On Tuesday, roughly $300 billion of market value vanished across SaaS, data, and software-heavy firms. The IGV Software Index was already about 30% below its late-September highs. Almost on cue, the usual debate flared again SaaS versus AI, this model versus that model as if markets needed yet another binary storyline.
At a high level, two forces are driving the repricing. The first is AI, but not in the simplistic “AI replaces SaaS” sense. Recent disclosures and real-world performance of frontier models have made it harder to assume software advantages automatically persist for decades. Markets are shortening how long they expect excess returns to last, pulling competitive fade forward and that hits SaaS valuations right where they are most sensitive: terminal value. The second force is macro, acting as an amplifier. With Kevin Warsh stepping in as Fed chair and openly discussing multi-trillion balance-sheet contraction, tolerance for long-duration, liquidity-dependent assets drops quickly. Put those together and you get rotation: long-duration software de-rates first, while capital gravitates toward nearer cash flows, stronger balance sheets, and clearer pricing power.
For now, let us focus on the first factor: what AI is doing to SaaS economics and valuation.
What we’re really watching is a repricing of uncertainty—or more precisely, a compression of time. Scenarios investors once expected to unfold over ten or fifteen years are being dragged into a five-year window. When time collapses like that, duration gets punished.
Part of the confusion comes from treating “SaaS” as a single sector. Competitive dynamics differ sharply along lines that now matter a lot: horizontal versus vertical software; systems of record versus systems of engagement; mission-critical workflows versus convenience layers; seat-based pricing versus usage or outcome-based pricing. Thanks to AI, those distinctions have moved from academic to existential.
This week’s frontier model releases helped clarify the real ceiling of current AI—what these systems can do reliably, and where they still fail. One telling example: Anthropic tested whether its latest model could automate an entry-level research or engineering role. The headline answer was “no,” though a minority argued it might become feasible with enough near-term scaffolding.
The productivity gains are real and in some cases eye-popping, anywhere from 30% to 700%—but they mostly come from a simple shift in how the work gets done. The AI takes the first pass, and a human steps in to review and correct. That’s speed and leverage, not full automation. These models are incredibly helpful, but they’re still uneven in reliability. They can act too forcefully, gloss over consent, and today they solve only a fraction of real-world root-cause analysis problems end to end. So what we’re seeing is best understood as a supply-side productivity shock, not AGI or mass job displacement. Not yet !! Humans are still very much in the loop.
The important caveat, though, is that progress here isn’t linear. Frontier models evolve in bursts, not baby steps. It’s entirely plausible that by the end of the year we see sharp, discontinuous improvements in capability. That doesn’t invalidate the current limits—but it does mean the boundary keeps moving, and fast.
It also helps to rethink what “capacity” means. In industrial markets, capacity is physical: plants, machines, throughput. In AI-era software, capacity is the ability to produce, replicate, and deploy functionality at near-zero marginal cost—and that rewires the economics of advantage.
We’ve been here before. Software has already lived through major transitions—on-prem licenses to subscriptions, boxed software to SaaS, CapEx to OpEx. Along the way, incumbents stumbled or disappeared: Siebel Systems, Lotus, Symantec. Software didn’t end then either. Business models changed and value pools moved. AI opens another chapter of reconfiguration: faster, cheaper ways to build, copy, and deliver capability.
Right now there are more narratives than hard signals. Data lags, stories race ahead, and investors fall back on imperfect but useful proxies. A few worth watching: engineering output per dollar, how quickly competitors reach feature parity, what happens to sales capacity as headcount meets automation, how often customers talk about building internally—and always the intensity of discounting. That last one, in particular, is the market’s blunt but honest clearing mechanism.
All of this feeds back into how software is built, how business models are delivered, and how companies buy. The common denominator is uncertainty. Not “Is software dead?” but “For how long can excess returns persist?”
Strip away the emotion and the answer becomes clearer. This wasn’t a demand shock. It was a duration shock. Markets aren’t denying software’s future; they’re shortening it. The debate only looks theatrical on the surface. Underneath, it’s about time—and the cost of getting it wrong.
Why are SaaS highly valued ?
