Timesheets to tokens

I have spent three years on this blog following one cost curve, from the week ChatGPT shipped to the summer the productivity gauge broke. The argument has been the same the whole way. Generation got cheap. Verification became the work. This year the curve runs into an entire industry, the one I grew up in, and the collision is worth naming plainly.

The Indian IT services model is an operating system, and the operating system is billable hours. You win a contract, you staff it with people, and you bill for the time those people spend producing software. The unit you sell is a human hour spent on generation. Stack enough of those hours under a fat pyramid of fresh graduates and you get a machine that turns headcount into revenue. It ran for two decades. It built cities.

Generation is the exact part of that machine that just fell toward zero.

You do not have to take my read for it. Vinod Khosla, on the record at a Bay Area forum last September, said all of business process outsourcing will be replaced, and all IT services will be replaced, within the next five years. People heard a provocation. After three years of watching the cost curve, I heard an accurate read of where the cost went. When the unit you bill for is the unit that just got commoditised, the business built on selling it does not get a discount. It gets repriced.

The repricing is already showing up, and here is the asymmetry that is the whole story. The deflation is a realized number in the financials. The rescue is an announced number in the forecasts. The pyramid base has thinned hard, with fresher hiring down sharply from its peak. Revenue growth has slowed to a crawl and the bellwether firms have started saying de-growth and deflation out loud on earnings calls, which is not language you use unless the polite denials have run out. That is the booked, cleared, realized side of the ledger. On the other side sits the offset everyone points to, the idea that cheaper software creates more demand and the firms pivot up the value chain. That side is real too. It is also still a forecast, framed as emerging, not yet cleared through a single annual report. A strategist does not ask whether AI hits Indian IT. He asks which side of that ledger has actually settled, and right now only one of them has.

There is a second blow landing in the same city at the same time, and it is the one that gets missed. The clients and the labs are now hiring Bengaluru directly. Global capability centers, the captive engineering arms that multinationals run in India, have grown into thousands of centers employing millions, doing the work in-house that used to be a vendor engagement. The frontier labs are opening offices in the same talent pool. The same engineer who was a billable line item in a services contract is now hired straight, at a multiple of the services rate, to do the work the contract used to wrap. Bengaluru is not being outsourced anymore. It is being hired. The middleman is being removed from both ends at once.

So what survives. This is where three years of the same argument pays off, because the answer is not new, it is just arriving at scale. The work that survives is the work above generation. Deciding what correct means and proving the machine met it. The spec, the test suite, the gate. Generation is the commoditised input. Verification is the defensible output, and it is the thing an entire services industry has under-sold for its whole existence because, in a world of cheap human generation, nobody had to charge for it separately. In a world of free machine generation, it is the only thing left worth charging for.

The honest counter, and I hold it firmly, because the doom version of this is lazy. Jevons cuts both ways, including for jobs. Cheaper software means more software gets built, and more software in the world needs more people around it, not fewer. The diffusion is slower than the headlines suggest. Adoption is near universal and real capture is rare, which means most firms have bought the tools and not yet rebuilt the work, which is exactly the gap a smart services firm can sell into. The 1.7 million people in Indian IT are not the liability. The arbitrage model is the liability. The people are the asset, if they cross from selling generation to selling verification before the budget runs out.

I have watched this exact repricing happen inside one of my own products. ReelMaya makes product videos and ad scenes for e-commerce, the kind of shoot that used to be quoted in studio days, a model, a photographer, a crew, all of it billed as human hours. It now comes out of a few image-model calls. The invoice did not get smaller so much as it changed units. What used to be a timesheet for a shoot is now a token bill for a render. And the work that holds its value is no longer the shoot. It is the eye that decides which of the generated frames is actually good enough to ship. Same lesson, one rung down. The meter moved from hours to tokens, and what it still pays a premium for is the judgment sitting on top of the generation.

That is the move, and it is the same move I have been describing for an individual engineer, a team, and now a country. Stop selling the hour spent generating code. Start selling the proof that the code is correct. The title of this is the whole thesis. Timesheets to tokens. The meter changed, and the meter does not care how many people are at the keyboard. What it pays for now is whether you can prove the output is right.