Article

Software Isn't Dead, and Neither Is Edtech. The Moat Just Moved.

·Connie Kang
·All posts

The short answer

The sentiment is not vibes. It's a real, recent repricing event with a coherent theoretical spine. In the first weeks of 2026, public software stocks fell ~30%, erasing every gain since ChatGPT launched, a sell-off the industry nicknamed the "SaaSpocalypse." Venture capital has simultaneously concentrated into AI to a degree never seen before (AI took 61% of global VC in 2025 per the OECD, and 81% of global VC in Q1 2026 per Crunchbase). So when an investor says "software is uninvestable," they're pattern-matching to a live market signal, not just a hunch.

But here's the thing worth understanding precisely: the theory does not say "software has no value." It says value has migrated off the code. Every rigorous framework, from 1990s information economics to Porter to Christensen to the VC power law, points to the same conclusion: when the cost of producing software falls toward zero, profit stops accruing to the software itself and concentrates in whatever remains scarce and hard to replicate. The investors saying "software is dead" and the investors saying "software is more defensible than ever" are, underneath, describing the same mechanism from opposite ends. That's the argument you can actually win.

What sparked the panic

The market event. Since the start of 2026, ETFs of public software companies dropped ~30%, with Salesforce, Adobe, Intuit, ServiceNow, and Veeva each down 25–30% in a matter of weeks (a16z, "Good News: AI Will Eat Application Software," Mar 2026). That price action is the proximate cause of the loud "SaaS is dead" chorus - sentiment follows the tape.

The origin quote. The intellectual seed was planted by Satya Nadella on the BG2 podcast (Dec 2024): "The notion that business applications exist, that's probably where they'll all collapse in the agent era, because… they are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents." When the CEO of the company that sells Dynamics says business apps will collapse, it gives the thesis institutional weight. ServiceNow's Bill McDermott echoed the "apps apocalypse" framing.

AI collapsing the cost to build. The premise under everything is that code is now cheap to produce. Cursor reportedly reached ~$1B ARR with ~300 people; Lovable became a unicorn in eight months with ~45 people. The vibe-coding stack (v0, Cursor, Replit, Lovable, Copilot) makes "what once took a team of five engineers and weeks of sprints" doable by one person in hours. If anyone can build your product in a weekend, the reasoning goes, your product isn't defensible.

The "thin wrapper" fear. Google's startup VP Darren Mowry (TechCrunch, Feb 2026) put it bluntly: "If you're really just… white-labeling that model, the industry doesn't have a lot of patience for that anymore… You've got to have deep, wide moats." The fear: a company that's a thin layer over GPT or Gemini is "one model update away from irrelevance."

Incumbents bundling the feature. The canonical cautionary tale is Chegg, a US edtech company that sold homework help and textbook-solution subscriptions to students. When ChatGPT made essentially that same service free, Chegg's revenue fell 24% then 30% year-over-year and the stock collapsed from a former ~$14B valuation to roughly a dollar a share (Chegg SEC filings). It's the cleanest example of a point solution whose entire value was absorbed by a general-purpose AI overnight.

The same pattern hit others. Jasper, an AI marketing-copywriting tool that was an early generative-AI darling, lost ground when Notion and Microsoft Office added native AI writing, removing the reason to buy a standalone tool. Stack Overflow, the Q&A site developers relied on for coding answers, saw traffic decline as GitHub Copilot answered the same questions directly inside the code editor. The lesson investors drew: if your feature can be absorbed by a platform that already has the customers, it will be.

The capital flight. The data is stark. AI captured 81% of all global venture dollars in Q1 2026 ($239B of $297B), up from 55% a year earlier; OpenAI, Anthropic, xAI, and Waymo alone took 64% of all global VC in one quarter (Crunchbase). Money is voting, and it's voting for models and infrastructure over application software.

Why it's more than hype: the theory underneath

This is where the sentiment stops being market noise and becomes an argument. Six frameworks chain together cleanly.

