SaaS is dead?
If you spend too much time on LinkedIn or X, you’ve probably seen one of the more annoying, trending narratives of the ChatGPT era: “SaaS is dead.”
And you, as an innocent, non-Internet-addict reading indie Substacks, may wonder, “how is this possible, if SaaS has ruled the 21st century?”
Let’s back up before continuing our regularly scheduled programming of hating on social media.
SaaS, or Software as a Service, is software licensed to other businesses via subscription. (Technically speaking, you could describe Netflix or Spotify as “SaaS”, but no one does. Because of Aggregation Theory, consumer tech companies end up being larger but fewer in number than their SaaS counterparts. Here’s a lengthier, more detailed breakdown of nuances between B2B and B2C.)
For the rest of this essay, I will use “software” and “SaaS” somewhat interchangeably, for reasons that will become clear by the end.
Software companies have unique economics compared to those who make and sell physical goods, because they have fundamentally different cost structures. For the widget maker, every new widget they make has an input, or variable, cost. It costs them more to make 20 widgets than 10 widgets, because of the inputs required to make those 10 additional widgets.
For a software business, they “make” the software once, and whether they sell it to 10 customers or 20 customers, their costs are basically the same. In other words, there is a (near) zero marginal cost to software.1
One of the key features of SaaS is that the software is all hosted and maintained by the provider, and distributed through the internet, often accessible through a browser or desktop/mobile app. So a way to think of SaaS is that the customer rents software, rather than owns it. As the internet proliferated, distributing software became trivial. No need for setting up a datacenter in your basement, no need for armies of IT consultants – just swipe a credit card and open a Chrome browser.
SaaS became a winning business model. The unit economics of a paying customer are insane, with gross margins in the 80% neighborhood. Add to that the fact that it’s “sticky” – customers rarely churn once they start using the product – and it’s no wonder why these companies made an entire generation of people rich.
Was it really this easy? No, of course not. You had to build a compelling product that lots of companies would pay to use, and it had to be sticky enough such that it would be immensely painful to replace. Doing this is hard, and requires hiring loads of really smart people who expect really smart person salaries. Since top talent is expensive, it ends up being expensive to actually build and market great software products.
Most successful SaaS companies had to raise venture funding and operate at a loss for years before the money-printing machine was ready to go brrr.
All that said, other than illegal things that prey on the darkest weaknesses of human physiology, SaaS was possibly the best business model ever.
___________
So why is SaaS dead?
Because AI, of course.
Most pseudo-pundit-thought-leader-influencers who infiltrate the algos make arguments that ultimately come down to one of three variants:
AI as input: AI tools make it much easier to replicate existing products. What previously took five engineers six months to build, might now only take two engineers six weeks. Large Language Models (LLMs) are not universally a 10x timesaver as some PPTLIs suggest, but it’s true that some projects might see this sort of efficiency improvement. (Or in the case of a non-engineer like myself, being able to code a functioning application over a weekend, the efficiency gains are asymptotic – it would have been impossible without the LLM.)
AI as output: AI-based products are going to kill their SaaS 1.0 equivalents because they’re way better.
BONUS: “I’m just copying and pasting what I think I saw from someone who had a lot of likes and follows, which they said on a podcast I believe, I didn’t actually listen to the podcast, but I read the captions for 90 seconds on the snippet that popped up on X – wonder why my entire X feed is just horizontal TikTok videos about AI, but, hey better ride the wave, right?”
I will ignore this last one because it injures my soul, but (1) and (2) are strong arguments.
LLMs have absolutely lowered the time it takes to build software. It’s hard to say if the gains disproportionately favor more experienced engineers or junior ones. On one hand, having domain expertise makes for a really rich experience with LLMs. On the other, they help anyone become an 80/20 expert on pretty much any subject matter.
Rather, the more useful framework is high-agency versus low-agency engineers. It really seems like people who put in the time and effort to maximize their productivity with LLMs are the ones reaping the most benefit. Fortunately, it’s high-agency people who typically try to start companies. So yes, I fully embrace the efficiency gains argument.
