A college student in 2021 paid Chegg $14.95 a month to solve a hard algebra problem. By late 2025 the same student typed the same question into a free chatbot and got a better answer in eight seconds. Chegg’s stock fell 99%. The NYSE issued a delisting notice. A $14.5 billion company became a $0.51 stub.
I keep thinking about that drop. Not because it is about education technology, but because it is the cleanest autopsy we have of what happens when a SaaS company sits in the middle and the middle goes away.
The Canary Was a Homework App
Chegg’s entire business was packaging access to something a model could now generate for free. No proprietary workflow. No regulatory moat. No operational data only Chegg held. The subscription was a toll booth on a road that got paved over.
The company tried everything. CheggMate, built on GPT-4 — which is a polite way of saying it competed with the platform using the platform’s own technology. A lawsuit against Google. A 45% workforce cut. Full-year revenue down 39% to $377 million. The “pivot” to Chegg Skilling was the diplomatic phrasing for an admission the original business was dead. [7]
A 99% decline is not panic. It is accurate pricing of zero defensibility.
Hold one question for the rest of this essay: Am I Chegg?
The $285 Billion Wake-Up Call
On February 3, 2026, Anthropic shipped Claude Cowork with thirteen MCP connectors and plugin templates across finance, HR, engineering, design, and legal. Each plugin replicated functionality entire SaaS companies sold as standalone subscriptions. Jefferies traders christened it the “SaaSpocalypse” before lunch. By close, $285 billion in software market cap had vanished. By mid-February the total exceeded $1 trillion.
Salesforce -38% year-to-date. Adobe lost $120 billion in seven weeks. ServiceNow -26%. Intuit -33%. [1] [9]
Two days later, OpenAI launched Frontier, positioned as “the semantic layer for the enterprise.” Agents that navigate existing applications and execute workflows without a human logging in. If a Frontier agent runs sales motion end to end, per-seat licensing loses its justification. Uber, State Farm, Intuit, and Thermo Fisher signed on at launch. [3]
A Fortune 50 company disclosed plans to cut Salesforce and ServiceNow spend by 60% by year-end. Not a recession. A re-architecture. [4]
Look at what funds it. Five hyperscalers will spend $660-690 billion on AI infrastructure in 2026, nearly doubling 2025’s $380 billion. [2]
$690 billion of infrastructure cannot stay neutral. You need downstream token consumption to justify the spend. The cheapest way to manufacture that demand is to absorb the layer between your models and the end user. That layer is middleware. It is what the last decade of SaaS got built around.
And every CTO I have spoken to in the past two months says some version of the same sentence:
“If we wait six months, can’t Claude just do it?”
The question puts a concrete test on the roadmap: what will a customer still need from this product when the platform can perform the task? The answer is increasingly that the platform can do it.
There is a reasonable objection here. The selloff is panic, not structure. Tech stocks crash and recover. The 2022 SaaS bloodbath reversed in eighteen months. SaaS spend is still projected to grow from $318 billion to $576 billion by 2029.
That limits what the selloff can prove. My thesis is narrower: SaaS does not die. The middle dies. Systems of record survive because they hold the data. Compliance-heavy platforms survive because liability cannot be outsourced to a probabilistic model. What gets vaporized is the layer between them — the workflow tools, the integration platforms, the thousands of $10M-$100M ARR companies whose entire job was making the seams less painful. The platforms are now making those seams disappear.
We Have Seen This Movie Before
Software is not the first middleware to learn this lesson.
In 2000, the United States had 124,000 travel agents. By 2012, 65,000. A 47% collapse over twelve years. The internet did not end them in a quarter. It eroded their commissions, then their volume, then their relevance. [6]
The detail that haunts me is the false recovery. In the late 1990s, travel volume hit record highs, and agent employment briefly rose with it. The chart looked healthy. The margins were dying underneath. Anyone reading the employment number in 1998 thought the industry was fine. Anyone reading commissions knew the funeral had already been booked.
SaaS net revenue retention is stalling while gross retention sits near 90%. Customers are not leaving. They are not expanding. The dollars that used to fund seat growth now fund AI tooling — the SaaS budget paying for its own displacement. [10] False recovery, beat for beat.