1. When copies cost nothing, the price collapses (information economics)

Shapiro & Varian's Information Rules (1998) established the cost structure of all software: high fixed cost to produce the first copy, near-zero marginal cost to reproduce it. In plain terms: it might cost millions to build the first version of a piece of software, but making and shipping the second, thousandth, or millionth copy costs almost nothing, there's no factory, no raw materials, just a download. ("Fixed cost" = the upfront cost to create it; "marginal cost" = the cost of producing one more unit.) Their line, information is "costly to produce but cheap to reproduce." Standard microeconomics then delivers the punchline: under competition, price falls toward marginal cost. If the marginal cost of an information good is ~zero, its price heads toward zero wherever there's real competition. Profit on the good itself gets competed away; it survives only on inputs that aren't freely reproducible.

AI's contribution is that it now also drives the fixed cost, the cost of creating the software in the first place, toward zero, not just the cost of copying it. That matters because the high cost of building was the one barrier still protecting software: even though copies were free, you at least needed a skilled, expensive team to make the original. AI erodes that too. (This extends the spirit of Jeremy Rifkin's Zero Marginal Cost Society (2014), which argued that technology keeps pushing the cost of producing more and more things toward zero, here, the "thing" is working software itself.)

2. When code gets cheap, value moves to whatever stays scarce (commoditize your complement)

First, the term. A "complement" is anything customers buy together with your product, hot dogs and hot dog buns, cars and gasoline, phones and apps. When a complement gets cheaper, people buy more of your thing (cheap buns → more hot dogs sold). So companies deliberately try to drive down the price of their complement to pump up demand for what they actually sell. That's "commoditize your complement." The AI version works the same way. Running AI requires huge amounts of computing power, which companies rent from cloud and chip giants like Amazon, Microsoft, Google, and Nvidia. For them, AI is the complement - the thing people use their servers for. So they're happy to see AI models become cheap or even free, because cheaper AI means more people building and running AI, which means more demand for the computing power they sell. They make their money on the servers underneath, not the AI on top, so commoditizing the AI layer is good for their business, even if it wipes out the startups built on that layer.

This dynamic was named by Joel Spolsky in his 2002 "Strategy Letter V." His law: "demand for a product increases when the prices of its complements decrease," so "smart companies try to commoditize their products' complements." The implication for the "software is dead" debate is the part that matters: when AI and code become the cheap commodity, the value doesn't disappear, it migrates to whatever stays scarce and hard to replicate alongside them (distribution, proprietary data, trust, embedded workflows). That's the thread the rest of this study follows.

One important caveat to this "AI goes to zero" logic. Many people argue the opposite is coming: that AI pricing will rise over the next few years, because the foundation-model labs (OpenAI, Anthropic, and others) have burned enormous amounts of capital and their investors will eventually demand a return, pushing them to raise prices rather than cut them. Both can be true at once, and it's worth separating two things. The per-unit cost of a given level of AI capability has been falling fast (the price of a fixed amount of "intelligence" keeps dropping as models get more efficient). But the labs can still grow revenue and raise headline prices by selling more advanced frontier models, premium tiers, and agentic products on top. So the commodity floor keeps dropping: yesterday's cutting-edge model becomes cheap or free, even while the newest frontier capability stays expensive. For the argument here, what matters is that baseline software-building capability becomes cheap and abundant; that's true regardless of how the labs price their premium frontier. (It's also worth noting the infrastructure giants, Amazon, Microsoft, Google, Nvidia, have every incentive to keep pushing the model layer toward commodity, which works against the labs' desire to raise prices. That tug-of-war is unresolved.)

Spolsky also pointed out a crucial gap in the "cheap kills everything" logic, and it's the foundation of the counterargument to the whole "software is dead" thesis. His example: "Even when the price is zero, the cost of switching from Microsoft Office is non-zero." In other words, a free competitor still doesn't win if leaving your current product is painful. Even if someone builds a perfect Office clone and gives it away, companies won't switch, they'd have to retrain everyone, convert every file, and rewire every workflow that depends on it. So making something free does not commoditize it if there's a real cost to switching away from the incumbent. That's the escape hatch: cheap production commoditizes the code, but it can't touch the switching costs, data, and embedded workflows that keep customers locked in. Hold onto that, it's the hinge the entire rebuttal later swings on.

