Master list of "will AI disrupt software?" bear and bull arguments

Probing the biggest question in software today.

These are organized in roughly descending order (strongest arguments on top, weaker on bottom, based on my personal read). I will update this periodically as I come across new ones.

For more on this topic see An anxious moment for software and Incumbents can use Cursor, too.

🐻 Bear case (AI as a disruptor):

1. Foundation model providers entering the SaaS application space: Hyperscalers and model providers (OpenAI, Anthropic, Google, Amazon, etc) expand upwards to compete into the application layer. This has happened most notably with coding assistants. Is this coming next for project management, marketing automation, DevOps, etc? This is a far more terrifying risk than #2 below, in my opinion.

2. Lower barriers to entry in software production: AI-assisted coding allows new entrants, especially low-end competitors, to enter the market more easily. Even if this doesn’t necessarily lead to incumbent churn, it could/will drive down price points.

3. Enterprise consolidation & vendor fatigue: The proliferation of AI solutions might frustrate CIOs enough for them to decide to just consolidate spend with the big infrastructure/application mega-vendors (Microsoft, Oracle, SAP) instead of screwing around with a bajillion point vendors, particularly in a world where AI output has serious compliance and security concerns. “Nobody ever got fired for buying IBM” comes back with a vengeance.

4. Obsolescence of the entire system-of-record and workflow software categories: This is speculative, but a future paradigm where agents autonomously accomplish tasks across unstructured data might entirely obviate whole categories of software that are premised on normal relational databases. Do we need a UI if there’s no human doing the clicks? Do we need a CRM with standardized fields if an agent can just “figure it out” on its own? This might not affect revenue in the short run but could kill terminal value for many public SaaS businesses.

5. AI sucks up attention from customers and capital markets: Even though “software” and “AI” really aren’t distinct categories, customer budgets could rotate toward projects requiring inference and away from “vanilla SaaS”. Investor attention on AI often comes at the expense of software (we’re already seeing this in the "long semis, short software” trade) which means increased cost of capital for SaaS companies.

6. AI inference causing “Red Queen” effects on gross margins: Incumbents and new AI-first entrants all incorporate inference to compete in the AI era, making these capabilities commodities that are required to compete but don’t provide differentiation. This would push value capture away from SaaS (which would see gross margins drop from 70-80%+ to 50-60%) and towards OpenAI/Anthropic/Google.

7. Seat-based pricing fragility: We’ll see pressure on per-seat models, and incumbents will need to transition into some sort of consumption paradigm. This feels like something that could be transitioned/managed over time, though.

8. Vibe coding: Small customers could vibe code their own solutions rather than pay a third party vendor. Hard to see many enterprises going this route, but could be a problem in the very low end of SMB.

🐂 Bull case (AI as a sustaining innovation for incumbents):

1. Distribution and brand matter more than ever: If software production costs continue to drop, then what’s scarce (and therefore valuable) is distribution, brand, reputation, and go-to-market. “Nobody ever got fired for buying IBM” becomes “Nobody ever got fired for buying Datadog”. Also, large players with big marketplace ecosystems (e.g. Salesforce, Atlassian) can distribute AI functionality while commoditizing AI point players.

2. Applications are the natural delivery vectors for AI: How is AI inference consumed? Via web applications of course, perhaps you’ve heard of ChatGPT? The SaaS application layer is the “last mile” for pushing AI into enterprises which means it becomes more important than ever. Functionally there may be no difference between an “AI-first entrant” and an incumbent that develops AI modules in its preexisting platform (put another way, AI is not a platform shift).

3. Developer leverage favors scale: If each software developer is about to become 2-3x more valuable, then value accrues to the businesses with the most developers. A company like Salesforce with 10,000 developers (or whatever it is) gains an enormous amount of leverage to push their roadmap and outpace new entrants.

4. Switching costs and inertia are very real: The entire premise of wildly successful businesses like Constellation Software is that nobody ever rips out their CRM or HRIS just to get incrementally better functionality or UX.

