Artificial Intelligence (AI) in the form of large language models (LLMs) has made quite incredible strides in coding capabilities in the past year. As a result, stocks of software companies are coming under pressure (the clumsily-named “SaaSpocalypse”). It seems that each time a new frontier model drops, the stock market reacts negatively. The underlying logic appears to be that “anyone can build their own software to do X” and that therefore existing companies selling software doing X are going to go out of business. I don’t think that checks out. AI is definitely causing massive changes to the software industry, but the Saaspocalypse looks overhyped.
As I have said many times, software development is about solving problems with computers. Useful software results from a long process of understanding user problems and needs and building a solution that works. Successful software does this in a way that is ergonomic and efficient for the user. Writing code is the smallest part of this process. Our existing software products are not the result of a few lines of idea in a prompt – they are the result of sometimes decades of iterative improvement and development.
Not something easily reproduced at home or at work by an amateur with an AI.
That is not saying that AI-produced code is not useful. It is very useful, and the technology is very powerful. But it will not obliterate the software industry. It makes creating code faster, and thus it should make developing software cheaper – which will have major consequences. Question is just what consequences.
Traditional Buy vs Build?
Traditionally, the fundamental question driving the acquisition of software products is buy vs build. Do you as a business (or consumer) buy an existing solution, or do you build one yourself? I would argue that we have seen a long shift from building to buying as the computer industry has matured. Maybe we will see a swing back towards building again, as AI makes code cheaper. Maybe.
Building your own software can be motivated in a few ways:
- There is nothing available to buy that solves the problem at hand. This is a common start to new software companies – beginning as an internal team and eventually spinning out as a startup.
- It can be cheaper to build than to buy. If you have the expertise in-house to build something, and available commercial solutions are priced higher than the cost of your own engineers, then building wins. Note that the full cost of the software is much more than just coding the first version – you have to include maintenance, support, training, integrations with other products, etc.
- It can be a strategic choice, to keep control over a key business function and not be beholden to a particular vendor or limited to what is going on the market. Bespoke software can be a key differentiator and competitive advantage.
So how does AI affect this?
Let’s look at some examples.
Specialized Vertical Software
Earlier in 2026, I heard presentation from Vitec, a company specializing in software for very specific verticals. Niche markets like “dentist practices in Sweden” or “real estate agents in Norway”. None of the markets are particularly large, but they all have peculiar needs and regulations and processes that Vitec have captured in their software.
The value of Vitec’s products is that they encode the particular needs of these niches. Their software has been tested by other people. There is a team of experts in the domain driving the software forward. As a buyer, you get something that works. Building something on your own might seem reasonable for a narrow niche, but if the niche is deep and complicated enough, off-the-shelf makes sense. Even if it is just sharing the effort over a small number of customers.
The Vitec view on AI-driven coding that they presented was that the productivity increase applies to all programmers. Thus, even if a potential user who decides to roll their own software get 10x faster thanks to AI, Vitec’s own programmers see the same increase. Thus, the distance between build and buy is basically maintained. Maybe not a flawless argument, but I can buy the essence.
BTW, I stumbled across a decent Reddit thread saying roughly the same thing.
Productivity Applications
I have a very hard time accepting that massive existing applications like Microsoft Office or the Adobe Creative tools can be replaced with build-your-own software. The sheer amount of functionality coded into these applications with 30+ years of history is a huge barrier. The applications have a rich ecosystem of plugins and extensions that make them even harder to replace. Going beyond just functionality, a lot of effort has gone into optimizing performance and ensuring the robustness of these often business-critical tools.
Clear case of “buy”. If nothing else, there are often open-source applications available in case price is the problem. But building from scratch just does not make sense.
On the other hand, are these kinds of applications really in the “software-as-a-service” category?
Algorithmically Complex Software
Another class of hard-to-reproduce software is software that embodies complex algorithms, software like compilers, databases, and EDA tools. These programs encode decades of experience and painstaking work to ensure they work correctly even in the strangest corner cases. For example, compiler warnings have been tweaked and tuned based on user feedback and cumulative insights as to what is helpful. Correctness and stability of output are very valuable.
Building this kind of software from scratch just for yourself seems quite foolish. Reinventing the wheel and all that.
This is also a good example of where “coding” shrinks in importance. Writing the code is not the hard part, it is coming up with the algorithms in the first place is. Software developers can spend entire careers in a narrow field and never run out of new ideas or room for improvement.