Before getting lost in today’s arguments, it helps to ask a simpler question: why was SaaS valued so highly to begin with?
Structurally, most SaaS businesses are still built around contracted, recurring revenue. Customers sign multi-year agreements, contracts renew automatically, and usage often expands over time rather than resetting to zero each year. Net revenue retention still matters. Churn is still visible early. From a cash-flow perspective, SaaS still behaves more like an annuity than a traditional product business. That fundamental engine hasn’t disappeared.
Growth is still largely embedded. Companies don’t win a customer once and move on; they land, then expand—through more seats, higher usage, add-on modules, or broader deployments across the organization. Mature cohorts are often still the most profitable. The idea that value compounds inside an installed base remains true, especially for software genuinely embedded in workflows.
Operating leverage also hasn’t vanished. At scale, gross margins are still high, infrastructure costs still flatten, and R&D and G&A still grow more slowly than revenue. Many SaaS businesses can still plausibly reach strong free-cash-flow margins once growth investment normalizes. In that sense, the long-term economic model that investors underwrote hasn’t suddenly broken.
What has changed is not the mechanics, but confidence around duration and certainty. Investors are no longer willing to assume every SaaS company enjoys excess returns for fifteen or twenty years, or that expansion and pricing power are automatic. AI has made substitution faster, reduced switching friction in some layers, and raised real questions about how long advantages last. The model still works but the margin for error is smaller, fade can be faster, and the bar for defensibility is higher.
So SaaS hasn’t stopped being SaaS. Recurring revenue, embedded growth, and operating leverage remain. What’s different is that markets are more selective and more skeptical about how long those benefits persist, and which companies truly deserve to be valued as long-duration assets.
So why are valuations compressing now?
The short answer is that the SaaS model always had a hair-trigger built into it. The very assumptions that justified premium multiples also made valuations extremely sensitive to small changes. If net revenue retention softens, the growth runway suddenly shortens. If discount rates rise, long-duration cash flows lose value quickly. If terminal margins come down even slightly, terminal value the part doing most of the heavy lifting collapses. When those inputs move, prices gap.
That fragility was always embedded in the math. What changed is that AI flipped the switch. It introduced credible uncertainty about durability, how fast features can be copied, and how long pricing power lasts. Once markets questioned those assumptions, the whole structure reacted.
From a capital-cycle perspective, it looks like a cycle turning slowly at first, then all at once. There’s a concept sometimes called the Seneca effect: systems grow gradually by layering on complexity, but they can unwind far faster once the reversal starts. Growth compounds calmly; decline cascades. Feedback loops kick in, confidence cracks, and the system doesn’t drift back to average—it falls off a cliff.
That’s essentially what you’re seeing in some SaaS valuations: not the end of the model, but a sudden repricing of how tightly everything depended on assumptions that no longer feel as solid.
How AI is changing SaaS ?
The first big shift is that the boundary between vendors and customers is blurring. Before AI, buy-versus-build typically favored buying SaaS: building internally was slow, expensive, and risky, so vendors accumulated real advantage. After AI, that logic weakens. Internal tools, copilots, and agents make it easier for teams to build “good enough” solutions themselves. The new competitor isn’t another startup—it’s the customer. From a capital-cycle point of view, that matters because entry barriers fall much faster than exit barriers, which is classic late-cycle fragility.
The second change is how stress shows up. SaaS doesn’t clear like industrial markets—no warehouses fill with unsold inventory. Pressure shows up economically: prices push down, discounting spreads, new ARR slows, net revenue retention drifts lower, and customer acquisition gets more expensive for each incremental dollar of growth. AI speeds all of this up by lowering switching costs and making substitutes easier to adopt. The adjustment isn’t one dramatic event; it’s a grind of margin pressure and growth disappointment.
Third, the gap between entry and exit barriers widens. On the way in, barriers are collapsing: AI tools, APIs, open source, and faster shipping make it cheaper and quicker than ever to launch something competitive. On the way out, barriers remain high: stock-based compensation, large sales organizations, organizational optics, and deferred revenue all make it hard to pull back quickly. Easy to enter, hard to exit is another textbook late-cycle pattern.
