The Great AI Unwinding: Why IT Is Quietly Collapsing

Two days ago, the largest consulting company on Earth had the worst single day in its entire stock market history. It lost nearly a fifth of its value in hours. And here's what makes no sense. It didn't lose money. It made more than last year. I've spent 15 years building the systems these companies run on. And this is the second time this year a single AI headline has wiped out Indian IT. So, when a company beats its numbers and still gets punished like this, I don't see a stock story. I see a warning for every one of us in tech. In the next few minutes, how a number in New York wiped out lakhs of crores in Mumbai overnight, why 95% of company AI projects are quietly failing, why Microsoft just canceled the AI tool its own engineers loved, and the one skill that turns this whole disaster into the biggest opportunity of your career. Stay till the end. The scary story flips. When Accenture spoke, India bled. In one session, over 1 lakh 35,000 crore rupees gone from Indian IT. The Nifty IT index hit a 3-year low. Infosys down almost 8%. TCS, Tech Mahindra, HCL, all red. Why does an American forecast crash Bengaluru? Because Accenture is the bellwether. And the number that scared everyone was outsourcing bookings down 15%. Outsourcing is the exact business Indian IT lives on. But the deeper tell was one line from management. Clients aren't adding money for AI. They're reallocating existing budgets. AI isn't new growth. It's the same rupee moved. Remember, Accenture's results were good. The market isn't pricing what they earned. It finds pricing what they're afraid of, that AI replaces this entire industry. So, here's the question this video answers. Is that fear actually true? The firing started long before this crash. Across TCS, Infosys, Wipro, and HCL, more than 42,000 jobs gone since 2023. Infosys first annual drop since 2001. TCS first in 19 years. This wasn't a bad quarter, it was the floor cracking. And the bosses said it out loud. The CEO of HCL, "The time is out for that model, double the revenue with half the headcount." Accenture CEO, "Reskilling is not a viable path for the skills we need, half the workforce." Reskilling won't work from the top. But not everyone agrees. Infosys' CEO is hiring 20,000 freshers and says AI expands work. Investor Vinod Khosla says the opposite. Infosys and TCS get replaced by AI agents. So, who's right? The data picks a side and it's not the one you'd expect. A few months ago, an OpenAI founder, Andrej Karpathy, spent 2 hours one Saturday letting an AI score every American job by how exposed it is to AI. Then, after it scared millions, he deleted it. For decades, we believed automation hits factory workers first. His map showed the opposite. The more digital your work, the more exposed. Software, accounting, analysis, support, top of the danger list. Electricians, nurses, plumbers, safest. The safest job in the AI era needs hands, not a laptop. Anthropic's own labor study found the same week. The most exposed workers are older, educated, and well-paid. Map that onto India. The World Bank says for the first time AI threatens white-collar service jobs, not factory jobs. And they live here, Bengaluru, Hyderabad, Pune, Chennai. NITI Aayog warns worst case, our tech services workforce shrinks from 8 million to 6 million by 2031. If that's your job, I know exactly what you're feeling. So, let me tell you the part nobody's telling you. Here's what the whole world is doing. Every company is racing to put AI into everything. Microsoft shipped an entire agent framework. Adobe put agents inside Word and PowerPoint. The belief is total. AI does the work cheaper, faster, forever. Markets believe it. They're pricing Accenture for death. CEOs believe it. They're firing. Companies believe it. They're spending billions. And when everyone believes the same thing that confidently, that's exactly when you check if it's true. So, I did. I built these systems. And what happens when the belief meets reality is not what they sold you. One disclaimer first. AI is not weak. It's breathtaking. The newest model reworked a 50 million-line codebase in a day. Work that takes a team 2 months. That's real. So, if AI is this powerful, why are 95% of company AI projects failing? You need to know what AI actually is. At the core sits a large language model. Strip the magic and it does one thing. Predict the next word over and over. It learned the patterns of language from a huge slice of the internet. It doesn't read like us. It breaks everything into chunks called tokens. And tokens are the meter. Every word in, every word out is tokens. And tokens are what you pay for. Teaching that base model your specific job is called fine-tuning. Here's what the panic misses. A raw model is not a product. It's an engine, not a car. To make it work, you wrap it in a harness. Tools, your private data through retrieval, guardrails, security, and a human in the loop. Good companies are building all of it. The model was never the threat. The question is whether anyone engineered the system around it. Most didn't. Watch what happens when they don't. Air Canada's bot invented a refund policy, no grounding. A Chevrolet dealership's bot was talked into selling an