5. Enterprises pay SaaS vendors for a lot more than just software: They pay the vendor to maintain it, to improve it, to support it, to secure it, and to make sure it’s compliant. These services are cost-effectively provided by scaled software vendors who can spread these costs across many customers. Any enterprises who want to take software development in-house need to take on all of these additional burdens.

6. So far, no evidence of AI disintermediation: Rather than facing higher churn and slowing growth, many scaled public software named have recently started re-accelerating. There’s no evidence of AI disruption in the printed numbers (yet).

7. Software is a dynamic business: Software companies are some of the most adaptable businesses imaginable — if the customers demand a slightly different product, they can build it; if a lead gen route dries up, they can pivot to a different route to market; if seat-based pricing doesn’t make sense, they can flip to consumption pricing. Thus we should not assume many permanent structural disadvantages in this industry, there’s always the opportunity to reinvent yourself (see: Microsoft).

8. AI apps have a margin disadvantage: Big incumbents have established franchises providing 80%+ gross margins and plenty of free cash flow. In contrast, many new AI entrants are running at negative gross margins just to win share and grab a toehold. This makes them highly dependent on external capital and very exposed to investor sentiment.

9. Scaled software businesses are the perfect consumers of AI: Nearly the entire cost structure of software businesses sits in knowledge-worker labor in OpEx, which means they have the most meat on the bone for AI-initiated operating leverage and net margin improvement. Also, these businesses are filled with the kind of people who know how to procure and implement AI.

10. Ability to fast-follow: Historically the way startups beat incumbents was via raw speed. Now, incumbents can more easily copy and scale the strongest features coming out of the new startup entrants, negating the speed advantage.

Tentative takeaways

My personal hunches on where this goes:

• Market-wide, SaaS gross margins will be headed downwards (due to inference costs in COGS) but net margins and cash flow margins will go up (because there is so much meat on the bone to compress labor in OpEx via AI). Gradually this will make the P&L’s of these companies look more like industrial companies, and it will be easier to value them on earnings rather than revenue.

• On balance, AI advances in software development will not provide a decisive advantage to either incumbents or entrants. The main outcome of all this innovation will be better software and greater consumer surplus for everyone, but moats for specific companies will come from other factors.

• The most dangerous thing that can happen to any software market is OpenAI deciding to enter it directly.

• I’m not sure that an “AI-first new entrant” is even a real thing. This whole discussion reminds me a little bit of the beginning of the Covid-19 pandemic, where every healthtech company bragged about being the first to have telemedicine capabilities… it turned out that all this meant is that every vendor implemented Twilio by the end of April 2020 and no differentiation was earned by anyone. Similarly, now everyone is using AI inference, and I don’t really see what makes a company founded in 2024 any more “AI-enabled” than one founded in 2004, everything else held equal. (I may be proven dreadfully wrong on this point.)

• With that being said, it does seem like we’re moving into a world where the purpose of software is to accomplish work, not to just be a workbench for humans to accomplish their work. I think every incumbent cloud software company has the opportunity to transition into this newer sort of “AI” business model, but not all will execute well enough to win. For the ones that do, I foresee enormous upside on their current market caps.

• We need to get away from talking about software businesses as if the only thing that matters for their moats is the cost of building software. There are so many other factors determining winners vs losers than whoever ships code the fastest. The fact that Lovable or Replit can vibe-code an app in 5 minutes doesn’t say much about what will happen to Atlassian (as one example).

• The companies with the most options to take advantage of the AI megatrend are probably the Mag7 businesses (minus Tesla which is sort of its own thing). If somebody were to ask me how to naively invest into the current moment without having to do too much analytical work, I’d probably suggest that they either just buy the hyperscalers (MSFT, GOOG, AMZN) or go long the cap-weighted Nasdaq-100 (QQQ).

• The entire landscape will probably look radically different in 6 months and half of what I’ve written above won’t make sense anymore.