More interesting than the efficiency gains, though, is that LLMs make for a bunch of user-facing product opportunities. What I mean by this is that there will be all kinds of new software products that have AI-based features.
To use a very mundane example: suppose you’re the head of accounting and one of your responsibilities is presenting monthly results to the CFO. It’s a stressful, time-intensive process where you mull over details, squinting at each row of the financial statements, looking for trends compared to last month or quarter or year. But now, you have an AI helper that’s directly embedded into your accounting system, that identifies those trends and anticipates questions the CFO will ask in the presentation.2
NetSuite, Sage, Oracle and every other legacy provider will launch various “AI-powered” features into their existing products. But they won’t introduce anything that cannibalizes their respective 80% gross margin golden geese – they will roll out new features that make their power users’ lives better, just like the example above. It’s easy to see this being successful, and they’ll be able to charge higher prices for their software, because their existing customers are locked in for life.
But there’s another fledgling class of accounting software, which is centered around AI capabilities. Imagine instead of employing a team of accountants, the controller or CFO simply trains an AI agent to ingest all sorts of unstructured invoices, receipts, bank statements, Slack messages and emails, and it produces pristine financial statements, stored in a tidy interface, with the same (or, let’s be honest, way better) “AI helper” capabilities as the legacy providers. If you’re a new company trying to minimize cash burn, wouldn’t it be a no-brainer to use this version of accounting software?
To conclude, I’ll offer my gut analysis on all this. These are weakly held positions – for now anyways, given how early in the cycle we are – but here goes:
Even if foundational model development slows down – ie, we are stuck with GPT-4o or Claude 3.5 Sonnet level models for the next five years – there are tons of useful SaaS 2.0 companies to be built.
These companies will differ from SaaS 1.0 in a couple crucial ways:
2a: AI as input: they’ll build faster and more cheaply, and thus have smaller teams. They will still require VC funding, though, because customer acquisition will cost money. (Yes, there are lots of intriguing sales and marketing AI agents, but I expect they make humans more efficient, not replace them.)
2b: AI as output: the products will be AI-native, using LLMs as primitives, not as bolt-on features.
Established companies will not suddenly change their ways – I think growth rates for legacy SaaS will taper, but profitability and retention will stay high because the vast majority of economic activity in five years…will be conducted by businesses that already exist today.
New companies building AI products are going to use each other’s B2B software, creating a virtuous cycle of adoption (which is exactly what happened with the SaaS 1.0 generation).
Distribution and differentiation will matter even more than before. This follows logically from the industry having a lower barrier to entry. Distribution is still massively important; startups will have to work with and/or around existing ecosystem lock-in. Differentiation can come in a lot of ways; user experience and pricing are two obvious ones.
If I’m right about all or most of these points, it will mean SaaS is not dead – it’s evolving. The landscape of B2B software will look different: new companies, new products, new pricing models…which is to say, healthy industry growth and innovation.
As long as businesses exist, they’ll need software. Better to focus on building the future of productivity, not doomscrolling about its demise.
An astute observer will notice this dynamic also applies to music, film and really any other product predicated on “bits, not atoms.” This is true, and it’s why many investment banks, as an example, lump in Technology, Media, and Telecommunications (TMT) as a single coverage domain.
Accounting software is an example of a “horizontal” SaaS product, in that it can be applied to all kinds of businesses (ie, customers). Many smart people think “vertical” software that focuses on specific industries, or verticals, will actually be the big winner of an AI-based SaaS 2.0. I think that’s probably right, but don’t want to get too in the weeds in this post. Speaking of weeds, I once used pet extermination software as an illustrative example in an old essay, and it confused some people, who thought I was saying you could use software to kill bugs, which is true if you’re talking about bugs in software, but very much not true, as far as I know, about killing actual physical, sentient insects. Really I was just talking about a niche tool that helped the local exterminator schedule their clients and track their jobs without the use of paper. Net net, I’ve decided to stick to examples in areas I know best, like boring back office accounting.