The economics of the software business has been really good for developers since the 1980s at least. By developers I mean anyone in the business of making software, not just individual engineers – developers in the Apple App Store sense. Before the 1980s software was mostly a consulting style of business (think IBM.)
Selling Information Doesn’t Cost Much
From the first time software sold to large markets, it was almost the perfect business. You developed it once and produced copies for almost nothing. In some ways not so different from the movie business after the inntroduction of VCRs, except that depending on what it did, software could command a large price compared to a video. The marginal price of distribution was very low.
You had marketing costs (as with most any product,) and costs for actual goods shipped. Boxing up diskettes wasn’t free but the next best thing to it. Instead of packaging up coffee-makers and their manuals selling for $24.95 each and paying for shipping (coffee-makers weigh something,) you were shipping out essentially empty boxes with manuals and diskettes and selling them for $24.95. Or $39.95. Or $59.95. Or more. Oh, maybe the $79.95 software required four disks.
That was phase one. Distribution over the internet was phase two: Even cheaper distribution, patches and upgrades available on demand and a world-wide market for not much more initial development cost. Under this model the customer’s hardware still does the computation. They provide the platform for your software and you have to adapt to what people out there have and can afford.
Selling Computation
Phase three: Software as a Service (SaaS) takes distribution to the next level. Run the software not on the customer’s system but rented infrastructure in the cloud, scaled to demand. You remove distribution headaches and disk-copying. It still may have larger cost per customer but makes up for it by enabling recurring revenue models (subscriptions basically.)
While companies took back responsibility for managing computer hardware, they made more money and the fundamental model was still one where the revenue – if you get popular – grows way faster than costs. Here the business sells (or re-sells) their computation resources in addition to the software product. Most of the computation going on in these products (think boring office suite types of apps) isn’t incredibly high. Really compute heavy work like games usually leans heavily on the customer’s system.
There is also the native phone and tablet app market that’s a sort of hybrid of phase two and three from the developer point of view. Developers can choose how to monetize their product. Sometimes they just charge an annual fee or one-time purchase. In reality though, the phone app market is more of an extractive model. The platform takes a big cut for no risk. They rent out the distribution channel.
But the real money is in games, including games of chance. Lots of games employ the in-app purchase model which I think of as similar to a subscription, but a variable amount only limited by the developer’s addiction engineering skills and the victim’s bank account. The ultimate software business model is probably the online casino or sports gambling app. You’re combining the guaranteed profits of a casino without the cost of a physical location or workers – a casino where anyone can visit at any time.
Selling Intelligence is not Cheap
Phase four is of course AI based SaaS. At the moment at least, running state of the art AI models on consumer hardware isn’t feasible, so SaaS is the only practical way to run an AI software business. For the first time since the 1980s developers have major ongoing expenses for their products. If the expenses go up (maybe the platform underprices their service? I bet they do,) your whole business model may collapse. This represents a huge shift away from the traditional way to plan a software business. It’s even worse than it sounds. Lots of investment in data centers going on now won’t build anything too lasting. All the hardware (servers and GPUs) will quickly become eclipsed by newer varieties. The ones running the AI services will need to charge for repeated rounds of building out to keep up. So rates will likely only rise from here.
Though inference costs per TOPS (a unit of computation, “Trillion Operation Per Second”) continue to drop a lot, as the models aren’t yet spectacularly capable for many tasks good products will demand much more than they currently use. Or at the least the best products will likely demand the most inference. As more becomes available, more gets used. It’s a bit complicated by different measures of inference: Do we mean sheer throughput for many tasks, or really fast throughput on a few tasks? There are different kinds of hardware out there and some new type could reshape the market if it enables a new kind of model. All this is to say, I don’t expect costs to drop a lot for AI app developers.
One approach is to offload the inference costs to consumers: You buy the app but supply your own Gemini (or whatever) account. I don’t think this will prove too popular since it exposes users to the (still subsidized) costs of using AI. Probably it’s the better approach for the environment though. I mean, one way or the other someone ultimately has to pay for the inference but companies will try to hide that cost to gain users for as long as they can.
Even if inference gets way better on consumer devices the quality available at a data center will be way higher. The question is, are there types of models that do a good-enough job on consumer hardware in the next three years? I’m skeptical of that outcome.
There’s an interesting second order effect on the business too: The cost to develop software drops because of AI coding agents. No question there’s already a slight effect and it will grow as the agents become more powerful. The development cost reduction doesn’t happen in a vacuum – everyone’s costs drop together. The net effect might not be too noticeable.
Here we’re talking about traditional software not primarily run by LLMs. If you think instead of a product mostly relying on AI for it’s capabilities and its code developed mostly by AI then you change the business model a lot: Cheap initial development with rather high marginal costs.
Driving Prices to Zero
But then there’s another effect: If a lot of vibe-coded products come out that are truly “good enough” all that’s left is for UI experts and product engineers to develop really good versions for niche use cases. Otherwise, software supply saturates a finite market.
Again, this concerns mostly traditional software and SaaS. I don’t know what valuable software will look like in a few years. Maybe only powerful AI driven software will be considered interesting. Who will own it? And will ownership of the software matter, or the platforms running it?