The travel agents who survived moved into niches the platforms structurally could not serve. Virtuoso, the luxury network, grew 211% during the collapse. Cruise specialists kept booking phone-first because 75% of cruise bookings stayed offline — too relationship-dependent and too margin-thin for Expedia to bother. The generalists — the people whose job was typing itineraries into a booking system — were eliminated.
The timeline skepticism is fair. Enterprise software has decade-long replacement cycles. None of that unwinds overnight.
It does not have to. The middleware collapse does not require migration. It requires CIOs to not renew. That decision happens every quarter, in discretionary spend, with no migration cost.
The Middleware Trap and the Four Ways Out
The trap, said plainly: you sit between a platform and an end user, aggregating features or packaging a workflow. The platform ships your features natively. You compete on switching costs and inertia. Inertia is a depreciating asset.
Bain, Forrester, and a Fortune-50 procurement deck converge on the same vulnerability profile: horizontal point-solutions, generic AI wrappers, products whose interface was a UI over a workflow that can be reconstructed from APIs and documents. [5] AI wrapper startups face a projected 85-92% failure rate inside five years through “Sherlocking” — foundation models absorbing wrapper features in a single product update. [7]
I want to be honest about the obvious challenge here. Frameworks with neat exits have a poor predictive track record. The skeptic’s version goes like this: in many categories, an 80% good-enough general agent eats the niche, your remaining 20% is not worth a separate subscription, and any four-position taxonomy is post-hoc storytelling.
That lands. 80% good-enough will eat a lot of categories. The surviving niches are the ones where the gap between 80% and 99% is regulatory liability, financial accuracy, safety consequences, or domain depth a generalist cannot legally or practically replace. That is not a long list. It is, however, a real one.
Four positions show up repeatedly in companies not bleeding right now:
- Proprietary Context Owners — value from operational data accumulated over years, which platforms cannot replicate from public training data.
- Infrastructure That Agents Call — APIs and rails AI agents must invoke to act in the physical or regulated world.
- Deep Workflow Owners — products so embedded in a professional process that ripping them out costs more than the AI savings.
- Trust and Verification Layers — human accountability and certification AI cannot self-issue.
Treat that list as directional, not deterministic. The boundaries are fluid. The underlying logic — platforms absorb anything their economics let them absorb — is not directional. It is gravity.
A last objection: critics call micro-SaaS a lifestyle category, and AI makes those products easier to build and easier to clone.
That misframes it. Micro-SaaS is not a counter-strategy to platform consolidation. It is the economic complement created by it. When platforms eat the middle, they expose a structural blind spot: markets too small for platform-scale acquisition, too specialized for generic agents, too relationship-dependent for self-serve. The defensibility is not features. It is operational intimacy.
Small Enough to Matter
Look at who was quietly profitable while the rest of the industry was bleeding.
Plausible Analytics serves 16,000 paying customers. It grew from $2.1M ARR in 2023 to $3.1M in 2024. Four people. Zero ad spend. Zero investors. No board. It competes with Google Analytics, which is free, dominant, and backed by the largest advertising company on earth. On price and distribution alone, that would seem an impossible competition. [8]
It works for one reason. Google will never ship a privacy-first analytics product, because privacy-first analytics contradicts the model that funds Google. The moat is not the code — the code is open source. The moat is positional. The platform’s own incentives prevent it from competing.
Three deliberate choices made the position possible: privacy by architecture, baked in at the data layer (not a feature toggle Google could ship next quarter); open source with paid hosting, which converts the code itself into trust; radical simplicity — one dashboard, not hundreds of reports — which is a deliberate retreat from the surface area platforms compete on.
Vertical and micro-SaaS players show 35-60% higher retention than horizontals in the same period the horizontals are losing market cap. [8] Plausible is the pattern, not the fluke.
The economics here echo the Fixed Pie Fallacy: when execution costs collapse, new markets open below the old viability line. Niches uneconomical at $5M minimum size become viable at $500K. Thousands of them.
What makes those niches defensible is the kind of context Fine-Tune Your Organization keeps returning to — atomic, proprietary, accumulated from real customers in real workflows. A generalist model can summarize the regulations. It cannot live inside the operation.
