For fifteen years, software had the best business model in public markets: roughly 80% gross margins, 33% free cash flow margins, and revenue you could see years out, because it was just seats times price times contract length. Investors treated ARR like an asset class and paid accordingly, often 20 to 30 times ARR at the peak, not 20 to 30 times earnings.
Then came the question every software investor is now forced to answer: if AI drives the cost of building software toward zero, what exactly were we paying 20x for?
Matthew Kanterman and I took opposite ends of that question on the latest Pixels & Profits. I am the bear. My view is that SaaS may not die outright, but it has been, and will stay, on a trajectory of rationalization: many existing SaaS companies get killed, and margins and multiples keep coming down across the board. Matthew's view is more nuanced: the business model is broken, but the companies aren't all doomed. Where an application sits in the organization, and how much critical data it controls, decides who survives.
We agree on more than you'd expect. We disagree on the part that matters most for your portfolio: timing.
What the 20x multiple was actually paying for
Matthew opened with the Wu-Tang version of the thesis: cash rules everything around me. SaaS printed money, and the reason investors paid up was visibility. You knew the customer count, the ACV, the gross ARR, and, most importantly, net retention. For a lot of these companies, existing customers grew 110% to 120% a year, through extra seats, upsells, and plain pricing power. Incremental gross margin on that new revenue was effectively 100%.
Strip it down and the multiple rested on three assumptions:
Revenue renews with almost no effort.
The next seat costs almost nothing to turn on.
Switching costs are high.
Break any one of those and the multiple has to rationalize. AI, in my view, breaks all three.
The SaaSpocalypse was the first repricing, not the last
The market already ran this experiment once. In the March 2026 "SaaSpocalypse," some software stocks fell 50% to 60%. The credit market was even more alarming. As Matthew pointed out, the broader leveraged loan index was down about 1.5% year to date, a bad year but not a disaster, while software leveraged loans were down almost 7%. Credit does not move 7% in a few weeks very often, and he recalled that the share of distressed or near-distressed software issuers was over 40% at the time, based on loan and bond yields. Investors were not just saying multiples were compressing. They were saying some of these businesses might not exist in six months. These had been the marquee assets of the stock market.
Then the bounce. Salesforce recovered strongly year to date. Even Atlassian came back more than Matthew expected. The story behind the bounce is simple: the first narrative was "AI can write any software," and then everyone discovered that a vibe-coded prototype is not a production system. I know, because I did it. I replaced Monday, our project management tool, with something I built myself. It was not easy, and it was not a weekend.
But the private market is where you see the honest marks. Bending Spoons is effectively buying private SaaS companies for the capital they raised: pay the VCs back dollar for dollar and take what's left for nothing. Miro, valued at around $17.5 billion in 2022, agreed to be acquired for about $1.355 billion in enterprise value, roughly a 90% markdown. Airtable, which peaked at over $11 billion, sold to the same buyer at a deep discount weeks earlier.
That's the repricing without the noise of a public ticker.
The seat is the wrong unit
Classic SaaS charges per human seat, often prepaid a year up front (that prepay is what finally pushed me to build my own Monday replacement). It expands by adding more humans, at almost no marginal cost.
Matthew's point is that the seat is no longer the right unit of value, and vendors already know it. Four of his core vendors are pushing him toward consumption pricing, from Salesforce to smaller AI tools like Rogo that started seat-based and are switching. Consumption isn't a bad model. Shopify clips a coupon on every dollar sold through its stores, and investors are comfortable with that. But consumption lowers switching costs and makes revenue much less visible. The core question becomes: what does revenue durability look like?
The reason this shift is happening is that the user is changing. Increasingly it isn't a human clicking around a website. It's an agent working on behalf of a human, through an API or an MCP control surface. If one person can spin up a hundred, a thousand, or ten thousand agents hitting your service, seat-based pricing works against the vendor, not for it.
And there's a second hit that rarely gets discussed: the hidden margin disappears. Every company I've worked at has unused seats, people who want access but never log in, licenses nobody cancels. SaaS vendors quietly profited from that slack. Matthew sees the same thing as a Claude Enterprise admin: there are power users and there's a long tail of light users, and under seat pricing the tail subsidizes the power users. Consumption pricing makes that visible. List-price theater, multi-year discounts, and negotiated minimums are a lot harder to hide when you're billing fractions of a cent per token.