SUV for $1. No guardrails. A delivery company's bot swore at customers and wrote poems mocking itself. No testing. Each time the model did exactly what a predict the next word machine does. The failure wasn't the AI. It was the missing harness. Point a raw model at your business and you don't get intelligence. You get confident fabrication. Scale it up. MIT, 95% of enterprise AI projects fail to deliver. A thousand company survey, 42% abandoned most of their AI up from 17. Gartner, 40% of AI agent projects canceled by 2027. The thing the market prices as inevitable is failing quietly everywhere. And the fear that AI replaces all of us tomorrow, physics says slow down. Researchers estimate a single AI request can burn on the order of a thousand times the power of a web search. One new AI data center eats a nuclear plant's worth of electricity. Microsoft is sitting on tens of billions in orders it can't fulfill. No power. The $500 target project just canceled a Texas site. This arrives slowly and every step needs a human to aim it. D. Remember I promised Microsoft canceled the AI its own engineers loved? Here it is. Microsoft gave its engineers Claude code, then killed it. Not because it was bad, but because the token bills blew through the division's entire annual budget. And Microsoft isn't even the worst case. Uber put Claude code on 5,000 engineers. Within months, 84% were hooked. Power users burning up to $2,000 a month. And it torched its whole year's AI budget in 4 months. And the kicker, token prices actually fell 60 to 80% that year. The bills exploded anyway because the better AI gets, the more you use and nobody was managing it. If giants like Microsoft and Uber can't manage their tokens, what chance does anyone else have? So, put it together, and the story flips. The market thinks AI replaces you. The data shows AI fails on and bankrupts the budget when it works. Both true at once because AI is powerful and unmanaged. The tool doesn't run itself. That gap is the biggest opportunity of the decade. Making a raw model trustworthy in the real world has a name, AI reliability engineering. It's the harness, the testing, the guardrails, the human in the loop, the monitoring that catches the model when it's wrong. Almost nobody does it well. That's why 95% fail, not because AI is weak, but because no one engineered it to be trustworthy. Its twin is token management. If tokens are what you pay for every word AI reads and writes, managing tokens is managing your AI budget, and Microsoft just proved almost nobody does it. The survivors will treat tokens the way a finance team treats money, measured, budgeted, optimized. That's a job. That's a skill. And right now, it barely exists. Here's the practice nobody talks about. Don't point AI at a goal and hit run. Do what a good doctor does. Diagnose first, write the plan, the spec. Then test that plan cheaply and let it fail 10, 50, 100 times on small token budgets until it's solid. Then spend the big tokens on real execution. Most companies do the opposite, execute blind and burn a fortune failing in production. Plan with tokens, test with tokens, then execute. And if you think this is just my opinion, last week Microsoft CEO Satya Nadella posted it to 28 million people. Every company, he said, must now build two kinds of capital, human capital and token capital, the model, data, and skills you actually own instead of rent. His words, "Without human direction, you have compute running in circles." He told enterprises to decouple their hard-won expertise from whatever model they rent and build private evaluations on their own data. That is AI reliability engineering described by the most powerful man in enterprise software. That is the entire game. So, here's the flip I promised. Yes, the old roles are shrinking. Python hiring down a third, Oracle down 70%. But pay for those skills rose 20 to 35% and demand for 16-plus years experience is up 28%. 50,000 open AI jobs in India, only 15,000 qualified. The engineers everyone wrote off, you were never the cost, you were always the cure. But only if you upskill, you stay the one who directs it. The human in the loop is not going away. So, what do you do? Today, a stack you can start this week. AI reliability engineering, the harness guardrails, human in the loop, token management, spec first testing. That's how you make today's AI work. And tomorrow, a deeper fix is being built. World models, AI that simulates reality before it acts instead of just predicting words. Two labs just raised over a billion dollars each for it. It might be what finally makes AI reliable by design. That's the next video. If this gave you the real picture, not the hype, not the panic, subscribe. One rule on this channel, don't trust, verify. The unwinding is real, too. If you build the skills nobody else has, I will see you in the next one. Wait, wait, wait. Quick disclaimer before you go. Yeah, AI helped me make this video. But every fact, every number, every bit of analysis, that's all me. I I the research and verification and then directed it. AI was the tool. Have you spotted the watermark? You can exactly know how it's made. Don't trust, verify. Even me.

Transcript source: youtube-owner-caption. Use the film as the primary record.