3. Easy to build means easy to compete away (Porter's Five Forces)

Michael Porter is a Harvard strategy professor whose "Five Forces" framework (HBR, 1979/2008) is the standard tool for judging how profitable an industry can be. He supplies the formal version of the investor's worry. His point: "The threat of entry puts a cap on the profit potential of an industry." In plain terms, if it's easy for new competitors to jump into your market, you can't charge much, because any time you make good margins, someone new shows up and undercuts you. How easy it is to enter depends on "barriers to entry": things that make it hard for newcomers to compete, like needing huge scale, a trusted brand, access to distribution channels, or regulatory approval. The fewer the barriers, the more competitors pile in, the thinner everyone's profit. Porter adds that "substitutes" (a different product that does the same job) are most dangerous when they're nearly as good for the price and easy to switch to. AI-built software is exactly that: a cheap, low-switching-cost substitute produced under collapsing barriers to entry. Put it simply: when software is cheap and easy to build, lots of competitors flood in, and prices and profits fall for everyone. That's just what Porter's framework predicts, which is why "software has no moat" is a serious argument, not only AI hype.

4. AI drains the shallow moats and leaves the deep ones (moat theory)

A "moat" is whatever protects a company from competitors, the reason rivals can't just copy you and steal your customers. Strategists have catalogued the handful of things that actually work as moats. Two well-known lists are Morningstar's "five sources of moat" and Hamilton Helmer's "7 Powers," and they overlap heavily. Boiled down, the durable moats are:

  • Switching costs - it's painful or expensive for customers to leave you (retraining staff, migrating data, rewiring workflows).
  • Network effects - the product gets more valuable as more people use it (everyone's on it, so newcomers have to be on it too).
  • Proprietary data or resources - you own something rivals can't get: an exclusive dataset, a patent, a hard-to-copy asset.
  • Brand and trust - customers choose you because they trust you, especially when the product is hard to evaluate themselves.
  • Cost/scale advantage - you're structurally cheaper to operate than anyone else because of your size.
  • Regulation - licenses, certifications, or compliance hurdles that legally keep competitors out.

The crucial insight is that AI does not weaken all of these moats equally. Some were never very strong to begin with, like having slightly better features than a competitor, which AI now lets anyone copy in days. Others, owning exclusive data, being the system a company can't afford to switch off, holding a regulatory license, are barely touched by AI, or even strengthened by it. So AI knocks out the flimsy moats first while leaving the real ones standing.

What AI commoditizes: raw feature differentiation, model capability itself, and UI-over-a-database. What it leaves alone, or even makes stronger: data only you have, regulatory approvals and hard-won trust, the ability to reach customers (distribution), network effects, and one especially important kind of switching cost. There's a difference between switching costs that are technical (the new software is hard to set up) and ones that are organizational (your whole team's habits, training, and accumulated knowledge are wrapped up in the current system). AI can help with the technical kind, it can write code to migrate your data. It does nothing about the organizational kind. A hospital won't swap out the system it runs on just because a cheaper one exists, because the real cost isn't the software, it's retraining thousands of staff and rebuilding years of accumulated know-how baked into how they use it.

Helmer frames this as a test: an advantage is only a real moat if competitors genuinely can't or won't copy it. "Better features" fails that test now, anyone can copy them with AI. "Being the system a hospital can't afford to rip out" passes it easily.

5. Startups get squeezed from below and above (disruption theory)

Clayton Christensen (a Harvard professor) explained why big, established companies keep getting beaten by cheaper, "good enough" newcomers. The pattern: a newcomer enters at the bottom of the market with a cheaper, simpler product. The incumbent (the big established company that already dominates the market) doesn't fight for those low-margin customers, it makes more business sense to focus on its high-paying ones, so it hands the bottom of the market over without a fight. But the newcomer keeps improving and creeps upward into better customers, and by the time the incumbent notices, it has "run out of customers" to retreat to. AI-built software fits this pattern exactly: it's the cheap, good-enough newcomer eating into single-purpose software tools (the kind that do one narrow job, like scheduling or expense tracking) from the bottom up.

This is already visible in edtech: the established players are racing to bundle everything into all-in-one platforms, absorbing what used to be separate point tools. That's the same squeeze from the other direction, a standalone tool that does one job now risks being out-flanked both by a cheap AI-built version from below and by an incumbent folding that feature into its all-in-one suite from above. (What survives this two-sided squeeze is the subject of the rebuttal below.)