The Platform Absorption Diagnostic
To assess your own exposure, trace what the platform could replace and what customers would still need from you.
Step 1 — Stack Dependency Mapping. Draw your stack from infrastructure up to end user. For each layer, ask three questions. Could that provider ship my core feature natively in 12 months? Do they have a distribution advantage that makes my switching costs irrelevant? Is my layer a natural extension of their product surface? Three yeses puts you in the absorption zone.
Step 2 — Three-Layer Context Test. Classify each feature’s value source. Structural context (public knowledge, best practices) is Google-able and trains itself into the next model release. Operational context (live workflow data) is defensible only as long as you control the integration. Proprietary context (years of accumulated domain knowledge from serving specific customers) is the only durable moat. Most founders confuse familiarity with proprietary depth — knowing that restaurants need table management is structural, knowing how this restaurant’s Tuesday lunch rush actually works after five years of POS data is not.
Step 3 — Four Positions Score. Rate yourself 0-10 against the four. You need a 7+ in at least one. A 4-5 across all four is the trap, not the safety. Claiming weak versions of all four is worse than committing strongly to one.
Step 4 — The Decision. Three paths. The hardest part is choosing C when C is right.
- A. Double down. 7+ on a position, genuine proprietary context, manageable absorption risk. Deepen the moat. Hire for domain depth, not generic engineering. [3]
- B. Pivot to the edge. Position exposed, but you have domain expertise for a niche too small for platform economics. Rebuild as micro-SaaS with direct human contact and operational embedding. Useful heuristic: TAM under $50M is below platform interest. [4]
- C. Exit or consolidate. Context is structural. Position is middleware. Absorption timeline under 18 months. Sell, merge, or wind down before the platform ships your feature for free. Sunk cost is not a strategy.
Am I Chegg? If your product is the friction between a model and something a model can already do, the answer is closer to yes than your last board deck suggests. If the platform’s own economics prevent it from entering your market, the answer is closer to Plausible.
The platform ate the middle. That part is settled. The interesting work left — the only work in this corner of software that survives the next five years — is on the edges. Where the giants will not bother. Where domain depth still beats raw capability. Where a four-person team can outlive a $14 billion company that thought packaging access was a moat.
Small enough to matter is not a consolation prize. In this cycle, it is the position.
References
[1] The $1 Trillion Software Carnage: How AI Agents Broke the SaaS Model. MarketMinute / FinancialContent, 2026. https://markets.financialcontent.com/stocks/article/marketminute-2026-2-24-the-1-trillion-software-carnage-how-ai-agents-broke-the-saas-model
[2] AI Capex 2026: The $690B Infrastructure Sprint. Futurum Group, 2026. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
[3] OpenAI Frontier Reshapes Enterprise Software. Fortune, February 5, 2026. https://fortune.com/2026/02/05/openai-frontier-ai-agent-platform-enterprises-challenges-saas-salesforce-workday/
[4] Fortune 50 procurement disclosure cited in MarketMinute, 2026 (see [1]).
[5] SaaS As We Know It Is Dead: How To Survive The SaaS-pocalypse. Forrester, 2026. https://www.forrester.com/blogs/saas-as-we-know-it-is-dead-how-to-survive-the-saas-pocalypse/
[6] Alderson, M. Travel Agents Took 10 Years to Collapse. 2025. https://martinalderson.com/posts/travel-agents-developers/
[7] Vertical SaaS Survival Guide: Competing Against AI Giants. Beancount, February 2026. https://beancount.io/blog/2026/02/02/vertical-saas-survival-guide-competing-against-ai-giants
[8] Plausible Analytics public revenue and customer disclosures, 2023-2024. Plausible.io. Cross-referenced with Forrester vertical SaaS market sizing (see [5]) and Bain SaaS analysis (see [10]).
[9] Anthropic Claude Cowork plugin launch coverage, February 3, 2026 (consolidated reporting; see [1]).
[10] Bain & Company. Why SaaS Stocks Have Dropped — And What It Signals For Software’s Next Chapter. 2026. https://www.bain.com/insights/why-saas-stocks-have-dropped-and-what-it-signals-for-softwares-next-chapter/