I unpack what this month’s software selloff is pricing in, and who stands to win as the seat model breaks.
It seems that every few weeks, someone publishes a speculative take about AI that sends the internet and the markets into a spiral.
This week, it was Citrini Research’s “THE 2028 GLOBAL INTELLIGENCE CRISIS“: an imagined scenario in which AI-driven job losses cascade into a total economic collapse. The memo describes a runaway negative feedback loop: AI substitutes white-collar labor, consumer spending collapses, financial contagion spreads through credit and housing markets, and a recession takes hold, all while the tech companies building AI massively profit.
The report went viral and tapped into investor anxiety that’s been building throughout the month. When Anthropic shipped new Cowork plug-ins in late January, it helped kick off a week-long "software-mageddon” that wiped roughly $1T from software stocks. Cybersecurity names like CrowdStrike fell double digits after Anthropic previewed Claude Code Security. IBM had its worst day since October 2000 after Anthropic announced that Claude could help modernize COBOL systems.
During the 2022 software selloff, rising interest rates pushed down valuations, but no one seriously questioned whether Salesforce or ServiceNow would still be valuable businesses in 10 years. This time, the selloff is about terminal value: whether the business models that justify SaaS valuations remain valid at all.
As I see it, two questions are being conflated: what happens to software as AI matures, and what happens to the broader economy as AI displaces knowledge work. The second question has real stakes, given how much of the modern economy runs on software and the people who build and use it, but it’s a different question, and I’ll focus here largely on the first.
On the macro question, I’ll simply note one pattern that has held across many major technology transitions: when a foundational cost collapses (compute, bandwidth, distribution), the economy tends to expand through the savings rather than shrink around them. The incumbents who relied on the old economics are hit hard. But the total market grows, and the productivity gains flow into new markets, new products, and new jobs that didn’t exist before. With aging populations and deglobalization increasingly weighing on growth, AI looks less like a threat to the economy than a necessary offset to those headwinds.
AI isn’t just collapsing the cost of intelligence: it’s making it infinitely scalable, and through agents, giving it the ability to act autonomously. That’s a larger surface area than any previous tech transition - which means the scale of both the disruption and the opportunity are much larger too.
What’s motivating the software selloff
In many ways, this month’s software selloff was a dam that was waiting to burst.
Public SaaS growth has broadly decelerated since 2021. Until recently, the market treated it as a cyclical slowdown, assuming growth would eventually return. With Anthropic’s announcements, suddenly the deceleration looked less like a temporary headwind and more like an early signal that the traditional SaaS playbook itself was becoming obsolete. The Citrini report gave that story a doomsday arc. Together, they gave the market permission to reprice what the fundamentals had been signaling for three years.
The fears underlying that repricing are worth separating out, because each one sharpens a different aspect of the incumbent vs. startup question.
➡️ The seat model is breaking
When a team can do more with fewer people, companies stop expanding their seat count. With agentic AI, pricing will increasingly shift from predictable per-seat subscriptions toward usage- and outcome-based models. For incumbents, this is a double problem: it threatens revenue while demanding investment in entirely new pricing architectures.
Agentic AI makes it possible to charge for outcomes, putting software in direct competition with the vastly larger services economy for the first time. The question is whether incumbents can restructure to capture this “services as software” opportunity, or whether AI-native startups will get there first.
➡️ Margins are under pressure
Traditional SaaS could replicate itself at near-zero marginal cost. AI products can’t: every interaction requires inference compute, and the more agentic the product, the more tokens it consumes. The fastest-growing AI-native startups are running gross margins around 25%, some negative - compared to the 70-90% that defined SaaS economics. Inference costs are falling, but token demand per query is rising as tasks become more complex and reasoning-intensive, so the net effect remains uncertain.
➡️ SaaS category boundaries are collapsing
As the nature of knowledge work changes, the boundaries between roles are blurring, as are the software categories built around them.
The GTM stack is a great example. Gainsight, Zendesk, Outreach, Clari, and Gong are separate tools for adjacent functions, each requiring separate budget conversations, implementations, and integrations. AI-native companies can now own end-to-end workflows that cut across all of them.
For SaaS point solutions that built businesses around a single workflow or use case, this is existential. But for startups, collapsing silos unlocks markets that couldn’t exist inside them. The “ops” wedge is a good example: RevOps, DevOps, and SecOps exist precisely because no single system of record ties these functions together. Previously they weren’t viable software markets: two or three seats per company at most. As the walls come down, builders can create products that traverse them.