Then comes what you might call the illusion phase: returns versus the cost of capital. The 2020–21 world assumed high ROIC, long-lived advantages, and slow fade. AI changes those economics in ways that don’t immediately show up in reported margins. Incremental returns on new investment begin to slip because replication and substitution risks rise. You can still look “healthy” on headline ROIC while new dollars earn less. One of the oldest capital-cycle rules applies: when reported ROIC looks fine but incremental ROIC is falling, the cycle has already turned. Markets usually realize late—and when they do, the repricing is sharp.
Another underappreciated piece is the hidden boom. Funding slowed after 2022, so it felt like de-risking. But effective capacity exploded anyway. Engineers became more productive. Build costs fell thanks to open source and large models. Sales capacity expanded via automation. Supply grew after the funding bust, not before it. That’s why valuations didn’t magically stabilize just because multiples looked cheaper—capacity was still coming online.
At the same time, the supply pipeline compressed. Old-school SaaS had a two- or three-year lag between an idea and a real competitive response. In the AI era, parity can arrive in months, and internal tools can replicate a large share of functionality surprisingly fast. When response times collapse like that, competitive advantages fade faster than markets expect. Investors were still pricing a tidy, linear decline. Reality rarely is. Check growth and churn under the microscope.
AI also reshapes the cost curve. It flattens it. Marginal competitors become viable, incumbents lose pacing advantage, and differentiation shifts away from pure feature depth toward harder-to-copy things like distribution, data gravity, compliance, and deep workflow embedding. The capital-cycle lesson is blunt: when marginal producers get stronger, industry returns fall—even if demand grows.
Finally, management behavior tells you where you are. You see sales teams being hired even as net new ARR slows. You hear “AI expands TAM” used to defend valuations. You watch platform acquisitions justified on narrative rather than incremental returns. Stock-based compensation papers over weaker economics. These aren’t signs of a fresh upswing; they’re coping mechanisms near the top of a cycle.
Put together, this is less about SaaS breaking and more about the system adjusting to a new reality. AI didn’t flip the table overnight, but it changed enough assumptions fast enough that the old playbook no longer works as it did.
How did the valuation reset actually happen?
It starts with time. Markets marked down how long they think SaaS companies can earn excess returns. Instead of fifteen to twenty years of advantage, investors shifted toward something like five to ten—with more doubt attached. That alone can crush EV-to-revenue multiples even if next year’s revenue numbers barely move.
Next comes fade. AI shortens the half-life of advantage. Features get copied faster, “good enough” alternatives show up quickly, and internal build becomes a live option. Returns on capital converge toward the cost of capital sooner. When that happens, terminal value, the piece doing most of the work in SaaS valuations—shrinks fast.
At the same time, the risk premium creeps up. Not only because rates may be higher, but because business models feel less predictable. Seat-based pricing looks shakier in a world of agents. Retention becomes more sensitive to build-versus-buy choices. Competitive intensity is harder to forecast. When the future turns foggy, long-duration assets take the hit first.
There’s also a timing problem. Costs move forward—more inference and R&D spend today—while benefits move backward. Monetization is uncertain, and customers increasingly expect AI features to be bundled or free. In DCF terms, near-term cash flows weaken while upside is pushed out, dragging NPV lower.
Then consensus breaks. Analysts stop agreeing on the story, forecast dispersion widens, liquidity thins, and multiples compress. Prices overshoot. That isn’t “panic”; it’s what happens when the map itself is in question.
Seen this way, it’s the capital cycle reasserting itself, accelerated by AI, and expressed through shorter duration, faster fade, and higher perceived risk—not through the disappearance of software demand.
To understand where we are we need to do a proper classifying. This dislocation isn’t uniform across “SaaS.” It’s segmented, and the segmentation that matters boils down to two axes.
The first is horizontal versus vertical. Horizontal software sells broadly across industries—CRM, HR, analytics, productivity, developer tools. Vertical software is deeply embedded in a specific industry—hospital systems, core banking, insurance administration, regulated ERP. Horizontal scale gives breadth; vertical focus creates entanglement.
