There’s a seductive story going around about AI and software development. It goes like this: AI has reduced the cost to build software by (let’s say) 75% or more, which means niche software that was previously uneconomical will now flourish. We’ll see a Cambrian explosion of products serving small, underserved markets. The long tail of niche software is finally here.
I think this story is mostly wrong, at least for commercial software, and the reason comes down to a basic decomposition that people aren’t doing carefully enough.
When we say “the cost of software has fallen,” we typically mean the cost of building it - engineering time, primarily. Let’s imagine that AI has genuinely reduced this by 75%. That’s nominally a dramatic figure, and it feels transformative.
The problem is that building software is not the same as running a software business. A typical SaaS company’s costs break down into something like:
Engineering and product development: 30-40% of total costs
Sales and marketing: 40-50% for growth-stage companies
Operations (infrastructure, support, compliance): 15-20%
General overhead and coordination: 5-10%
If AI cuts your build costs by 75%, but build costs were only 35% of your total, you’ve achieved a 26% reduction in overall costs. That’s meaningful, but it’s not the revolution that “75% cheaper” implies. The difference between a $600k/year cost base and a $444k/year cost base matters, but it’s not going to suddenly make a $200k/year market viable.
Here’s where it gets worse. The implicit assumption in the “long tail” thesis is that AI reduces distribution costs too. You can generate marketing content, automate outreach, and run campaigns more efficiently. This is true at the level of production - you absolutely can produce more marketing materials more cheaply.
The trouble is that marketing and sales costs aren’t really about production. They’re about winning a competition for finite attention. There are only so many eyeballs, only so many hours in the day, only so much mental budget that potential customers have for evaluating new tools. When AI makes it cheaper for everyone to produce content and run campaigns, you haven’t reduced the cost of winning. You’ve escalated an arms race.
Think of it this way: if everyone gets a megaphone, the cost of being loud drops to zero, but the cost of being heard might actually increase. You now need to out-shout everyone else who also has a megaphone. The input costs fell but the competitive equilibrium stays roughly the same, or perhaps gets worse as the noise floor rises.
This isn’t speculation. Customer acquisition costs in SaaS have been rising for years, even as marketing tools have become more sophisticated and accessible. More automation, more AI-generated content, more programmatic advertising - and yet it costs more than ever to acquire a customer, not less. The same dynamic will likely apply to AI-powered marketing: everyone gets the same weapons, nobody gains a lasting advantage, and we end up back where we started except with more spam in everyone’s inbox.
This dynamic gets more extreme over time, not less. Build costs have a floor somewhere near zero - or at least approaching just the compute costs for running the AI. Sales and marketing costs have no such floor, because they’re fundamentally about a scarce resource (attention) that doesn’t expand just because AI gets better.
As build costs compress toward zero and S&M costs stay stubbornly high, software businesses start to look more like consumer packaged goods or fashion brands. The product itself becomes almost commodity, and all the value and differentiation shifts to distribution, brand, and customer relationships. The moat stops being technical and becomes entirely about go-to-market. We’re already seeing this in some SaaS categories where the products are near-identical and the winners are simply whoever has the best sales machine.
Taken to its logical conclusion, this suggests a bifurcated market. On one side, a small number of giant players who can afford the S&M arms race and achieve the scale needed to amortise those costs. On the other, a long tail of software that simply isn’t sold at all - internal tools, personal utilities, community projects, bespoke automation. The middle ground of small-to-medium software businesses serving niche markets gets squeezed from both directions: not enough scale to compete on distribution, not niche enough to avoid competition entirely.
This framing suggests that AI’s cost reduction is genuine only in contexts where you’re not competing for attention:
Internal tooling, where you’re building for your own organisation and distribution costs are zero by definition
Bespoke automation, where the audience is yourself or a small known group
Software for communities you’re already embedded in, where discovery happens through existing relationships
Situations where “good enough” beats “better than competitors,” because you’re trying to clear a bar rather than win a market
The common thread is that these are situations where you don’t need to shout. The software finds its users because they’re already there - you’re building for yourself, for colleagues, or for a community you belong to.
AI is creating a new kind of generalist - someone who can sketch a UI, write backend logic with AI assistance, spin up infrastructure, and query data without deep specialisation in any of these areas. Previously, building a product required a team: frontend engineer, backend engineer, designer, infrastructure. Now one person with good judgment and AI tools can do work that once required five.
This person is enormously valuable. The question is: valuable for what?
Probably not for building another identical SaaS product in a crowded market. The distribution costs haven’t changed, and now you’re a one-person company trying to out-market well-funded competitors who also have access to AI. The AI-enabled generalist might instead be most valuable building internal tools, automating workflows within an existing business, or creating bespoke software for contexts where distribution is already solved.
The long tail of software might arrive, but it could look less like “thousands of small SaaS companies serving niche markets” and more like “companies building their own internal tools rather than buying generic ones.” The software still gets built, but it never enters the market. It’s produced and consumed within the same organisation, or within a community small enough that word-of-mouth suffices.
The optimistic read on AI and software costs has been that we’ll see a flowering of small software businesses serving long-neglected niches. I suspect the reality is more constrained: we’ll see a flowering of software, but much of it will never enter the market as products. It will be built inside companies for their own use, built by communities for their own members, or built by individuals for themselves.
This is still valuable - possibly more valuable, since internal tools don’t carry the deadweight costs of marketing and sales. But it’s a different vision than the one being sold. Not a long tail of software businesses, but a long tail of software that’s never sold at all.
The megaphone problem doesn’t go away just because the megaphones got cheaper. As build costs fall toward zero and attention stays scarce, the problem actually gets worse. The future of software might be abundant, but the future of software businesses looks increasingly winner-take-all.