← ALL WRITING
ESSAY / 26 OCT 2026

Seat-Based Software Grows With Hiring, and Hiring Has Stopped

Software stocks and loans are being marked down on the idea that AI is replacing the products. The numbers point to stalled seat growth and an overleveraged 2021 vintage instead.

Danial Hazarika6 MINUTE READ
MARKETS / SOFTWARE / CREDIT
Written October 26, 2026. Figures are as of that date. This is research, not investment advice.
CONTENTS ↘

The story the market settled on

Software stocks fell almost 30% between October 2025 and February 2026, according to the Bank for International Settlements. The explanation that took hold, usually under the name SaaSpocalypse, is that new AI models can do much of what software-as-a-service products do, at a fraction of the cost, and that companies will increasingly build their own tools instead of paying for seats.

The losses are showing up in private credit too. Private credit funds had lent more than $500 billion to software companies by the end of 2025, about 19% of all their direct loans. In a Reuters analysis of filings from 44 business development companies, the lenders marked down 81% of their software loans this year, against 40% of their loans in other sectors. About 4% of borrowers now have loans valued below 80 cents on the dollar, up from roughly 1% a year between 2023 and 2025.

The losses are real. The explanation for them is where I think the market has it wrong.

What the displacement story needs to be true

If AI were replacing software, you’d expect to see companies ripping out systems of record and running internal AI tools in their place. That isn’t happening at any scale. The best-known example points the other way. In 2024, Klarna’s CEO told investors, “We just shut down Salesforce. Within a few weeks, we will shut down Workday.” In March 2025 he clarified that Klarna had consolidated its systems and data, and wrote, “So no, we did not replace SaaS with an LLM.”

Where AI-native competitors are real, in customer support, coding and some vertical workflows, they’re mostly winning new budgets and new projects. They aren’t forcing companies to cancel tools they already have installed, like Miro or Airtable. Those products got more useful over the last two years. What changed is that the number of seats stopped growing, and only that second fact shows up in the loan marks.

What the numbers show instead

Seat-based software grows with white-collar employment. From 2010 to 2022 it compounded on two things at once: more knowledge workers every year, and more tools per worker as each team bought its own. Both have stopped. Tech headcount has been flat to down since the 2023 layoffs. Indeed’s postings for IT infrastructure, operations and support roles were about 30% below their pre-pandemic level in March, and the US economy added only 29,000 jobs in September. Procurement teams are consolidating vendors rather than adding them, and CIO budgets are flat.

A seat model with no seat growth can only grow by adding customers or raising prices, and that’s what the reported numbers show. HubSpot ended June with 306,446 customers, up 14% from a year earlier, but its average subscription revenue per customer rose only 4%, to $11,800. Customers are being added. Existing customers aren’t buying many more seats.

The vendors’ own pricing changes point the same way. Salesforce now sells Agentforce through credits at $0.10 per standard action. Zendesk charges $1.50 to $2.00 per automated resolution, HubSpot charges $0.50 per resolved conversation, and ServiceNow bills per AI action. Companies don’t move away from the pricing model that built their industry while the unit it’s priced on is still growing. They move when it stops.

AI does have a role in this, as a competitor for budget rather than a replacement for the software. With IT budgets flat and AI spending rising, something has to give, and what’s giving is seat expansion and new hiring. That’s a much weaker claim for AI than the one in the loan marks, because budget moved from seats to AI isn’t new value. It’s the same money in a different line item.

Why the lenders prefer the AI story

Whether or not anyone intends it, the AI explanation does two useful things for the managers who tell it.

It makes the software losses a sector event instead of a vintage event. The loans in the most trouble were mostly written around 2021, when buyout firms paid 10 to 15 times annual recurring revenue for horizontal software companies and borrowed six to seven turns of leverage against them, on the assumption that growth above 20% would last. When growth falls to single digits, that math fails without any help from AI. Calling it AI disruption keeps the problem inside one sector and away from the rest of the loan book.

It also turns the losses into a pitch. The old software loans are failing, the story goes, because AI is winning, and AI infrastructure is what the same managers are now raising money to lend against. Told that way, the losses on the old loans become an argument for making the new ones.

The redemption data gets a similar reading. Blue Owl has limited withdrawals from its private credit funds to 5% a quarter, and when requests came down, that was presented as the pressure easing. The other reading is that investors who wanted out learned they couldn’t get out quickly and stopped asking, and that a 5% cap met partly from a credit facility and resubmitted requests is a queue that hasn’t cleared.

Where it goes next

If this is a vintage problem rather than a software problem, it won’t stay in software. Buyouts from 2021 in healthcare services, business services and consumer companies were financed with similar leverage on similar rate assumptions, and they’re now carrying rates they were never underwritten for. Reuters’ 4% figure already covers all borrowers, not just software ones, and Benefit Street Partners’ own description of the trouble, “over-levered horizontal software and services businesses,” already includes services. The next round of BDC filings will show whether non-software borrowers are joining the list of loans marked below 80.

It also weakens the case for lending to AI. That case needs AI to be taking spending from somewhere, and software was the cleanest example to point to. If software slowed because hiring stopped, the evidence that AI is displacing spending elsewhere gets thin. In June, Torsten Slok, Apollo’s chief economist, noted that companies in healthcare, banking, energy and manufacturing were spending on AI without it showing up in their profit margins. Private credit is also the largest outside lender to the AI build-out, as the previous post laid out, so the same lenders are underwriting AI on revenue that hasn’t yet been taken from anyone.

And it undercuts the idea that software is the place to hide if AI stumbles. That rotation assumes the two are opposite sides of one story. If software’s problem is slower seat growth, it doesn’t recover when AI sells off. Both can fall at the same time for different reasons.

What would prove this wrong

The displacement story gains ground if white-collar hiring picks up and seat-based revenue still doesn’t follow. That would mean something other than headcount is holding seats down. Large companies replacing systems of record with internal AI tools, at scale and on the record, would count too.

The vintage argument weakens if non-software buyouts from 2021 come through the next two quarters of BDC filings without joining the list of loans marked below 80.

Sources

← ALL WRITING