The second axis is system of record versus system of engagement. Systems of record are authoritative and often regulated. If they fail, the business is at real risk, and switching costs are concrete. Systems of engagement sit closer to the surface—interfaces, workflows, collaboration, user experience. These can be swapped out piece by piece, and that’s where AI bites first.
Put those axes together and the valuation reset stops looking random. Horizontal systems of engagement—design tools, project management, collaboration, marketing workflow—are UI-forward and convenience-heavy. Switching costs are usually lower and pricing is often seat-based. In an AI world where feature replication is fast and internal “good enough” tools are viable, these names are most exposed. If multiples compress here, it’s not obviously hysteria; it may be the math catching up.
Contrast that with vertical, regulated, audited systems of record. These behave differently. Substitution is slower, switching costs are real, and pricing discipline tends to hold. Competitive advantages last longer because the software isn’t just a tool; it’s part of the operating fabric. These businesses rarely look exciting, but they sit behind moats AI doesn’t easily tunnel through.
A third bucket is easy to muddle but shouldn’t be: AI-native platforms and infrastructure tooling. That is a different cycle with a different definition of capacity. Lumping AI infra in with legacy SaaS misses the point.
If that segmentation is roughly right, a practical investment framework follows.
First, classify the business. System of engagement or system of record? Engagement layers deserve shorter duration assumptions and faster fade. Systems of record—especially in regulated settings—deserve more patience on duration.
Second, examine the moat. Moats built on compliance, auditability, data gravity, embedded workflow control, and ecosystem distribution are more durable. Moats built mainly on feature depth are fragile. Features are now among the easiest things to copy.
Third, stress-test the pricing anchor. Seat-based pricing is riskier in a world of agents and automation. Usage- or outcome-based pricing can preserve value more effectively as “humans per workflow” falls.
Fourth, reverse-engineer what the market assumes. Lock near-term consensus cash flows, assume returns fade to the cost of capital by some year N, and solve for N implied by today’s price. Then ask if that duration is defensible. If the market still assumes long duration in AI-exposed engagement layers, there may be more downside. If it assumes unrealistically short duration for deeply embedded or regulated systems, that’s where opportunity hides.
Fifth, don’t ignore second-order effects. Valuation collapses don’t stay on spreadsheets. They drive talent outflow, slower decision-making, internal caution, and longer customer sales cycles. At that point, the capital cycle becomes endogenous: the repricing itself accelerates CAP erosion.
And finally, don’t wait for earnings misses. Watch the leading indicators: net revenue retention (especially expansion), CAC payback stretching, heavier discounting at renewals, more “we’re building internally” commentary, faster time-to-parity from competitors, and shifts in seat counts and pricing models. These are late-cycle signals in software form.
If there’s one clean takeaway—more hypothesis than proclamation—it’s this: AI looks like a supply shock that quietly expands capacity, shortens response times, and speeds up competitive fade. When capital cycles turn, markets do what they always do: shorten duration assumptions, steepen fade rates, raise uncertainty premia, and compress multiples hardest where terminal value once did most of the work.
SaaS mechanics still work: recurring revenue, embedded growth, operating leverage. What broke is certainty about duration.
So what? Let’s back-test it.
We took that “software meltdown” basket and treated it like a real-world lab. No hero calls, no price targets—just one hard question: if you ran the framework before looking at the charts, would it have flagged the fragile names as fragile and the durable names as durable?
The unlock is classification. Before touching returns, force each company into two buckets: horizontal or vertical, and system of engagement or system of record. Once you do that, the drawdowns start to look almost predictable.
Start with horizontal systems of engagement—the design tools, project management, collaboration, marketing workflow layers. These are UI-forward, convenience-heavy products with relatively lower switching costs and seat-based pricing. In a world where feature replication is faster and internal tools are suddenly viable, these names were always going to be the most exposed. That’s exactly what shows up: sharp, correlated drawdowns, often 50% to 75% off the highs. Multiples compress far faster than near-term fundamentals. That’s not random. That’s duration being repriced—classic CAP compression.
Now compare that to horizontal systems of record—the platforms that run core workflows: workflow engines, financial systems, ERP-like gravity wells. These also got hit, but they weren’t vaporized. You see more like 30% to 45% drawdowns and generally better relative strength than the engagement-layer names. The framework predicts that: AI can pressure duration, but it doesn’t erase the competitive advantage period for something deeply embedded, audited, and operationally existential. The market took a haircut to CAP. It didn’t take CAP to zero.
