Has the economics of software run off a cliff with AI?
Don't look down!
Most businesses have only ok scale characteristics. If you make chocolate bars, you take in cacao (commodity), do a bunch of work to it, add some margin, and ultimately ship a product. Each product has marginal cost (the cacao). The only real advantage from scaling up is that fixed costs, like the cost of processing machines and distribution, can get amortized across more sales. But fundamentally, there is marginal cost to each sale, and the margins stay fairly low.
Software pre-internet already started to break that model. Programs had to be put onto disks and shipped in boxes, so there was still some marginal cost, but mostly once you built it, it was “free” to sell it. It’s interesting to note that the costs of distribution led to a business marketplace where things like feature scale or channel ownership mattered the most.
The internet broke that. Distribution is free (almost), so now it’s super easy to build once and ship as much as you want, and the market became global because there was almost no difference in cost to serve anyone anywhere in the world. The economics shifted to attention, and first-mover advantage and network effects started to dominate.
In both cases, though, the economics meant that it was worth paying a lot to software engineers. Programs were big and hard to build, speed and skill mattered, so paying an engineer a lot of money to build once and then ship many times made sense. Eventually, with billion-user platforms, it even made sense to pay huge amounts of money for tiny, incremental features, since any new usage easily made the money back.
But now…all of that has changed. Some kinds of software are much easier to build (vibe coding) and this circle will only expand. There is an incremental cost now to shipping AI based software (inference) - each new user costs money to serve, and that money is “used up” - the electricity to serve the inference doesn’t accrue in any permanent sense, it’s just consumed. There is also an emerging trend that the best engineers are starting to consume as much in inference costs do develop software as their salaries (sometimes even multiples - it’s not hard to imagine a world where the best engineers cost 10x their salaries in inference, and are expected to do 50x or more the work a single engineer can do today).
All of that is pushing software to look more like that chocolate company above, where there isn’t as much leverage from scale, where margins compress, and where the “line worker” (average engineer) isn’t that valuable. In fact, that’s a prediction I will make, that software companies, and the software industry, will come to look much more like manufacturing: a few highly talented engineers will plan and oversee complicated “factories”, and most engineers will be low-paid “line workers”.
Anything that is transient, easy to reach with vibe coding, idiosyncratic, etc, is likely to get squeezed hard economically. Engineers and product designers who can’t make effective use of AI to build economically useful software that can scale well will also struggle. The days of adding a small feature to a huge product are likely ending.
In the engineering teams I work with, I see a real bifurcation. There are hardcore engineers who disdain AI based coding. This group is, fortunately, getting smaller, but still doesn’t make much effective use of these tools. At the other end, there are product managers and designers who aren’t technical enough to use the tools. Neither of these groups is thriving or seems likely to.
In the middle, there are people who are some mix of creative and engineer - 60/40, 40/60, either is fine. Those folks seem really well positioned to make use of these new tools. They imagine interesting things, use the tools effectively, are constantly experimenting, and have high leverage. I suspect this kind of “product engineer” is what the future of software engineering will mostly look like.
I personally lived through the arrival of the internet as a coder. It was very clear to me that, to survive, I had to understand both the technical and business implications of that shift. AI is breaking the most fundamental assumption of software: that it scales at almost zero cost. This will absolutely change the dynamics of the whole industry, probably in unpredictable ways. It’s a fatal error as an engineer or product leader to ignore this challenge right now.
This reminds me of Ben Thompson's aggregation theory. Software and the internet took the marginal cost of distribution of digital products do zero and shifted the structure os markets on its head. Before you needed to dominate supply, after you needed to dominante demand.
Now, what happens when it is the marginal cost of building software that goes to zero?
I personaly think that supply will be mutch more granular, given that the cost of building software is now near zero. The cost to maintain a software business will be low enough to justify the existence of very niche software businesses. Product and design people are well positioned to thrive, if they learn how to use these ever better AI tools to build software.
Your prediction holds only if teh software produced uses expensive inference that costs the development company. Using AI to build conventional software will not incur those costs. If vibe coding reduces skills and salaries, then this may even increase profits. If the cost of inference is borne by teh user, e.g., as Apple intends by having inference on Edge devices, again, the prediction fails. Lastly, if the cost of inference, at least for the products sold, is significantly lower, then your prediction could still fail.
Any scenario that retains the zero-cost distribution and marginal cost tending to $0 will ensure the current software model will continue.
What I do see is if vibe-coding massively increases the volume of software developed, like smartphone apps, then oversupply vs demand will drive down unit revenue, and approach classic manufacturing economics. Consumers will benefit, as software packages, once expensive, are approaching very low prices due to volumes and competition. Lotus-123 was what, $400 in the mid-1980s? Google Sheets is free. Even the MS Windows Office suite is very low-cost compared to what it was 30 years ago.
Andreesen is still correct that "Software is eating the world". The bad news for most companies is that the global market is finite and saturating, but supply will continue to increase, and hopefully start to end monopoly profits. If governments end the extortionate cut of platform app sales, or China internationalizes its OSs to compete with US companies, we may even see the decline of some of today's BigTech companies.