Something has gone wrong with how we're using AI.

The bills come in. The headlines tell the story. Microsoft's internal numbers say running AI now costs more than the engineers it was built to serve. Uber spent its entire 2026 AI budget by April. GitHub is walking away from flat-rate plans. Three Mile Island is being brought back online to power a single data center.

Or maybe nothing has gone wrong. Maybe we've just arrived at a growing pain moment.

This isn't a failure. This is what fast growth feels like.

The question is no longer whether AI works. It works. The question is whether we're truly making the most of it. And the answer is that we have not yet paused to ask.

It's time to do just that — take inventory of where we are.

We've spent three years in a race for quantity. More parameters. More tokens. More agents calling more agents in deeper and deeper loops. Each layer doing work the layer below could have done for a fraction of a cent, in a fraction of a second, on a fraction of the power. The race made sense when it started; nobody knew yet what these models could do. Now we know.

A great deal of what gets billed as "AI work" is not AI work at all.

Fetching a web page is not reasoning. Stripping the navigation off an article is not reasoning. Turning HTML into clean markdown is not reasoning. These are clerical tasks, the kind a small, deterministic program has done elegantly for thirty years. Asking a sixty-dollar-per-million-token reasoning model to do them is asking a brain surgeon to alphabetize the waiting room.

We have built the most expensive computer in human history and pointed it at filing cabinets.

The opportunity here is larger than a cost-cutting exercise. A new layer of the stack is emerging — deterministic infrastructure built and priced specifically for AI agents. Not the old web, repackaged. Purpose-built primitives that do the predictable work at machine speed, settle for fractions of a cent, and leave the language models free for the work only they can do.

The right shape for an AI system has three layers, not one. Deterministic infrastructure at the bottom, doing the predictable work — reading, parsing, formatting, sorting, validating. Language models in the middle, doing what only they can do — reasoning, judgment, the genuinely novel cut. Humans at the top, doing what only humans can do — bringing taste, intention, accountability.

Each layer earns its keep. Each layer is essential. Each layer leaves the layers above and below free to do what only they can do.

This is not a smaller idea than AI. It is a more honest one.

It is also a more responsible one. Investors deserve software that doesn't burn their capital on string manipulation. Employees deserve tools that get faster and cheaper every year, not slower and more expensive. The grid deserves to keep the lights on for humans, not waste energy on matrix multiplications that could have been a regular expression.

The labs building the great reasoning models are not the villain in this story. They have built something revolutionary, and we owe them the discipline to use it well. Every kilowatt a deterministic pipeline returns to the pool is a kilowatt a frontier model can spend on actual frontier work.

This is what we mean when we say Skim should cost less than two-tenths of a cent. Not because reading the web is unimportant. Because reading the web is solved, and we should take advantage of this and price it accordingly.

The frantic-paced race for quantity is winding down. A saner cycle is beginning — one in which every layer of the stack gets sized correctly, priced correctly, and asked to do only what it does best. The winners will not be the ones who burn the most tokens. They will be the ones who burn the right tokens, on the right work, at the right layer of the stack.

The most powerful tools are the ones that know what they are for.