Another important question when it comes to algorithms is pure and simple basic correctness. You can ask an AI to cook up an algorithm, but how do you know it is actually correct? Just throwing test cases at it is not enough, analysis is needed. A good example of this is Joachim Strömbergson asking an AI to cook up a block cipher. It did it. But it is not something you should use since there has been no deep analysis of the actual security of the result.
1980s Analogy
One interesting analogy for how vibe coding moves the buy vs build balance is to look back. When I started computing in the 1980s, it was not uncommon for commercial software for home computers to be developed by a single person (or at most a handful). A cutting-edge computer game for a home computer of the era could be built by a good programmer and a graphics artist. Many “professional developers” were essentially users who had a good idea and the dedication and skills to execute on it. The distance between commercial software and home-brew software was quite small technically speaker.
However, the software business was still doing great. If you saw a game reviewed in a magazine or running on a demo machine in a computer shop you could in theory sit down and try to replicate it on your own. But why would you when you could buy the ready-made product and enjoy it immediately? It might just be eight kilobytes of code and data, but it was eight kilobytes that had been selected and designed by someone who knew what they were doing. It encoded knowledge, taste, and information.
Coding Your own Little Solution
I see AI making it easier, faster, and cheaper for technically-savvy people to code their own custom applications. But this is not likely to compete with existing software but rather cover cases where “buy” was never an option.
For example, a friend of mine needed to facilitate small-scale trading of privately-held shares. Buying a commercial solution was unaffordable as nobody had a price model that would work – even the cheapest solutions would result in massive per-trade costs that would make the whole exercise pointless. Implementing known trading algorithms in a small custom program that could be run when needed is not all that hard, and with AI it becomes really easy.
Honestly, I think this is place where a lot of the ideas about the Saaspocalypse comes from. There is a depressing number of services out there that solve small well-understood problems with well-known solutions but where building your own software made no sense in the past. With AI reducing the cost for roll-you-own, such products will be under pressure.
But as a product manager I feel that those products have themselves to blame. If you have something that is so easily replicated, you did not really create much value in the first place. Products must have some kind of unique selling point to make sense in the market.
Killing the Small Consultant
The above discussion does point at one type of software business that AI will most likely kill. That is the classic small consultancy business that solves small problems for small customers. Things like creating a web page for a small business. Or coding up the stock trading script discussed above. There are assignments that have low information content, no real differentiating features, and a high proportion of coding. An AI agent can do the job well enough as it is really just rehashing known patterns – and AI is pretty good at doing that.
It is sad to see this go away, as this is often where budding computer programmers and entrepreneurs got started in the business.
Creating new Competitors?
Another way that AI will impact software companies is that increased coding productivity could make it easier to create competitors to existing software. After all, if you can use an existing software package as a guide, building a new implementation is a lot easier than coming up with the same from scratch. Borrowing the distilled knowledge and learning that went into the original software, but without “borrowing” any of the code.
For example, Cloudflare created EmDash as a “successor” to WordPress. Essentially, a new implementation in a different language of the same functionality. They also improved the plug-in architecture and security. It also happens to be technically tied to CloudFlare infrastructure so that you have to be a CloudFlare user to actually deploy a site using EmDash. Sneaky.
Is this a threat to WordPress? Good question. It is not a complete replacement, as it cannot use existing WordPress plugins (change of language and architecture). It is more a framework for programmers rather than a complete user-facing solution. It has technical merit and seems to make sense if you are building a new site from scratch – or are willing to invest in a migration.
The unanswered question is what the effect of programming AI had on the creation of EmDash. Clearly, CloudFlare could have built this at any time in the past. Did AI make it significantly cheaper to do so? CloudFlare says it did, but it is hard to measure since they did not also do the same entirely manually.
Summary
AI is definitely a shock to the software industry and ecosystem. Changing the economics of code creation and increasing programmer productivity (at least for some types of tasks) will cause an upheaval.
Will we see a surge in in-house software replacing commercial software? Maybe, but I would argue only for cases where the value of the commercial software being considered is low. The needle might move back towards build rather than buy for a lot of simple things – but the cost of long-term maintenance must always be kept in mind.
In many fields, AI will benefit bought and built software equally, and likely just maintain the current equilibrium. Software that encapsulates a lot of knowledge and learning will be far less likely to be replaced by home-grown software. After all, what you buy is not the code, but the value of the accumulated knowledge built into existing software.
I do think it will get easier to set up competing businesses, as the time needed to get to functional parity with the current state of existing software will be lower than it used to be. This effect seems very real, but there is still a lot of value in the incumbents. Some things just take time to learn, even if the coding part is sped up 10 times.
Yes, that’s a great encapsulation of how this AI era will play out.