Yes, some vendors will benefit, the Jevons case where thousands of agents simply drive more usage. (Matthew has muted "Jevons paradox" on X; he's heard enough of it.) But I think the blade cuts more toward SaaS companies than away from them.
The bear case: AI deletes scarcity
Here's my core argument. SaaS sold scarcity. The multiple assumed software was hard to build, hard to copy, and very hard to leave. AI deletes that scarcity.
Features that took a team a year now take a team weeks. Today not everyone can vibe code a production system, and the gap between "prototype that looks cool," which is what spooked investors, and "production-ready, reliable, usable" is real. But it is closing. Technology takes time to diffuse, and the models keep getting dramatically better. If your moat was shipping a feature first, or that refactoring is expensive, that's no longer a moat. It's a GitHub repo and a series of prompts.
And it isn't only custom builds. Open source is increasingly better than the commercial product. I recently switched from a paid voice-transcription app I'd bought a lifetime license for to its open-source alternative, because the open-source version was simply better. Most SaaS is built for a broad user base, so it's bloated. An organization often needs only a specific slice, and it can now build or install exactly that slice, customized for its use case.
"System of record" is a moat. It's not a permanent one.
Silicon Valley has settled on a convenient line: regular SaaS is in trouble, but system-of-record SaaS is safe. A system of record is the business truth: customer and employee data, orders and charges, the financial ledger.
They're not wrong, today. In theory, good architecture separates the data from the layer that operates on it (the "D" in SOLID is dependency inversion), so swapping the layer on top should be easy. In practice it never is: there are workflows, permissions, audit trails, employees trained on one way of working, and M&A-driven Frankenstein stacks. Matthew adds regulation and oversight on top. Imagine the diligence call: "Who was running audit controls on the books last month?" "Claude was."
Today, switching off a Salesforce-class CRM can take a large company a year and $500K to $1M; a Workday or SAP-class migration can take one to two years and millions. Mid-market companies are an order of magnitude less.
My argument is that this number only goes down. AI is already rewriting legacy systems once considered intractable, like COBOL. Context windows are huge and growing. A year and a million dollars becomes six months and $500K, then three months and $50K. As that happens, CIOs are forced into a simpler calculation: what value am I getting, and what does the alternative cost? When CIOs think that way, multiples come down.
Matthew's pushback is the strongest point against me, and I agree with it: it's not just dollars, it's risk. CIOs are extremely risk-averse. He cited a large medical device company whose SAP migration to S/4HANA crashed on go-live day and cost it hundreds of millions of dollars. That fear will be the biggest thing keeping customers on incumbents, and keeping their multiples artificially high. But anyone who has worked at a big company knows the pattern: everyone fears firing the person who obviously needs to go, and afterward everyone says it should have happened sooner. The fear is real. It's also temporary.
Incumbents can't ship on the new clock
My last bear point is organizational speed. Why does Outlook still feel the same? Why does everyone complain about Jira? Incumbents aren't going to ship on the new clock that AI enables. The best AI-native builders today can ship on the order of 2,000 PRs a month versus roughly 20 for a typical engineer. That's about 100x in speed and cost, and while it takes time to diffuse, it will.
I'll give a concrete example. My vibe-coded project management tool now integrates a wiki and an MCP server, so my agent writes our documentation for us. I won't switch back to anything, because no vendor offers that. Why can the CEO of a small game studio build this when the big vendors can't? Because their engineers aren't incentivized to make the big changes they need to make.
Matthew's fair counter: their customers move at the same slow pace. Very few organizations move like a startup or a frontier lab. So I expect the revenue loss to come from the bottom up: startups and SMBs switch first, and larger enterprises follow over time.
Trends to watch
1. The economic depreciation rate of software is collapsing. This is Matthew's framework, and I think it's the most important idea in the episode. Take AppLovin: it built the best ad-targeting model in its market with a multi-year head start. Now the next model is more expensive to build (compute isn't cheap), and every competitor has the same tools to catch up fast. Even if AI cuts development costs, competition forces you to rebuild far more often. If you have to rebuild your product every two years instead of every five, the window to earn back your investment shrinks, and IRRs fall for every company that builds and sells software.