Then there’s vertical engagement software, which sits in the middle. These businesses benefit from domain nuance and industry embedding, but often still sell workflow and UI differentiation rather than regulatory authority or true data gravity. So the selloff is real, but less synchronized. There’s more dispersion and compression is slower. In capital-cycle terms, that’s delayed clearing rather than instant collapse.
The most defensible bucket—vertical systems of record—behaved close to how you’d expect. Regulated, audited, mission-critical systems where failure isn’t an option. Even here prices dropped, but damage was smaller, relative strength better, and stabilization often came earlier. Classification alone explains a large share of the cross-sectional carnage, which is already a meaningful validation.
Next is magnitude. One core claim is that the big move wasn’t primarily “earnings collapsing.” It was duration being marked down. The tape supports that. Plenty of companies had growing ARR, high NRR, and only modest tweaks to forward revenue, yet stocks still fell 40% to 70%. That only makes sense if the market is saying: you can grow, but you won’t earn excess returns for as long as we thought. Terminal duration down, fade rate up. These companies weren’t punished because they were about to run out of cash; they were punished because their future got compressed.
You can see the incremental ROIC trap as well. Many of the hardest-hit names still reported healthy headline margins and superficially decent unit economics. Underneath, expansion slowed, discounting crept in, and competitors hit parity faster. That’s what it looks like when incremental returns on new investment are falling even while the installed base still looks good. Markets often sniff that out before accounting makes it obvious. That’s capital-cycle behavior.
Moats show the same pattern. The worst performers cluster in products that are UI-heavy, workflow-light, feature-differentiated, and not protected by compliance or audit burden—bullseye for AI replication pressure. Regulated and deeply embedded systems held up better. The blunt truth remains: feature depth isn’t a moat anymore, at least not by itself.
Pricing is the underrated killer. Seat-based engagement software fell fastest and deepest. Businesses tied to usage, outcomes, or embedded process control were materially more resilient. In an automation world, the pricing anchor determines whether you can still capture value when “humans per workflow” declines and agents do more of the work. The market clocked that quickly.
Even without formal MICAP math on every name, the shape is visible. The biggest losers were the ones valued as if they had very long competitive advantage periods. After the drawdown, many are still implicitly priced for five to ten years of excess returns, which may still be generous for horizontal engagement software. The uncomfortable implication is that some compression was necessary, and in certain corners may not be finished. On the flip side, some vertical systems of record now imply very short CAPs, which is where selective opportunity can appear. Not “SaaS is cheap” as a blanket statement, but “some duration got over-punished.”
You also see the clearing mechanism exactly as expected. Software doesn’t clear like oil or steel. There’s no inventory pile and no plant shutdown to reset the market overnight. It clears economically: NRR softens, expansion slows, discounting rises, customers talk about internal builds, sales cycles lengthen. Those signals show up before earnings finally break, and markets move ahead of the reported damage. Late-cycle behavior, again.
Then the second-order effects kick in, and that’s where it gets nasty. Once valuations compress, talent drifts toward AI-native firms, innovation cadence slows inside incumbents, risk aversion rises, and customers hesitate. Feedback loops become self-reinforcing. The cycle turns endogenous: the repricing accelerates CAP erosion.
Net net, the framework passes this back-test unusually cleanly. It explains who gets hit hardest, why “earnings are fine” doesn’t save you, why AI accelerates compression rather than rescuing multiples, why dispersion beats rebound, and where duration genuinely survives. What it doesn’t do—and shouldn’t pretend to—is call the bottom or hand you timing.
The uncomfortable but straightforward takeaway is this: the selloff wasn’t random, it wasn’t just vibes, and it wasn’t anti-software. It was the capital cycle doing what it always does when supply expands in ways people don’t fully see, duration assumptions prove wrong, and competitive advantage half-lives shorten. The value of the framework is that it doesn’t just explain the move after the fact—applied ahead of time, it keeps you away from the worst landmines.