The subsidy era in AI isn't ending with a price increase. It's ending with the thing you bought getting smaller while the invoice stays the same. Last week users of one of the major consumer agents started hitting usage limits that had never existed, mid-plan, with no notice, and the explanation from the top was that the cost of serving the product had run ahead of what it earned. The plan price didn't move. The capacity behind it did. Every flat-rate AI plan carries the same unwritten term, and the enterprise version is worse than the consumer one, because team plans are how most companies budget AI and almost none of them have a capacity floor in the contract. We call it the shrink clause. It says the vendor can cut what your plan delivers without changing what your plan costs, and you'll find out when your people hit the wall.
What is the shrink clause?
It's the unwritten term in every flat-rate AI plan that lets the vendor reduce delivered capacity while holding the price.
We've written two clauses before this one. The availability clause is access ending on the vendor's timeline. The crossfire clause is access ending over a fight between other companies. The shrink clause is less visible and more common than either. Nothing ends. Your seats still work. What changes is how much work each seat can do before it's throttled, which model answers by default, how long the context window stays, how many agent runs you get in a day. Those are the real product, and they live in a settings table the vendor controls, not in your contract. A price increase gets a renewal conversation. A capacity cut gets a support ticket, and by the time finance hears about it the quarter's work is already behind.
Why is this happening now?
Because flat-rate plans were priced below the cost of serving them, and the gap is coming due.
The economics were never a secret. A heavy user on a two-hundred-dollar monthly plan can consume compute that would bill at several times that on the API, and the difference was covered by investors who wanted the users. That worked while growth was the goal. It stops working the moment a vendor needs the product to pay for itself, and several now do. The first lever every vendor reaches for isn't the price, because a price increase is visible, triggers renewals, and sends customers to competitors. The quota is invisible until someone hits it. So the quota moves first. For a consumer that's an annoyance. For a company running agent workflows through team seats, it's a capacity cut to an operation that was sized on yesterday's limit.
Isn't this just a pricing change by another name?
No. A pricing change tells you what happened. A capacity cut makes you discover it.
The distinction matters for budgeting. When a vendor raises prices, your cost per unit of work goes up and you can see it, model it, and decide. When a vendor cuts capacity, your cost per unit of work goes up by exactly the same logic, but the invoice doesn't change, so nothing in your finance system flags it. What flags it is a workflow that stops at two in the afternoon, an agent that fails because its quota ran out mid-task, or a team that starts rationing a tool you're paying full rate for. We described a cousin of this in default drift, where spend moves on the vendor's release schedule. The shrink clause is the mirror image. Spend holds still and the capacity underneath it moves, and both leave you running a stack whose economics you didn't set.
How exposed is a typical company?
More than it knows, because the exposure is the size of the subsidy, and nobody has measured the subsidy.
Most companies run two kinds of AI spend side by side. API usage, metered and visible, where you pay for what you consume and the shrink clause can't touch you. And plan seats, flat-rate and bundled, where the price is fixed and the capacity is whatever the vendor says it is today. The second category is where most of the headcount sits, and it's also where the subsidy lives. If your team's actual usage on a flat plan would cost three times the plan price at metered rates, you're sitting on a subsidy the vendor is going to collect, and the shrink clause is how it collects. A company that has never run that calculation has no idea how big the hit is or where it lands. The ones that have can see which teams are living on borrowed capacity and plan for the day it's withdrawn.
What does protection look like?
Price every plan at metered rates, write a capacity floor into any contract that matters, and keep the workflows you can't afford to lose on usage you control.
Start with the measurement. For each flat-rate plan, estimate what the team's real consumption would cost at API prices. The gap is your subsidy, and it's also your exposure. Where the gap is small, the plan is fairly priced and the shrink clause can't hurt you much. Where it's large, you have a decision. Then the contract. Enterprise agreements can specify delivered capacity, model tier, context limits, and notice periods for changes, and most companies have never asked, because nobody thought the quota was the product. It is. Finally, the architecture. Workflows that matter to the business shouldn't run through a plan seat at all. They belong on metered usage behind a boundary you own, where the cost is visible and the capacity is whatever you're willing to pay for. We've made that argument from the cost side in the frontier freeze. The shrink clause makes it from the reliability side. The plan seat is for people. The production workflow is for the meter.
What should you do about the shrink clause?
The question you're really asking is which of your AI spend is sitting on a subsidy that's about to be collected. Run one afternoon of arithmetic and you'll know.
List every flat-rate AI plan your company pays for. For each, pull the usage your team actually generates and price it at the vendor's metered rate. Rank the plans by the gap. The top of the list is where the shrink clause will land first, because those are the plans the vendor is losing the most on. For each of those, decide whether the work belongs on a seat at all, and for the work that stays, get the capacity written down before the next renewal. The vendors will keep shrinking plans, because it's the cheapest lever they have and most customers never notice until the wall. The companies that measured the subsidy will have moved their real work off the seat before the wall arrived. Everyone else finds out at two in the afternoon.
If you want your AI spend sorted into what's subsidized and what's yours, with the production work on a meter you control, start with an AI Blueprint or reach us at contact@theyor.com