Businesses Are Replacing SaaS Subscriptions With Self-Built AI Tools

TL;DR: A company publicly deleted a $600,000-a-year Salesforce bill by building its own CRM instead of renting one. Days later, investor Chamath Palihapitiya said the entire $5 trillion software market, licenses and services combined, is getting rebuilt the same way. Developer Theo (t3.gg) put a number on how low the bar has fallen: an entire startup can now be a markdown file handed to an AI coding tool. Cooper Simson’s own version of this story is smaller but just as telling: an agency quoted him $600 a month for an Instagram DM agent, so he built one himself with Claude for about $13 a month. None of this requires a computer science background. It requires knowing what a self-built stack is made of, and where to start.

The problem: rented software is built for everyone, so it fits no one

Every SaaS product you pay for was designed to serve thousands of different businesses at once. That is the whole economic model, one codebase, one set of features, sold to as many customers as possible. It is also the reason so much of it feels like a compromise. You pay for modules you never touch, you adapt your workflow to fit the software instead of the other way around, and every price increase lands on your desk with no negotiation.

That compromise just got a very public price tag. In mid-July 2026, VC Harry Stebbings posted about a company that had walked away from Salesforce entirely. The company built its own CRM, one shaped exactly around how its team and its AI agents actually work, and its $600,000-a-year software bill went to zero. The post did over half a million views, not because deleting a subscription is rare, but because deleting a subscription of that size, replaced by something the company built itself, is still rare enough to stop people mid-scroll.

The $5 trillion thesis

A week before that, Chamath Palihapitiya made a much bigger claim. On his X account, he argued that the entire software market, roughly $1 trillion in licenses plus $4 trillion in services, is merging into a single $5 trillion category that gets rebuilt by AI-assisted, self-built tools. He is not just commenting on this from the sidelines. He said he is betting his fund, 8090, directly on it. That post pulled nearly a million views.

Whether the exact number lands at $5 trillion or somewhere else, the direction of the claim matters more than the figure. Two of the loudest voices in software and venture capital, in the same month, both pointed at the same shift: the economics of renting software are becoming harder to justify next to the economics of building it yourself.

The proof this isn’t just talk: a markdown file and a Claude Code session

Skepticism is fair here. Big claims about AI replacing entire categories of software are common and rarely hold up. What makes this one different is how boring the mechanism actually is.

At the AI Engineer 2026 conference, developer Theo (t3.gg) described watching the bar for building software drop in real time. His framing, paraphrased from his talk: the whole product can just be a markdown file that you hand to Codex or Claude, and the tool builds from there. That clip did over 260,000 views on its own, mostly from developers who recognized exactly what he meant.

A markdown file is nothing more than plain text with some formatting. What used to require a spec document, a project manager, a design pass, and a development team now starts as a page of notes handed directly to an AI coding assistant. That is the actual mechanism behind the Salesforce story and behind Chamath’s thesis: the cost of building custom software collapsed, so the cost of renting generic software stopped making sense by comparison.

A smaller, more personal version of the same story

The Salesforce number is dramatic, but the pattern shows up at a much smaller scale too, which is actually the more useful data point for most people reading this. Cooper Simson, who runs the content and consulting brand Actionable AI, went shopping for an Instagram DM automation tool for his business. An agency quoted him $600 a month to build and run one.

Instead, he built it himself using Claude. The agent runs 24 hours a day, replies to comments, qualifies leads, and routes people to the right resource, and it costs him roughly $13 a month to operate. Same job, same uptime, a fraction of the cost, and it fits his exact workflow instead of a generic template someone else designed.

That is the same shift as the Salesforce story, just at a scale almost anyone reading this can relate to. You do not need a $600,000 enterprise contract to feel this. You need a $50 or $200 monthly subscription for something a self-built tool could replace, and increasingly, that tool does not require you to know how to code.

What a self-built stack actually looks like

This is the part most coverage of this trend skips: the practical shape of what “building it yourself” means in 2026. You are not writing raw code line by line. You are directing an AI system that writes and assembles the code for you, describing what you want in plain language the same way you would brief a freelancer.

A minimal self-built stack has five pieces:

The builder. This is the AI system that actually writes and assembles your software from plain-language instructions, the same role a development team used to play. You describe what you want, it writes the code, tests it, and fixes what breaks.

A database. Anything your tool needs to remember, customer records, orders, conversation history, has to live somewhere. Modern database platforms are built specifically to pair with AI-assisted development, so the builder can set up and manage the tables for you instead of you writing raw database queries.

A place to deploy it. Once your tool works, it needs to run somewhere other than your own laptop, so it stays live even when your computer is off. Deployment platforms built for this handle that with a single command instead of a server you have to manage yourself.

Connectors, often called MCPs or skills. These are the pieces that let your AI builder actually talk to other services: your email, your calendar, your ad accounts, your social platforms. Without them your tool can only reason about text. With them, it can take real actions in the tools you already use.

Memory and context files. Plain text files that describe your business, your preferences, and your rules so the AI system does not have to be re-explained every time you start a new session. This is the closest thing to institutional knowledge a self-built system has, and it is often just a folder of notes.

None of these five pieces require a formal engineering background to use. They require knowing they exist, and knowing roughly what each one is for.

Where to actually start

If this is your first time building something instead of buying it, do not start with the most ambitious version of your business. Start with the smallest real thing that would save you money or time today.

An internal dashboard that pulls a few numbers you currently check manually is the easiest first build, because it only reads data, it does not need to change anything for you. A lightweight CRM that tracks your leads and their status is a reasonable second step, more useful but slightly more involved since it needs to store and update information. A comment-to-DM agent, something that watches your social posts and automatically responds to and qualifies interested people, is the most ambitious of the three, closer to what replaced that $600 a month agency quote, and worth attempting once the first two have built your confidence with the process.

In every case, the starting point is the same: describe what you want in plain English, as specifically as you can, and let the builder handle the implementation. The gap between “I have an idea” and “I have a working tool” has never been this short.

When you should still buy the software

None of this is a blanket argument against SaaS. There are real categories where renting still beats building, and pretending otherwise would be dishonest.

Anything compliance-heavy, accounting software tied to tax filing requirements, for example, carries regulatory obligations that a self-built tool has no business taking on casually. Anything that requires direct bank account syncing or handles regulated financial data benefits from a vendor whose entire job is staying current with security audits and compliance certifications you cannot reasonably replicate yourself. And any tool in a genuinely regulated industry, healthcare records, financial advice platforms, anything with legal liability attached, should stay with an established vendor until the self-built tooling in that specific space has matured and been independently audited.

The honest version of this trend is not “never pay for software again.” It is “stop paying full enterprise price for something a plain-language conversation with an AI builder could replace, and reserve the rented software for the categories where the rental itself is the value, not the convenience.”

The takeaway

A company deleted a six-figure Salesforce bill. A billionaire investor is betting his fund on the same pattern repeating across a $5 trillion market. A developer described the entire mechanism as a markdown file and a coding assistant. And on a much smaller scale, a $600-a-month agency quote turned into a $13-a-month self-built tool that runs exactly the way one person’s business actually works.

The tools you rent were built for everyone, which is exactly why they fit no one perfectly. Software was never the real advantage. The workflow built around it was. Building your own tools means you own that advantage instead of renting it from someone else, one subscription at a time.

Comment “VIBE” and I’ll send you everything :)