The Subsidy Cliff

You are not paying for AI. You are being paid to use it.

Go look at the pricing page of any major AI lab. OpenAI, Anthropic, Google, pick one. What you are seeing is not what AI costs to deliver. It is what these companies are willing to lose, per API call, per enterprise seat, per integration, to make sure you build your work around their product. The prices are customer acquisition costs wearing a product catalog's clothes.

This is not a secret. It is the business model. And if it sounds familiar, it should.

We Have Seen This Before

From 2020 to 2022, zero-interest-rate capital flooded every part of the economy. Not just technology. Capital was essentially free, and organizations spent accordingly. They hired aggressively. They expanded into markets they had no business entering. They built operational structures that assumed cheap money was a permanent condition.

It wasn't. And the correction was brutal.

But the interesting lesson from ZIRP is not that the correction happened. Everyone knows that part. The interesting lesson is who survived it and why.

The organizations that emerged strongest were not the ones that avoided spending during the cheap period. They were the ones that spent aggressively on things that stayed valuable after the subsidy ended. Companies that used free capital to hire genuinely excellent people, build real technical advantages, and establish durable market positions came out ahead. They treated the subsidy as an accelerant, not a foundation.

The organizations that got destroyed were the ones that mistook the subsidy for the economy. They hired bodies instead of building capability. They expanded headcount as a substitute for building systems. They optimized for a subsidized environment instead of building for the real one.

The same split is forming right now with AI. And most organizations are on the wrong side of it.

The Dependency Is the Product

I noticed this in my own work before I understood the pattern. We had been running a particular AI workflow for about four months. API costs were negligible. Then I did a back-of-the-envelope calculation on what the workflow would cost if pricing moved even two or three times higher. The number was uncomfortable. Not because the workflow was wasteful, but because we had designed the entire process around the assumption that inference was practically free. We were making architectural decisions, staffing decisions, product decisions, all downstream of a price that someone else was choosing to charge below cost.

That was the moment I started thinking about this differently.

The usual framing gets this wrong. "AI will get more expensive, so be careful" is too simple and leads to the wrong response.

The subsidy is not an accident. It is a deliberate strategy. Every decision you make at current pricing, from an engineer's workflow to a CFO's budget projection to a procurement team's vendor contract, becomes harder to reverse as time passes. The dependency accumulates the same way regardless of where in the organization it forms.

AI labs know this. That is why they are spending billions to make sure the prices feel irresistible right now. Every quarter you operate on subsidized inputs, your switching costs climb. Your habits form around specific tools. Your architecture assumes specific providers. Your budget models bake in costs that are not real market prices. It all gets harder to unwind, and that is the point.

This is the same playbook Uber used to destroy taxi economics before repricing the ride. The same playbook every venture-subsidized marketplace has ever run. Get the adoption first. Reprice later.

A CFO building next year's AI budget on this year's pricing is making the same mistake as the startup that hired forty people because money was free. She is planning around an input cost that someone else is temporarily absorbing. When that absorption stops, every line item downstream shifts. Procurement teams locking into three-year tool contracts at current rates face a version of the same problem. So does any startup whose unit economics only work because a venture-backed lab is eating the real cost of inference.

The ZIRP Survivors Spent. They Just Spent Differently.

The answer is not to pull back. That would be the wrong lesson from ZIRP too. The organizations that waited out the free money era did not win. The ones that spent wisely did.

This is the cheapest AI will ever be relative to its capability. The window where you can build deep organizational competence at subsidized rates is open right now. It will close. When it does, the organizations that used the window to build real skill will have a massive advantage over the ones that waited.

But there is a critical distinction between two types of spending. And the way to see it clearly is to ask one question: what breaks when the price doubles?

Some AI investments get more valuable as prices rise. An engineer who develops genuine judgment about when AI output is good and when it is wrong. A team that builds the discipline to evaluate, not just consume. An organization that develops the institutional muscle to integrate AI into how decisions get made, not just how code gets written. These investments compound. They do not depend on a specific price point.

Other AI investments are only viable because the inputs are cheap. The workflow that throws every question at a model because it costs pennies. The enterprise transformation program that looks brilliant because experimentation is essentially free. The product whose margins only work because a venture-backed lab is eating the real cost of inference. When the subsidy ends, these don't just become more expensive. They become untenable.

The ZIRP survivors were not cautious. They spent. But they spent on competence, not dependency. The AI subsidy survivors will be the same.

The Uncomfortable Math

Here is the part that most organizations do not want to confront.

We have been building AI strategies on top of pricing that cannot persist. The major labs are burning through capital at rates that require either enormous price increases or revenue models we have not seen yet. The current pricing is a bet by investors that market dominance will eventually justify the losses. That bet may pay off. But the organizations absorbing the subsidy are not the ones who will benefit when it does.

When pricing normalizes, and it will, the repricing will not be gradual. ZIRP did not unwind slowly. Interest rates moved, and suddenly the entire economics of a generation of companies changed overnight. Organizations had weeks, not years, to adjust.

AI repricing will follow the same pattern. A model upgrade that costs twice as much to run. A pricing tier change that resets enterprise contracts. A provider that stops subsidizing a product line because the unit economics never worked. These are not hypothetical scenarios. They are the expected trajectory of a market where every major player is losing money.

The organizations that will handle this well are the ones that already know the answer to the question: if our AI costs doubled tomorrow, what would we cut, what would we keep, and how fast could we adapt? If you cannot answer that question today, you are building on someone else's foundation.

What Remains When the Price Changes

We have been here before. The specific technology changes, but the pattern does not.

Cheap inputs create a window. The window rewards investment. But it punishes organizations that confuse subsidized access with sustainable strategy.

The honest truth is that nobody knows exactly when the repricing happens, or how sharp it will be. We are making our best guesses based on a pattern we have seen before, applied to a market that moves faster than any previous version. We could be wrong about the timeline. We are not wrong about the direction.

The organizations investing in AI right now are doing the right thing. The question that keeps nagging at us is simpler than it sounds: when the price tag finally reflects reality, will what you built still hold?