SaaS Is Dead, Long Live SaaS

The SaaS obituary has been running for about two years now. The argument is familiar by this point. AI writes code, so why pay per seat for software your team could generate over a weekend? Multiples compressed, “software is dead” became a panel topic, and every application company got repriced as if a customer with an AI subscription is a customer about to churn.

I think the obituary was filed early, and I think the reason is that it misidentifies what SaaS companies were selling in the first place.

The Last Time Code Got Cheap

This has happened before, almost beat for beat. Until the mid-1950s, programming a computer meant writing instructions in the machine’s own language… slow, error-prone work, closer to craft than engineering. Hardware was the expensive part and programming was the rounding error. Then computers got cheaper, and something inconvenient happened… programming costs started to overshadow hardware costs. Higher-level languages showed up and made writing programs dramatically easier, everyone celebrated, and the cost of software stayed enormous anyway, because writing the program was never where most of the cost lived. It lived in maintaining the thing, changing it, integrating it, and keeping it running while everything around it moved. IBM eventually unbundled software from hardware in 1969, which is roughly the moment the industry formally admitted that the code was a product of its own, with its own bill.

The industry has a habit of celebrating the collapse of one cost right as the real cost migrates somewhere else. Cheap hardware revealed expensive programming, and cheap programming, courtesy of AI, is revealing expensive ownership.

What You Were Actually Paying For

The SaaS-pocalypse thesis gets one thing badly wrong. Nobody was ever buying code. Code has been close to free at the margin for decades… you could fork an open source project or hire a contractor, and companies mostly didn’t, because the code was the least of what they were paying for. The contract buys a counterparty. Someone else carries the uptime, the security patches, the compliance certifications, the integrations that break when a partner changes their API, and the edge cases already discovered by ten thousand other customers. Someone else answers the phone when the thing falls over during quarter close, which is when things fall over. You’re renting an obligation, and the whole point is that the obligation belongs to somebody who isn’t you. There’s a reason procurement negotiates the SLA harder than the feature list.

AI changed the economics of exactly one item on that list. Producing code got cheap, and everything else stayed priced where it was, which matters because everything else was always most of the bill. The vibe-coded internal replacement makes this vivid. The build is cheap on Friday and expensive for the next five years, because now you own a codebase nobody fully understands, written partly by a model, maintained by whoever hasn’t quit yet. Congratulations, you’re a software company now, involuntarily, in a vertical you never wanted to enter. Total cost of ownership didn’t drop; it moved onto your balance sheet, which is less an argument against SaaS than an advertisement for it.

The Part of the Obituary That’s True

I want to be fair to the bears, because some of the category really is dying. If your product is ostensibly a form sitting on a database, with shallow workflow entanglement and switching costs measured in an afternoon, then yes, your customer can now generate a good-enough replacement, and eventually will. Seat-based pricing makes it worse, since AI shrinks headcount before it shrinks work, and the invoice tracks heads. Thin SaaS is in genuine trouble, and no pricing model saves it. The market’s error was reading the death of the thin layer as a death sentence for the whole category, and marking everything down together.

The companies on the other side of that line are systems of record, sitting inside workflows so entangled that ripping them out means re-certifying processes that took years to certify the first time, while years of data pile up underneath. And, awkwardly for the obituary writers, they’re the best-positioned buyers of the technology that was supposed to kill them. They own the workflow and the data, and they already have distribution into a captive base, so AI lands in their P&L as a feature they sell rather than a threat they absorb. I’ve written before that hard to build and hard to displace are different properties. Complicated-system SaaS scores low on the first and high on the second, which is roughly the best quadrant available to a software company.

So what’s left looks like a mispricing. The market repriced an entire category for a cost collapse that only hit the cheap part of it, and somewhere in that markdown are companies whose ownership moat got stronger over the past two years while their multiple got cut in half. Markets are quick to price the first-order story and slow to price the second-order one, and the gap between those two is where I’d rather be looking than wherever the crowd went next. For my purposes that logic also points forward, at the vertical systems of record still being built today in domains too messy for the weekend demo to survive contact.

What Would Change My Mind

The ownership moat holds only as long as AI is better at writing code than at owning it. If agents get genuinely good at maintaining codebases they didn’t write… holding context across years, catching regressions, carrying the pager at 3am, surviving an auditor’s questions… then ownership costs collapse too, and the line between thin and thick SaaS stops protecting anyone. The early signal will be an enterprise actually decommissioning a system of record and living to tell the story, rather than another internal tool built at the edges. I haven’t seen one yet.

Until then, the obituary reads to me like it was written by someone who thought the typing was the expensive part.