing doesn't get the luxury of a slow debate. Contact centers are ~95% labor cost, and voice agents now resolve routine calls in under 90 seconds, 24/7, with zero attrition. Gartner projects conversational AI cuts contact-center labor costs by $80 billion in 2026, and that by 2029 agentic AI resolves 80% of common customer-service issues autonomously.

The market has priced it: Teleperformance fell 19% in a single day back when Klarna bragged its OpenAI assistant did the work of 700 agents, and short interest has since climbed past 12% of float. Concentrix dropped 25% in one session after cutting guidance. Capgemini bought WNS for $3.3 billion explicitly to rebuild it around “agentic operations.” The Philippine BPO association — an industry of ~1.9 million jobs — revised its own 2028 employment forecast downward, from a hoped-for 2.5 million to roughly flat. When an industry lobby cuts its own headcount projection, believe it before you believe any CEO.

Firms like [24]7.ai, whose entire pitch is conversational automation, are in the awkward position of promising to bring the disruption themselves while owning none of the underlying models. That's not a moat, that's a franchise agreement with your replacement.

One honest caveat: Klarna itself over-rotated and re-hired humans for complex cases. Tier-1 call volume is going away fast; the buildings won't be empty by 2028. But the headcount line points down and it is not coming back.

The future: three sortings

The adapters

The path exists and a few firms are actually on it. It runs through owning things: vertical IP in regulated, liability-bearing domains where the accountability moat from the last post actually applies. Wipro raising its stake in insurtech Aggne to 80% and running FCA-regulated life-and-pensions administration is coherent. TCS's sovereign-AI and data-center play (100MW+ with OpenAI) is coherent — infrastructure is something you own. HCLTech buying into Sarvam is coherent. Notice the pattern: the coherent moves involve equity and assets, not partner badges.

The zombies

The default path. Keep reselling frontier models at thinning margins, keep announcing reskilling numbers while cutting the reskilled, keep the dividend firehose running, and slowly discover that outcome-based pricing — which clients are already demanding, and which LTIMindtree's CEO says clients are “excited” about (of course they are, it's their money) — priced your revenue down without you building anything a client couldn't get elsewhere. Zombies don't collapse. They just become smaller every year with excellent margins, like a beautifully managed melting ice cube.

The macro problem nobody owns

IT-BPM is ~7% of India's GDP, the largest chunk of its services exports, and for thirty years it was the escalator that turned engineering graduates into a dollar-earning middle class. Chandrasekaran has already said out loud that hiring won't track revenue anymore. If revenue survives on an AI-delivered, outcome-priced model, it survives with far fewer people — and GCCs, hiring 5 lakh a year at a premium, absorb only the top slice. The arithmetic for the other two million graduates a year does not close, and no earnings call is going to close it.

Conclusion

Put the whole thing in one paragraph. These firms spent twenty years choosing margins over R&D, fired the one CEO who bet on OpenAI before it was cool, and are now — with stocks down 30% and clients demanding AI discounts — announcing dealership agreements with the very companies whose products triggered the selloff, timed suspiciously to the worst weeks of the crash. The layoffs are “not about AI” in July and the company is “half a million AI agents” by June. A quarter million employees are trained in the technology that's thinning their ranks. The dividends flow, the R&D doesn't, and across the water China went from meme to shipping half the world's open-source AI in eighteen months.

Transformation and panic look identical from outside — both involve reorganizations, big announcements, and the word “AI” said many times. The difference is internal: transformation is building toward a position you understand; panic is performing understanding for shareholders while the position erodes. The stock market, to its credit, has stopped grading the performance and started grading the position. That's what down-33% means.

The rope sellers never learned how rope is made. Someone invented a machine that makes it free, so they signed a deal to distribute the machine, issued a press release about their rope heritage, and paid out the rope money as dividends. The market read the press release, looked at the machine, and sold.

References

Stock market

TCS statements & financials

  • Moneycontrol — K. Krithivasan on July 2025 layoffs (“not because of AI”)
  • TCS AGM coverage, June 2026 — N. Chandrasekaran “half a million AI agents” remarks (also referenced in the Reuters / MarketScreener partnership note)
  • TCS Q4 / FY26 results — revenue ₹2,67,021 crore, net profit ₹52,820 crore, 25% operating margin; FY26 dividend ₹39,571 crore — TCS Investor Relations

Partnerships

History & R&D

  • Vishal Sikka, India Today AI Summit (February 2026) — on the 2015 OpenAI donation — India Today
  • Infosys 2017 resignation coverage; Sikka resignation letter (“drumbeat of distractions”) — Infosys press release
  • Company annual reports — R&D as % of revenue for TCS/Infosys/Wipro vs Accenture; hyperscaler R&D budgets from public filings (SEC EDGAR)
  • Deccan Chronicle — India's IT rise and the Y2K origin story

GCCs & workforce

  • NASSCOM–Zinnov “GCC Value Orbit” report (July 2026) — 2,117 GCCs, 2.36M professionals, $98.4B revenue — NASSCOM
  • Business Today — GCC hiring crossing 5.1 lakh in 2026
  • Storyboard18 — Infosys Q4 FY26 headcount cut of 8,440; attrition data

BPO / voice AI

China

  • OpenRouter / a16z token-usage study — State of AI (December 2025); Chinese open-source models ~30% of global usage (also covered via South China Morning Post)
  • Forbes — China's DeepSeek V4 and Qwen Reshape the Open-Source AI Race (April 2026): Qwen 1B downloads, 180K derivatives, February 2026 download share
  • Startup Fortune / WAIC 2026 coverage — Xi Jinping's 29-nation AI alliance, Global South adoption
  • Digital in Asia / AI in Asia on China's AI stack; USCC “Two Loops” staff paper (March 2026)

India AI

  • IndiaAI Mission budget (₹10,371 crore) and disbursement reporting (~₹400 crore by early 2026)
  • Economic Times — Sarvam AI $1.5B valuation, HCLTech $150M investment (June 2026)

On this site