Pedro M. Dominguez pedromdominguez.dev

01 What changed

The clever part got cheap. The part that is yours did not.

For most of the last decade, the expensive thing about software was the software. That is ending, and it changes what a small business can own.

The cost of running an AI model has fallen about 1,500× in six years2, and the lead any one company holds now lasts weeks rather than years — publish a result and rivals match it almost immediately. Intelligence is turning into something you buy by the gallon, like electricity.

When the clever part becomes cheap, the valuable part becomes the part nobody else has: how your business actually works. Your customers, your prices, your schedule, the fifteen small decisions your staff make without thinking. In the industry that is called the application layer — the software a business actually touches. It is the one layer a roofer or a church can genuinely own.

The real business model is not in the foundational model anymore. It's at the application layer above and it's in the infrastructure below, whether that's the cloud or whether that's chips.
Chamath Palihapitiya — All-In Podcast, 24 July 20261

Two honest notes about that quotation. It says the application layer and the infrastructure below — both, and the second half matters as much as the first. And he is describing where investors should expect value to sit, not arguing that your business should own its software. That second step is mine, and the rest of this page is me making it rather than borrowing his authority for it.

02 Renting or owning

Two ways to have software.

Almost every business in Oklahoma City is on the first path, usually without having chosen it. The second one only became affordable recently.

Renting

  1. Your business
  2. A monthly subscription
  3. Software built for ten thousand businesses

The vendor owns the software, your data and what gets built next.

Owning

  1. Your business
  2. An open-source foundation
  3. Software built for you, with AI

You own the software, your data and how the work actually gets done.

Renting is not a swindle. It was the only thing that made sense when building anything custom cost tens of thousands of dollars — far better to split the cost of one product between ten thousand businesses. The price of that bargain is that the software is built for the average of those ten thousand, and you shape your business around it: you change how you take bookings because that is how the software takes bookings.

What has changed is the cost of the alternative. Building software specific to one business used to be the expensive path. It is becoming the affordable one, and the thing you get at the end is an asset rather than a receipt.

03 Cheap is not worthless

None of this means the rest of it stops mattering.

It would be easy to overstate this argument, so here is the limit of it. Chips, data centres, power, databases and the models themselves are not becoming worthless — the same research that describes models getting cheap calls infrastructure “the most concentrated layer in the stack”2, and the sentence quoted above puts real value below as well as above.

The claim is narrower, and it survives being pushed on. AI has collapsed the cost of assembling software — the plumbing, the forms, the tests, the tedious parts. What is left, proportionally, is the part that is specific to one business. So a larger share of the value ends up in the layer a business can own, even while everything underneath it stays valuable and stays rented.

04 What it is built on

A foundation small enough to read is a foundation you can leave.

This is where the technical choices earn their place — not because they are clever, but because of what they cost you later.

The software runs on Deno, an open-source runtime that starts with permission to do nothing at all and has to be handed each capability explicitly. The web server is @std/http from the JSR standard library. Every piece of information arriving from outside is checked by Zod before anything else sees it. Beyond that it uses what the runtime already provides rather than reaching for a framework, and it is built the way Unix tools are: small pieces, each doing one job, that you can understand one at a time.

The economic argument is the one that matters to you. A foundation you can read is a foundation you can leave — no vendor can re-price it, discontinue it, or decide your business is no longer a priority. There is nothing underneath it that expires and forces a rewrite in three years. And because the plumbing is small and standard, AI handles almost all of it, which frees the effort for the only part that is actually yours: how your business works.

  1. 01 Internet
  2. 02 Nginx
  3. 03 Deno
  4. 04 Your functions
The entire request path. Four hops, one process, no orchestration layer to operate at two in the morning.

05 The objections

Everything wrong with what I have just argued.

If these are not answered honestly, nothing above them is worth reading. Some of them I can answer. One or two I can only concede.

AI writes insecure code.
Often it does. Veracode put more than a hundred models through eighty coding tasks: 45% of what came out had a security flaw in it, and that number has barely moved in two years. So the code is not trusted because a machine wrote it. Everything arriving from outside is checked before anything else sees it, the server starts with permission to do nothing and is given back only what the job needs, and the whole suite runs on every change. This site is built that way too — the source is public, and you can read it.3
Who maintains it when you are busy, or gone?
The honest answer is that this is the strongest objection on the page. Day to day, the $20 a month is exactly this. Beyond that: the code is yours, it is small enough for another developer to read in an afternoon, and it is built on tools thousands of people already know. That is a genuinely better position than a custom system nobody else can open — but it is not the same as no risk, and I will not pretend otherwise.
Custom software rots. Ours will become a mess.
Big ones do. The defence here is size: a system small enough to hold in your head is one that can be changed in an afternoon three years from now. There is no framework underneath it waiting to force a rewrite, and very little borrowed code to keep in step.
What about the rules — cards, health data, anything regulated?
Then it is out of scope unless it is designed for from the start, and I will tell you that in the first conversation rather than the last. Card payments go through a processor that already carries that burden. Regulated records are a different kind of project with a different price, and pretending otherwise would be how people get hurt.
Do we really own the data?
Yes, and it is worth being concrete. The domain is registered in your name. The site runs on hosting in your name. The database is a file on a machine you pay for, and you can take a copy of it whenever you like. Nothing is held anywhere you cannot reach without me.
You still depend on somebody. Deno, a host, an AI company.
True, and worth saying plainly: nobody is independent of the internet. You rent a server, a domain, and — where the software uses AI — somebody's models. The claim is narrower than independence. It is that the layer holding your work, your customers and your way of doing things belongs to you, and that everything underneath it can be swapped for something else without rewriting that layer. Renting compute is not the same as renting your own operations.
Can a business our size really carry custom software?
Not alone, and I would not suggest it. That is what the monthly fee is for. Owning it means the asset and the exit belong to you — not that you are handed a repository and wished luck. The difference shows up the day you want to change something, or the day you want to leave.

06 Sources

Where every number here came from.

Each one was read at the source rather than taken from a summary. Where a claim could not be confirmed — an episode number, in one case — it is not stated.

  1. All-In Podcast — “The Fight Over Open Source AI…”, 24 July 2026 Chamath Palihapitiya, around ten minutes in, on models commoditising “much faster than anybody thought” and where the business model has gone. Full transcript.
  2. Chamath Palihapitiya — “Deep Dive: Where Value Accrues in the AI Stack” A six-layer map of the AI economy, researched in the first quarter of 2026. The source of the 1,500× fall in the cost of running a model, and of the point that infrastructure is “the most concentrated layer in the stack”.
  3. Veracode — GenAI Code Security Report (2025, updated spring 2026) Over 100 models across 80 coding tasks: 45% of the code they produced introduced a security flaw, and the pass rate has barely moved in two years.
  4. Cledara — average SaaS spend per employee, 2026 Benchmarks for what companies pay to rent software. Note that “small business” in surveys like this means firms far larger than a roofer or a church — read it as direction, not as a bill anyone here is receiving.

The whole argument, in three steps.

  1. A small open-source foundation
  2. AI collapses the cost of building
  3. You own the layer where your value lives

A local business does not have to be only a customer of the AI economy. It can own the software through which AI does its work.