The SaaSpocalypse was not a market event. It was the moment Wall Street caught up to a structural truth operators already knew: software built for humans to operate has a shrinking future.
In 2020, I was writing TensorFlow tutorials. Step-by-step walkthroughs. Code blocks with explanations of every parameter. There was an assumption buried in every one of them, so obvious that I never thought to state it: a human would read this, learn it, and do the work.
That assumption quietly expired. The market just took longer to notice.
In February 2026, it noticed all at once. $285 billion erased from SaaS valuations in 48 hours. The software ETF down more than 20% from its peak. Thomson Reuters posting its largest single-day decline on record. The financial press needed a name for it and landed on "SaaSpocalypse."
The name suggests a catastrophe. I would describe it differently: it was a correction of category. The market stopped pricing SaaS companies as software businesses and started pricing them as what many of them actually are — labor-adjacent businesses whose unit of value was about to be automated.
| Signal | Figure |
|---|---|
| Erased from SaaS valuations in 48 hours | $285B |
| Net new ARR across cloud software, year over year | −29% |
| Enterprises actively cutting SaaS vendors | 82% |
| Enterprise SaaS spend per-seat by 2030 | -40% |
A Seat Was Never the Product. It Was a Proxy.
Strip away the dashboards and the integrations, and the per-seat model rests on one premise: each seat represents a person doing work that the software makes faster.
That premise held for fifteen years because there was no alternative to the person. The software could assist the human; it could not replace the human. Pricing per seat was pricing per unit of irreplaceable labor.
Agentic AI broke the premise, not the software. The software still works. The dashboards still render. What changed is that the work behind the seat — reading, triaging, drafting, reconciling, following up — can now be done by an agent that does not need a seat, a login, or a lunch break.
When the work moves, the pricing unit has to move with it. Seats price access. Agents demand that you price outcomes.
This is why February was not about earnings. Atlassian, Salesforce, and Intuit were not punished for bad quarters — some companies that beat earnings fell anyway. They were repriced on a forward question: in a world where five agents replace fifty seats, what is this revenue actually worth? The honest answer was: less than the multiple implied.
The Cracks Were Visible Long Before the Crash
When I was publishing those tutorials in 2020 and 2021, the average enterprise ran over 110 SaaS applications. At the time, that number was read as a sign of a healthy market. I read it differently even then — I wrote about data maturity precisely because I kept seeing the same pathology: tools that did not talk to each other, reports that contradicted each other, decisions made on dashboards nobody fully trusted.
The sprawl was never strength. It was deferred cost. Every additional tool added a seam, and every seam leaked visibility, speed, or accountability.
AI did not create that problem. It removed the tolerance for it. By early 2025, the average stack had already shrunk from 112 apps to 106. Net new ARR across the cloud software universe had fallen 29% year over year. CFOs had stopped asking "who needs access?" and started asking "what changed because we paid for this?"
Those are different questions. The second one is much harder for a seat-based business to answer.
Two Kinds of Software Companies Will Exist. The Middle Will Not.
From inside Vizio AI and Vizio Ventures, the bifurcation that analysts describe in the abstract is already an operational fact. The dividing line is not who uses AI — everyone uses AI now. The line is where intelligence sits in the architecture.
| Dimension | Legacy SaaS | AI-Native |
|---|---|---|
| Where AI lives | Bolted onto the product | Is the operating layer |
| Pricing unit | Seat | Outcome |
| Core promise | Your people work faster | The work gets done |
| Revenue as agents improve | Shrinks — fewer seats needed | Grows — more outcomes delivered |
| Moat | Eroding — features replicate in days | Compounding — intelligence deepens with use |
The proof points are no longer theoretical. Zendesk charges per resolved ticket — nothing for failed attempts. Monday.com replaced a 100-person SDR team with agents and cut response times from 24 hours to 3 minutes. Intercom bills only when its agent closes an issue. These companies have transferred performance risk from the customer to themselves. That is what confidence in your own system looks like, expressed as a pricing model.
The companies stuck in the middle — legacy architecture, AI features, seat pricing — carry the costs of both worlds and the advantages of neither. The market has started pricing that position accordingly.
The Question I Ask First Is Different Now
The distance between my 2020 writing and my work today is not technical. The technical foundation — how models learn, how data flows, how systems hold together — is still load-bearing. What changed is the order of questions.
In 2020, I started with: what can this model do?
Today, I start with: where does this organization lose visibility, speed, judgment, or accountability — and what would have to be true for that to stop?
Capability-first thinking produces demos. Outcome-first thinking produces operating leverage. The SaaSpocalypse is, at bottom, the market forcing that same reordering onto an entire industry at once.
The crash did not change what good software is. It made it impossible to keep hiding what bad software was: access without accountability.
Three Questions Before Your Next Renewal
1. Where does intelligence enter this workflow? If the answer is "it doesn't — people read the output and decide," you are paying for a viewing window, and viewing windows are exactly what agents replace first.
2. Does this seat create accountability or just access? Access is a commodity now. Accountability — knowing what happened, who owns the next step, and when it moves — is what still commands a price.
3. Does the value compound? If the tool is exactly as useful in month 36 as in month 1, it is static. Static tools get rebuilt in a weekend. Compounding systems do not.
The organizations that will remember 2026 as a clarifying year rather than a painful one are the ones using the repricing to audit their own operating model — not just their vendor list. The vendor list is downstream. The operating model is the thing.
AI only matters when it changes how an organization sees, decides, and acts. Everything else — the features, the copilots, the chat windows — is still just software, and software just got repriced.