The subscription was an insurance policy

Flat-rate software sold you cost certainty for free. That coverage is being withdrawn.

When AT&T stopped selling unlimited data in 2010, the customers who still had the old $30 plan would not let go. They turned down phone upgrades to keep it. They stopped calling customer service in case it flagged their account for a switch. Most of them never came close to using the unlimited part. That was never what they were holding. They were holding the one thing the plan still guaranteed: the certainty of what the bill would be.

AT&T eventually throttled the heaviest of those subscribers, which is what an insurer does when it cannot reprice a policy it can no longer afford. It rations instead.

Software has been selling the same policy for twenty-five years, and it is being withdrawn right now, on roughly the same terms.

The seat was an insurance product, though nobody called it that. The vendor charged every customer the same premium and absorbed the demand variance, because serving a heavy user cost roughly the same as serving a light one. Marginal cost near zero made the underwriting free. So the flat subscription bundled two products: the software’s capability, and the predictability of its cost. The second one was invisible because it cost the vendor nothing to provide. Procurement cycles and annual budgets assumed usage risk was the vendor’s problem, and per-seat licensing depended on it.

This was an odd insurance business. A normal insurer lives in fear of its heaviest claimants, and the entire actuarial profession exists to find them and price them. Software had no such problem. The heavy user filed claims constantly, in the form of usage, and every claim cost nothing to pay, so the vendor could insure everyone without underwriting anyone. Flat-rate software was the only insurance business in which writing the policy was free, which is also why nobody noticed it was insurance. A policy that costs nothing to honor never has to announce itself.

AI broke the underwriting math, because now every query has a real marginal cost. The enthusiastic customer, once free to serve, is a loss.

Cursor was the first to say so out loud. In June 2025, its $20 Pro plan went from 500 fast requests plus unlimited slower ones to $20 of usage metered at model API rates, and users burned through the allotment in days. Michael Truell apologized and refunded the unexpected charges, treating the episode as a communication failure. What he was apologizing for was a risk transfer executed without naming it. The plan still cost $20. What moved was who carried the demand variance, and the customer found out by receiving a bill they had never had to think about.

GitHub showed what the same move looks like when a company knows exactly what it is doing. In June 2026, it replaced Copilot’s request units with credits metered against token rates and left every plan price untouched. The subscription did not get more expensive. It stopped being a subscription and became a prepayment. And GitHub drew a line most of the industry had not thought to draw: code completions still come with the plan and consume nothing, while agents and premium models draw down the allotment. That line runs precisely between predictable cost and volatile cost. Sorting your own catalog by risk and pricing each side accordingly is underwriting, whatever the invoice calls it.

Anthropic showed what it looks like to know and not be able to stomach it. Fable 5 sat inside the Pro and Max plans for about two weeks before being pulled onto metered credits at double the price of the flagship it replaced. Then Anthropic extended the deadline, and then extended it again. A company that reschedules a cutover three times is caught between subscribers who will not forgive the repricing and economics that cannot absorb it. The flinching told you more than the decision did.

Not every software company is standing in this line. A CRM that stores records and renders views still costs almost nothing to serve, and its seat is safe. The exposure runs exactly as far as inference does, through every product that spent the last two years bolting on a copilot. That is a subset of the category. It is also where nearly all the new money is going.

It would be easy to read this as opportunism, but the balance sheets say otherwise. Amazon alone is planning something near $200 billion in capital expenditure this year, and the largest builders together are spending close to double last year’s. When the cost of serving a customer grows at that rate, revenue that does not track usage becomes a solvency problem. No underwriter can carry a book whose most enthusiastic customers file the largest claims.

Flat pricing worked by pooling every customer into one price, so the light users quietly paid for the heavy ones. That pool cannot re-form once each customer can see what they actually cost, which is what makes this permanent rather than cyclical. Insurance has a name for what happens once risk becomes visible: adverse selection. Light users can now see what they consume, so they demand cheaper plans or leave. Heavy users go hunting for whatever flat-rate offers remain, precisely because those offers still subsidize them. Any vendor still selling unlimited attracts exactly the customers who make unlimited unwritable. Metering did not merely change the price. It destroyed the conditions under which the old price was possible, and no rollback or apology restores them.

Enterprises absorbed a shift like this once before, with cloud, and it took a decade and a whole discipline to manage. But cloud variance landed on the infrastructure budget, which engineers already owned and could attribute to a workload. AI usage lands on the software line, which companies procure annually, budget flat, and hand to people with no telemetry. There is no stable unit to plan in either, because cost per token does not map to cost per outcome.

So buyers reach for the one lever they can see, which is the model itself. Cheaper models are gaining ground, chosen on price rather than on what they can do. Trading capability away for cost predictability is what a buyer does when a risk is unpriced and unmanageable. Because that premium is paid in capability rather than cash, it stays off the invoice, so nobody has to account for it.

The rest of the bill is already visible. KPMG’s most recent survey of senior leaders found that nearly half of organizations have rephased AI deployments because the costs outran the expected value. That is not a story about disappointing technology. It is a story about a line item that used to be fixed, has started moving, and is owned by people with no way to make it stop.

Somebody will sell them the certainty back. It will come packaged as committed-use contracts, spend caps, fixed-price wrappers around metered intelligence, insured tiers with real ceilings. The shape hardly matters. What matters is that whoever sells it is writing the policy the flat subscription used to include for free. This time it has to be underwritten by someone who has looked at the customer’s workload and priced the risk of serving it. Software is about to acquire the one function it has never in its history needed, which is an actuarial one. Vendors will have to know which customers are expensive before they sign them, the way an insurer does.

That sorts buyers into two groups, and it has already begun. A company that can measure its consumption, attribute it to a workload, and demonstrate what it costs to serve walks into that negotiation as a known risk and gets priced like one. Everyone else arrives with nothing, and an underwriter handed no data prices for the worst case, because that is the only prudent thing to do with an unknown. This is the oldest asymmetry in insurance. It is about to hit a category that has never once had to think about it.

The advantage this year goes to whoever does the unglamorous thing first. Somebody is going to learn the unit economics of your workloads and find out which of your teams are the heavy claimants, and the only question is whether it is you or the vendor setting your renewal price. The buyers who get there first will find they are cheaper to serve than the market assumes, and they will be paid for the proof. The buyers who wait will pay the premium charged to everyone who could not show otherwise.

AT&T’s customers understood the plan was never really about the data. Software’s customers are about to understand that the subscription was never really about the software. It was the certainty. And certainty was the one thing that could not survive a real marginal cost.