But Bain's 2025 Technology Report adds the decisive nuance. In Christensen's framework there are two kinds of innovation: a disruptive one helps a newcomer overturn the established players, while a sustaining one just makes the existing big players better at what they already do (a faster engine in an existing car, say, it reinforces the incumbent rather than threatening it). The catch with AI is that it can be both. For a startup it's a disruptive weapon, but the big incumbents can also simply bolt AI onto the products they already sell to millions of customers: Microsoft adding Copilot to Office, Salesforce adding its Agentforce agents. Used that way, AI isn't overturning them; it's strengthening them, because they get the same AI capability plus all the customers and data they already had. So point-solution startups get squeezed from both directions, disrupted from below by cheap AI-native entrants, and out-bundled from above by incumbents. That two-sided squeeze is the real content of "software is dead," and it's why the threat feels different from past platform shifts.

6. Why a perfectly good business still might not get funded (the VC power law)

This is the piece most people miss, and it's the actual reason investors specifically talk this way. Venture returns follow a brutal power law: a16z's analysis of Horsley Bridge data found ~6% of investments generated ~60% of all returns; Cambridge Associates found the top ~2.5% of deals drive the bulk of all industry value. Bill Gurley: "Venture capital is not even a home run business. It's a grand slam business." (A baseball metaphor: in baseball, a "home run" is already a great result, the batter scores at least one point in a single swing. A "grand slam" is the rarest, biggest version of that, a home run hit at the exact moment that scores the maximum four points at once. Gurley's point: it isn't enough for a VC to back solid, moderately successful companies (ordinary good results); the whole fund depends on hitting the one rare, enormous, once-in-a-portfolio winner that pays for everything else.)

Peter Thiel's Zero to One turns this into a rule: because returns are so skewed, VCs must back companies that can escape competition into monopoly economics. His "competition is for losers" thesis holds that competition erodes margins to zero, so the only venture-grade business is one that becomes a durable near-monopoly. Now connect it: a business that is easy to build is easy to compete with, which means compressed margins, which means it can never become the monopoly outlier the power law requires. It may be a perfectly good cash-flow business, and still be un-venture-fundable, because it can't return a fund. That is the precise mechanism behind "software is no longer investable." It's not a claim about value; it's a claim about defensibility-adjusted value at venture scale.

The case that software is doomed, at full strength

To argue against it credibly, state it at full strength:

Software's only durable scarcity was the difficulty of building it. AI is dissolving that scarcity. By information-goods economics, price follows production cost toward zero. By Porter, collapsing entry barriers cap industry profit. By commoditize-your-complement, the model labs are actively incentivized to make your layer free. By disruption theory, you're squeezed from below by weekend-built clones and from above by incumbents bundling the same feature with distribution you can't match. And by the VC power law, even if you survive as a business, you can't become the monopoly outlier that justifies a venture check. The 30% software drawdown and the 81% AI funding share aren't anomalies, they're the market correctly pricing all of the above.

That's a serious argument. It's wrong in its conclusion, but only because of one move it skips.

The rebuttal: the moat moved, it didn't disappear

The honest counter, and notably, the one the best primary sources make, is not "software is fine." It's "the moat relocated from code to the things AI can't generate," and AI is a tailwind for whoever owns those things.

Distribution becomes the binding constraint. Box CEO Aaron Levie (June 2026): "Software has a bit of its own law of conservation of mass. If you make one thing free and abundant (development), then the new problem of discoverability and market differentiation (GTM) just costs more as a result." a16z's Alex Rampell's classic formulation, "the battle between every startup and incumbent comes down to whether the startup gets distribution before the incumbent gets innovation." When building is free, distribution is the whole game.

Owning the data and being the "system of record" survives. A "system of record" is the one official place a company keeps the data it runs on, the database everything else checks against (think the patient records a hospital trusts as the truth, or the customer list a sales team works from). Once you're that, you're very hard to remove. Greylock's Jerry Chen puts it well: "The new moats are the old moats", own the system of record for something a business genuinely depends on. But not all data counts. There's a real debate over whether "data" is a moat at all, and the resolution is about which kind: data that anyone can scrape off the internet is worthless as a moat, but data that only you have and that keeps growing as customers use you is gold. Veeva (18+ years of pharmaceutical-industry data) and Bloomberg (40 years of financial data) have exactly that, datasets no competitor and no AI can simply recreate, because they were built up over decades of being the system everyone used.

