The Scaling Law

Satya Nadella calls it the successor to Moore's Law. The curve is the easy part. Knowing where it crosses a cliff, and that the cliff is already on the calendar, is the rest.

One thing first, or a researcher will stop you. The scaling law is Kaplan's, then Chinchilla's. Test loss falls as a power of compute, parameters, and data. That is the real result, and it is narrow. Use it loosely and you will be corrected.

Nadella took the term and bent it. In his version the Scaling Law is the next Moore's Law, capability doubling on a clock. His sharpest line is that the work is bending the curve and then putting more points up the curve. He is careful to call it empirical, not physical. Not gravity. Something that has held, the way Moore's held, until it doesn't. His version is the one you can plan against.

Almost everyone plans against it wrong.

They read it as a forecast. Feed in compute, read out capability. True, and useless, because it stops at the curve.

The curve is continuous. Performance climbs a few percent at a time, slow enough to ignore. Value does not climb with it. Value sits behind a wall, and the wall does not move when the curve climbs five percent. Then the curve climbs five percent one more time and the wall is gone.

Eighty to eighty-five buys nothing. Eighty-five to ninety opens a use case. Ninety to ninety-five hands you an advantage. Ninety-five to ninety-nine moves an industry. Same slope the whole way. Four different worlds.

The labs publish the curve. Knowing which wall sits on your slope, and at what point the curve walks through it, is the part nobody publishes.

The Ladder Nobody Watches

People read this as a metaphor. It is the shape of how value actually arrives.

Most people judge AI on where it is today. They run the model, the output is not good enough, they move on. The people who win read the same slope and see the wall coming, and they start building while it is still in the way.

Reading The Threshold

It starts with the wall. The exact number where the economics flip for your one use case. Fraud review automates when false positives drop under the level a human will sign off on. Code generation replaces a contractor when it gets right often enough that you stop checking the work. Name the number. If you cannot name it, you do not have a threshold.

Then you put the wall on the curve. Where the published slope sits today, against the number you named. The gap between them is already drawn. You are not forecasting it. You are reading it off.

The move almost nobody makes comes next. You build before the crossing. The pipeline, the data, the workflow, while the model still fails the test you are building for, because it feels insane to build for a customer of zero. It is insane right up until the model crosses. Then the infrastructure is standing and everyone else opens a planning doc.

Then you move on the crossing. By the time a capability is being discussed, the people who read the threshold shipped a quarter ago. The gap between research prototype and business necessity used to be years. Months now. Whoever waits for the announcement is reading the result of a race that already ran.

The History Is Already Written

Enterprise software sold one thing for twenty years. Implementation complexity. The nine-month integration was the moat. The toll booth. You did not buy the software, you bought the year of consultants who made it work.

The wall here was simple. Could a model write a complete, production-grade internal system, end to end, correctly enough that a company would trust it. For a long time the answer was no, and the no was the entire business model of the services industry.

Then the curve crossed the wall. A system quoted at twelve crore and nine months got built by a small team with Claude Code in two months for one crore. A tenth of the cost. A third of the time.

A tenth of the cost.

The companies that read this saw it before the case study existed. They were not waiting for proof that the wall was gone. They had already moved their people off implementation and onto the thing implementation used to gate. The ones still selling the nine-month timeline found out the way the turkey finds out.

Where It Fails

The curve is smooth until it is not. Empirical, not physical, was the warning, and the day it stops holding it owes you nothing. Some capabilities arrive early, sideways, before the slope says they should, and a reading that waited for a clean crossing misses them. Reasoning showed up as a dimension nobody had on the ladder. Architectures reset the trajectory and the old curve stops describing anything.

The exponent lies by task. The slope that holds for one capability is the wrong slope for the one beside it. Borrow a multiplier from someone else's domain and you build for a crossing that lands a year late, or never.

The smooth part is most of the curve, and that is the part you get to read. The wall you can see coming is the one worth building for. The one that drops out of the sky was never going to be timed by anyone, including the lab that shipped it.

Define the wall too loosely and you build for a crossing that never sharpens. Place it on the wrong curve and you are early by a year you cannot afford. Mostly the future is not what gets you. What gets you is reading the slope of the thing beside the thing you meant.

The labs draw the curve in public. The wall is on your slope, and the crossing is already on the calendar. Most people are still reading the model.