AI Made Software Cheap to Build and Expensive to Rent

9 min read

Lovable reached $200M ARR with fifteen employees, about $13.3M of revenue per head. Over roughly the same stretch, SaaS prices rose 11.4% while general inflation ran at 2.7%.

Almost nobody has put those two numbers next to each other.

The first says building software got dramatically cheaper. $13.3M per head would have looked like a typo five years ago, and whatever you think of vibe coding, the cost curve it exposes is real.

The second says your bill went up anyway. Businesses now spend around $7,900 per employee per year on SaaS, a 27% increase in two years, and 2026 renewals carry a line item that did not exist before: an AI pricing uplift of 20 to 37%.

Production cost fell through the floor. Retail price went up by double digits. That gap will not quietly close. It is the shape of the next few years, and it is why a model everyone declared dead, pay once and own it, is coming back.

Subscriptions were a cost argument. The cost collapsed.

Subscription software was justified by a reasonable argument. Software is a service rather than a product. It needs servers, patches, and continuous development, so you are funding an ongoing relationship instead of buying a binary. Fair enough. For twenty years that was mostly true.

But that was always a cost argument, and the cost just collapsed. When a fifteen-person company can service two hundred million dollars of revenue, “we need your $49 a month to keep the lights on” is doing less work than it used to.

What replaced it is a leverage argument. The reason you are on a subscription in 2026 is that your data is in there, your team’s habits are in there, and the switching cost is high enough that an 11.4% increase does not move you. Vendors price that switching cost, because they can.

You can watch that pricing power get used. 78% of IT leaders were hit with unexpected AI or consumption charges in the last twelve months, and 61% cut planned projects because of price increases. Those companies are absorbing a bill they did not choose and cannot easily refuse.

Agents move the meter from seats to usage

Here is the part that should worry you more than the price rises, because it is structural.

Per-seat pricing is dying, and not as a prediction. A Forbes tech council piece in March 2026 called it the seat apocalypse: teams shrinking from 50 to 10 while software consumption grows 10x. Pure per-seat is now roughly 8% of the market, and Gartner expects at least 40% of enterprise SaaS spend to move to usage, agent, or outcome-based models by 2030. The 8% figure comes from one analysis, so read it as an order of magnitude rather than a measurement.

The reason is obvious once you see it. Seats priced humans. Agents do the work without occupying a seat, so the value the software delivers holds steady or rises while the number of people you are billed for falls. No vendor will accept that, so the meter moves from your headcount to your usage.

Think about what that means with an agent in the loop.

Under per-seat, your bill was bounded by how many people you hired. It was predictable, and you could plan around it. Under usage pricing, your bill is bounded by how much your AI thinks. Every exploratory query, every retry, every time an agent reads your CRM twice to be sure, all of it becomes a line on an invoice.

The whole promise of agentic work is that you let the machine be thorough, because being thorough is suddenly cheap. Usage-based pricing takes that exact behaviour and makes it the thing you are charged for. You end up with a financial incentive to keep your agents lazy.

That is a pricing model at odds with how you now want to work.

Why perpetual licences died the first time

The anti-subscription case is usually made badly, so here is the honest history.

Perpetual licences did not lose on ideology. They lost because they had a real flaw: updates stop, support expires, and the software goes obsolete in three to five years, forcing an expensive repurchase. “Own forever” often meant “own a fossil”. Vendors hated the revenue lumpiness, buyers hated the upgrade cliff, and subscriptions solved both.

There is a sharper version of the objection aimed straight at products like ours. A one-time licence on AI software is a promise nobody can keep, because inference costs money forever and no single payment covers an unbounded future compute bill.

That objection is correct, and it is why most “lifetime deal” AI tools are quietly insolvent. It is also the objection local-first architecture answers, but only if you actually build local-first, which most of them did not.

The demand never went anywhere, incidentally. Even mid-subscription-boom, over half of companies expect perpetual licence revenue to grow, specifically for cost predictability, indefinite usage rights, and independence from vendor changes. What was missing was a way to serve that demand without bankrupting the vendor.

Three things that make buy-once solvent again

None of them was true in 2015, which is why the model failed then and can work now.

Build cost collapsed. A small team can now build and maintain surface area that used to need forty engineers. The revenue base required to sustain a product dropped by an order of magnitude, which is what lets a one-time price cover it.

Distribution stopped needing a sales team. The other half of the SaaS cost structure was go-to-market rather than engineering, and that is collapsing too.

Compute moved to the edge. This is the one that answers the objection above. If the software runs on your machine, there is no per-user server cost for the vendor to recover forever. And if you bring your own AI key, the inference bill is yours, at cost, uplifted by nobody. The unbounded cost that makes lifetime AI licences a lie only exists when the vendor sits in the middle of every request.

Take the vendor out of the middle and the arithmetic works.

Local-first is what makes the price possible

Ion Alpha runs on your machine. Your data lives in a local database you can open, back up, and read without asking us. The AI runs against your own provider key, so you pay inference at cost with no markup, and we have no meter to run because the requests never come through us.

The pricing follows from the architecture. We can sell Ion once because we are not carrying a per-user server bill or a per-query inference bill on your behalf. A cloud product making the same promise would be underwriting your usage forever, and would eventually have to break it. Most do.

It fixes the incentive problem too. When your agents are thorough, that costs us nothing, so we have no reason to want them lazy. Run them all night. The only bill that scales is the one you pay your model provider directly, at their price.

The agent runtime follows the same logic. Flynn, the agent embedded in Ion Alpha, is Apache-2.0 open source and ships as a single binary you can keep, fork, or run without us entirely. It seals every run into a signed, tamper-evident record of what it did and under what authority, in the open Provetrail format, so what you own covers the account of the work as well as the work.

As for the old buy-once flaw of owning a fossil: that was a build-cost problem, and build cost is exactly what fell. Continuing to ship is affordable now in a way it was not when perpetual licensing died.

What owning your tools costs you

Owning your tools means owning your problems.

Your backups are yours. If your laptop dies and you had no backup, there is no support desk with a copy. Local-first means you are the operator, and some people genuinely do not want that job. For them SaaS is the right answer, and we would rather say so than sell you something you will resent.

You also get no vendor to blame. When a cloud tool breaks, someone else is contractually on the hook. When your local system misbehaves, it is your machine, your OS, and your afternoon.

And “own forever” is a claim about the software you have, not a promise of infinite free future work. We will be straightforward about what is included and what future major versions cost, because the alternative is the lifetime-deal trap everyone is rightly sceptical of.

We think the trade is worth it. We also think you should make it with your eyes open, which is more than the renewal quote in your inbox is offering.

$7,900 a year against a machine you own

$7,900 per employee per year, rising 11.4% annually, with a 20–37% AI uplift stapled on at renewal and a meter about to start counting how hard your agents work.

Against that: a machine you already own, running software you already paid for, with your data in a file you can copy.

For twenty years the first column bought you something the second could not offer. AI spent eighteen months erasing that difference, and then the price went up anyway.