The "thin wrapper" insult doesn't hold up. The most common put-down of an AI startup is "it's just a thin wrapper around ChatGPT", meaning it's a flimsy shell with someone else's AI doing all the real work, so it has no value of its own. The point here is that this critique is weaker than it sounds, because every successful software product is a "wrapper" around some commodity underneath. Andrew Chen (a16z): "Salesforce is a wrapper around relational databases. Shopify is a wrapper around payment rails." Salesforce didn't invent databases and Shopify didn't invent online payments, yet both are enormously valuable, because the value was never in the raw layer underneath. It's in everything built around it: making the product trustworthy, compliant, and actually usable for a specific job. The AI model you're "wrapping" is interchangeable and replaceable; the carefully-built system around it that makes it work for real customers is not. As a16z's Joe Schmidt puts it: "The model is fungible underneath; the system of work is not", i.e., you can swap the AI engine, but you can't easily swap the whole working system a business has come to rely on.

Industry-specific software for regulated, high-stakes fields is the safest ground of all. ("Vertical" software means software built for one specific industry, healthcare, law, education, finance, as opposed to general tools anyone uses.) The argument has three layers:

First, software encodes how an organization actually works. Hebbia's George Sivulka calls software "a stored process… a social contract", it's the agreed-upon way a group of people get their work done, written into a system. Replacing it is painful "not because the interface is hard to learn, but because the institutional knowledge stored inside is genuinely costly to reconstruct", i.e., the software holds years of an organization's accumulated rules and judgment that nobody wants to rebuild from scratch.

Second, the demanding, unglamorous parts of an industry are exactly what protect you. Sapphire Ventures: "The grit is what hides the market. It's also what makes it defensible." The compliance requirements, edge cases, approval processes, and hard-won institutional standards of a regulated field are genuinely difficult to serve well, which is precisely why generic tools and AI labs tend to stay away from them. That complexity isn't a nuisance to be dismissed; it's real, it matters to the people who live in it, and it rewards whoever takes it seriously enough to do the work properly. Their line "Models win demos. Wedges win pilots. Systems win markets" captures the progression: a flashy AI demo impresses, a narrow useful tool ("wedge") gets you a trial, but only a fully-built system that handles all the real-world mess actually wins and holds a market.

(This may sound like it contradicts the earlier disruption point, that startups win precisely by starting small and narrow at the bottom of the market. It doesn't, the two describe different stages of the same journey. Christensen explains how you get in: enter narrow and cheap, where the incumbent isn't paying attention. Sapphire explains what you must become to actually win and stay won: you can't stop at the narrow wedge, because a thin single-purpose tool is exactly what gets squeezed, by cheap AI clones from below and all-in-one suites from above, as in the edtech example earlier. The winning path is to use the narrow wedge as your way in, then build it out into the full system that's genuinely hard to displace. The wedge is the entry; the system is the moat.)

Third, more software is coming, not less. Veteran tech leader Steven Sinofsky's flat rebuttal (a16z, Feb 2026) calls the "software is dead" conclusion "Nonsense… We need vastly more software, not less," and predicts "domain experience", deep knowledge of a specific industry, "will be wildly more important than it is today," because that's the part AI can't supply on its own.

A startup can pick a business model the incumbent literally can't copy. This is what strategists call "counter-positioning": you adopt an approach that would hurt the incumbent's own business if they tried to match it, so they don't. The clearest example is pricing. Most established software is sold "per seat", a fixed fee for each user who logs in. An AI-native startup can instead charge "per outcome", only when the software actually gets a result (say, per support ticket resolved, rather than per support agent). Customers often prefer paying for results. But the incumbent can't easily follow, because switching to outcome-pricing would shrink the large, reliable per-seat revenue it already collects, it would be undercutting itself. So the startup gets a pricing model the incumbent won't match.