2. Paying for taste instead of software. An endpoint exercise, not a prediction: if software is nearly free to produce, you stop paying for the software and start paying for a workflow, a point of view, and taste. That's a word Silicon Valley borrowed from games. You'd buy a team's perspective on how project management or recruiting should work.
3. The API becomes the product. The control surface used to be a website. For agents, it's an API. Matthew points to Salesforce's new Claude-based offering, which he notes is the first "Force" product made by someone else in Salesforce's history, as a company disrupting itself before it gets disrupted. It may keep Salesforce relevant, but consumption shifts to the model's interface instead of Salesforce's own.
4. Organizational hardness. Many companies will struggle to become AI-native because AI blurs every role. "PM," "engineer," and "marketer" stop being fixed boxes; the right role depends on the person (Harry is a great PM-engineer, Sally a great engineer-marketer). Flexible organizations win. Rigid ones, which I'd call organizationally hard, won't make it. This is a Darwinian era.
5. Private equity becomes the AI enablement layer. Matthew expects a wave of software take-privates, and something less obvious: LBOs of traditional businesses specifically to make them AI-native, because that transformation is too disruptive to do as a public company. PE is sitting on more than $1.2 trillion of dry powder and is hiring AI operating partners. The value isn't glamorous. If an industrial distributor cuts its receivables by two days with AI, that unlocks real cash flow. In Matthew's words, this will matter more to the economy than whatever the frontier labs announce tomorrow.
I'd add that we'll see acquisitions for the customer list, relationships, and distribution, where the buyer guts the internals and brings in an AI-native team to rebuild the product fast and cheap. A company that's organizationally hard, with a product people tolerate but are stuck on, is exactly that kind of target.
6. Economics shift to the long tail. Smaller teams will carve out specialized niches that broad SaaS never served well. Not Salesforce-sized businesses, but smaller, durable ones with dedicated audiences. Vertical AI tools (Rogo in finance, Harvey in legal) are already better than general models at their workflows. The open question Matthew asked me, and it's the right one: building it was never the hard part. Can you keep it running? My honest answer is that the prototype is easy, and robustness at scale under many users is the hard part, but AI is getting better at exactly that too.
Where we each land
I think SaaS is ass. System-of-record companies are safer, with more friction and a bigger moat, but it isn't sustainable; it's a matter of time. SaaS as a category stays under attack, many companies go out of business, and multiples keep declining. The SaaSpocalypse rebound was a head fake. It's coming back.
Matthew sees divergence, not a wipeout. Last time, everything got dumped and investors sold first and asked questions later. This time, separating the survivors from the companies that get disrupted out of business is where the alpha is. For once, real stock-picking alpha, not a category bet. His closing advice: be prepared for volatility, especially if the foundation-model IPOs arrive this year. We saw it with SpaceX, and those would be SpaceX on steroids.
Each view has a clean falsifier:
I'm wrong if switching costs and CIO risk aversion hold up longer than the improvement curve. If migrations stay year-long, million-dollar, career-risking projects through the next several model generations, system-of-record companies keep their multiples.
Matthew's divergence trade is wrong if the friction collapses faster than the market can tell winners from losers. Then it isn't a stock picker's market, it's another category-wide repricing.
Key Takeaways
The 20-30x ARR multiple rested on three assumptions: effortless renewal, near-zero marginal seat cost, and high switching costs. AI pressures all three.
The SaaSpocalypse was a preview. Stocks fell 50-60%, software leveraged loans fell almost 7% versus about 1.5% for the broader index, and private marks like Miro (about a 90% markdown) show where value really sits.
The seat is dying as the unit of value. Agents, not humans, increasingly consume software, which pushes pricing toward consumption, lowers switching costs, and erases the hidden margin of unused seats.
System of record is a moat, not a fortress. Switching costs are falling from a year and $1M toward months and tens of thousands, but CIO risk aversion is the real brake.
The economic depreciation rate of software is accelerating. Rebuilding every two years instead of five compresses IRRs across all software.
Watch the bottom of the market first. Startups and SMBs switch first; enterprises follow.
Private equity becomes the AI enablement layer for the middle market, through software take-privates and LBOs that make traditional businesses AI-native.
Bear vs. divergence: I expect a broad decline; Matthew expects a stock picker's market. Either way, expect volatility.
This article is for informational and educational purposes only and is not investment advice. Figures are as discussed at the time of recording and should be verified against current data before acting.