In addition, the more agents operating across those categories, the more software you’ll need to manage, orchestrate, and interpret what they’re doing. Agentic systems create new kinds of complexity, which in turn creates new categories of software demand, which startups are well-positioned to fill.
➡️ Value is moving up the stack
Agents sit between the user and the system of record, doing the work and delivering the outcome. That makes the agent layer - not the system of record below it - where customer attention and budget increasingly flow. The system of record underneath becomes commodity infrastructure: still necessary, but competing on reliability and cost rather than commanding a premium. You don’t need to rip out Salesforce to make it worth a fraction of what it is today.
The bull case for incumbents is that data gravity is real: AI is only as good as the context it can access, and companies with decades of proprietary data have a head start that’s hard to close. Our view is that owning the data isn’t the same as being positioned to exploit it, and that incumbent systems of record weren’t designed to store the kind of data an agentic workforce needs. More on this in the next point.
➡️ The nature of moats is changing
SaaS lock-in has historically been built on two kinds of friction: the cost and risk of migrating data at scale, and the user habituation that comes from teams building workflows around a specific interface. Agents weaken the second kind: users interact with the agent rather than the underlying system directly, which makes the interface less sticky. This dynamic affects SaaS broadly: any tool whose value lived primarily in its interface is vulnerable.
What replaces friction as the moat is context: the organizational decision histories and reasoning patterns that accumulate as an AI system learns how a specific company thinks and operates.
The bull case for incumbents is that they have years of proprietary data and domain experience that new entrants can’t quickly replicate. The bear case (and our view) is that incumbent systems of record were optimized for a human workforce, not an agentic one. They record outcomes, not the reasoning behind them.
For a human workforce that carried context in their heads, that was okay. But agents need a context graph: a reasoning layer that captures not just what a company has done, but how it thinks. Retrofitting that capability onto a legacy architecture is a fundamentally different engineering challenge than building one from scratch. We believe that the startups that solve the context graph problem first will be very hard to displace.
Our view: The market for software will be 10x bigger in 10 years, and AI-native startups will capture much of that growth.
This month’s selloff is the market catching up to a realization that’s been forming for a while. In many ways, it confirms the thesis we laid out in our context graphs p.o.v.
As in past tech transitions, when a foundational cost collapses, the incumbents who built their businesses on the old economics get repriced while the total opportunity expands. That’s what’s happening now.
At Foundation, we believe that the market cap of software will be 10x larger in 10 years. Many of the top 10 companies by market cap in 10 years will be AI-native startups being built today. Many incumbents will also reinvent themselves, and many will acquire startups to do so. Proofpoint’s acquisition of our portfolio company Acuvity is an early signal of how this is already playing out in security.
🪦 Who loses
The middle tier of SaaS faces the most pressure. Companies in the $100M-$1B ARR range (2018-2021 vintage SaaS, still valued on 2021 expectations) lack the capital and customer relationships of the larger players, and they also weren’t built for the new paradigm like AI-native startups. They’re too big to move fast and too small to absorb the repricing while investing in what comes next. Some will become acquisition targets; many others will simply not make it.
Incumbent systems of record face a different but related threat. Enterprises aren’t going to vibecode their own ERP systems next quarter, and these platforms aren’t disappearing. But as agents increasingly sit above them and capture the value of the work being done, the system underneath becomes foundational but not differentiated. The ones that survive will be those that evolve from passive ledgers into active reasoning platforms - but this move is hard to do quickly.
🏆 Who wins
It’s our view that AI-native startups will be the primary beneficiaries, for the reasons laid out above. They’re unburdened by legacy architecture, existing revenue bases to protect, and the incentives that lead incumbents to retrofit rather than reinvent. Unlike model providers, they can also stay model agnostic and route to the best model for each task.
This is the moment for founders with genuinely outsized ambition: not incremental improvements on existing categories, but startups going after massive problems that were previously too complex, too expensive, or too risky to tackle with software alone. Two people and a compute budget can now build what used to require 100s of engineers over years. The cost of competition has collapsed, which means small ideas are no longer sufficient: if you can do 10x what you thought, so can everyone else.
The founders who will define the next decade of enterprise software are the ones who looked at this month’s selloff and saw not a warning, but permission to think bigger.