Pricing is just the clearest example, counter-positioning can come from any business-model choice the incumbent can't mirror without damaging itself. A few others: giving a product away for free or open-sourcing it, which an incumbent can't do without killing the license fees it lives on (the classic case is Netflix's flat-rate, no-late-fee model, which Blockbuster couldn't copy because late fees were a huge chunk of its revenue); selling directly to customers when the incumbent depends on resellers or a sales channel it can't bypass without alienating them; or building an AI-native product from scratch when the incumbent's older architecture and existing customers make a rebuild too risky. The common thread is always the same: it's not that the incumbent can't figure out how to copy you, it's that copying you would cost them more than ignoring you.

The two moats work differently, though, so it's worth being precise about who each one favors. Counter-positioning clearly favors the startup, by definition it's an advantage the incumbent can't copy without hurting itself.

Lock-in is different: it favors whoever owns the system of record, whether that's an incumbent or a startup. The dynamic is simply that once an organization has years of its operations and data built up inside a system, that system becomes genuinely hard to replace: not because anyone is trapped, but because it has become part of how the organization actually runs. Today that usually describes the incumbent. So why is it an opportunity for a startup? Because the goal is to become that system, to win the customer with counter-positioning (the model the incumbent won't match), then earn a place so deep in how they work, through real value and accumulated shared history, that there's simply no reason to leave. The startup uses counter-positioning to get in the door, and earned indispensability to stay.

Worth noting the flip side, because it's the healthy version of this: a vendor that holds onto customers only through inertia, high switching friction but no ongoing value, is exactly who's exposed as AI makes switching a little easier. Durable lock-in has to be earned by continuing to be worth it, not just engineered. That distinction is the whole difference between a partner and a captive, and it's the version that actually lasts.

And the money confirms it where defensibility is real: Harvey (legal) at an $11B valuation, Abridge (healthcare) at $5.3B with >90% clinician retention, Sierra at ~$15B. Menlo's enterprise survey found AI application spend hit $19B in 2025 and that 76% of AI use cases are now bought, not built, meaning that when a company needs AI software, three out of four times it pays a vendor for a ready-made product rather than building its own in-house. That directly undercuts one of the loudest "software is dead" fears, that AI would let every company just build its own tools and stop buying software. In practice, companies are doing the opposite: buying more. Defensible application software is raising more, not less.

What this means if you build in education

The synthesis lands squarely on the kind of company worth building.

The thing that dies is undifferentiated, non-AI-native, moat-less SaaS, UI over a database, a feature pretending to be a product, a thin wrapper with no proprietary data or switching cost. "Software is dead" is shorthand for "commodity software is dead," and it always was a weak business; AI just made the weakness undeniable and instant.

The thing that wins owns at least one scarcity AI can't manufacture: a proprietary, continuously-refreshed dataset; a system of record embedded in an institution's workflow; regulatory/trust/compliance ground that takes years and relationships to earn; or distribution into a market that's painful to reach.

Here's the part edtech founders sometimes get tangled on, and it's worth stating plainly: the moat and the mission are the same thing. In a consumer app, a moat can be extractive, a way to trap users and squeeze them. In regulated education, the sources of defensibility are the sources of value. The reason an institution can't easily leave a system is usually that the system holds years of their data and is woven into how they actually operate and improve, which is also precisely why it helps them. The reason standard-setters and regulators come to trust a product is that it genuinely makes a painful, high-stakes process less painful and more useful. The accumulated dataset that competitors and AI can't recreate is the same dataset that lets an institution see how it's really doing over time. You don't build the moat instead of serving schools; you build it by serving them so well, and embedding so deeply into something that matters, that the value compounds and becomes hard to walk away from. A defensible business and a genuinely beneficial product aren't in tension here, in this category, they're produced by the same actions.

So the honest framing of the whole "software is dead" debate, for anyone building in education, is this: commodity software is finished, and that's the right bar. The theory the skeptics are quoting says value migrates to distribution, proprietary data, embedded workflow, and trust in regulated fields. Those are exactly the places where, in education, doing right by institutions and building something durable turn out to be the same work.

Sources (highest-credibility, primary where possible)

The thesis / catalysts

The economic theory

Moat / strategy theory

Disruption + VC power law

The counterargument