In Retool's 2026 survey, 35% of respondents had already replaced a software subscription with something they built themselves. Retool's crowd skews builder, so the true number is lower. Even discounted, the direction holds: software is now the cheap part.
The code is not the moat. It never was. So what is it that people are paying for?
The result is what you sell now
The seat was always a stand-in. You charged per login because you could not charge per result. AI breaks the stand-in, because the thing doing the work is no longer the person holding the seat.
Watch how the sharpest AI companies price. Intercom charges $0.99 each time its support agent resolves a ticket, with no seat fee for the agent itself. Sierra, the agent company Bret Taylor co-founded, says it gets paid only when its agent completes the job; an unresolved conversation is, in most cases, free. Maxio, which sells usage-based billing software and surveys SaaS pricing, puts AI-native companies among the highest adopters of usage-based pricing. Bessemer, which funds many of them, now lays out a base fee plus outcome credits as the model to grow into, instead of a seat count.
This is not free money. Serving the work burns real compute, so the old margins are gone. Even the cohort Bessemer calls its "Shooting Stars" keeps only about 60% of its revenue after the compute bill, and the fast-growing ones far less. Software has a cost of goods again. You feel it every time a customer's usage spikes and your model bill spikes with it. The winners will be the ones who can point to a result worth more than the compute it costs, and charge for that.
There is a catch inside the model, and it is measurement. An agent that solves ninety percent of a problem before the customer gives up counts as zero under strict outcome pricing, so vendor and customer end up arguing over what a result even is. You cannot sell an outcome you cannot measure.
The shift goes further than pricing. In a March 2026 piece, Sequoia tells AI companies to sell the work itself, not the tool, and its math is blunt: for every dollar spent on software, six are spent on services.
The plumbing is the business now
If your customers can rebuild your app over a weekend, the real work is everything that keeps it alive once people depend on it. It has to stay up and keep the data straight. It has to control who sees what, and answer to someone when it breaks. AI made the first version nearly free. It did nothing to the cost of the tenth month.
The bill is already arriving. Retool, which sells tools to govern exactly these apps, surveyed senior technology and security leaders in May 2026: 93% were at least somewhat worried about AI-built tools running in production, and only 5% were very confident they could see everything running inside their own company. In the year before the survey, roughly one in five reported a production incident caused by a tool someone vibe-coded, meaning built by prompting AI rather than written by an engineer. Every one of those numbers is a business. Someone has to host the disposable app and keep it compliant. That market grows because building the app got easy.
Vercel, which mostly sells the unglamorous work of deploying and running other people's software, raised $300 million at a $9.3 billion valuation in September 2025. When the software is disposable, the ground it runs on is the business.
Some of the rip-and-rebuild wave, though, rides on AI sold below what it costs to run. When that correction comes, part of the "just build it yourself" math gets worse, and some of those weekend replacements quietly go back to buying. Bet on the layers that survive it.
The moat is closing the last mile
Anyone can prompt a model into a novelty, a clever demo or a party trick that works once. The novelty is free now. The scarce work is the finishing: taking AI from almost-working to just-working for one real, specific case. When the last mile closes, the product stops feeling like a tool you operate and starts feeling like magic. Magic is what people pay for.
The spending proves it. HeyGen sells avatar video with the lip sync handled, and says it passed US$200 million in annual recurring revenue in June 2026, doubling in eight months. Its founder's pitch is that people do not want video that looks or feels like AI. The same purchase is happening in hospitals. Abridge turns the conversation between a doctor and a patient into a finished clinical note inside Epic, the dominant hospital records system, and was valued at US$5.3 billion in June 2025, double its valuation of four months earlier. Transcribing a doctor is a demo anyone can build. A finished note inside the hospital's own system is the last mile, closed.
I am living a small version of this. Generating an avatar that looks like me and a voice that sounds like me took hours. But putting them together, with the lip sync held, is the genuinely hard part, and that is what I would pay for. The work that closes the last mile is the whole business. It is the part AI cannot cheaply finish, and the part your customer will not prompt their way past.
The context compounds
Here is the position I hold most strongly: when the code is cheap, the durable defensibility is context and memory. Two kinds. Curated domain knowledge, meaning the edge cases and unwritten rules of one industry, its language and its workflows, collected and encoded where a general model gets them wrong. And a company's own context: what the product has learned about one business and how it decides. Once a product holds that context, it is hard to migrate and hard to unpick. That is the stickiness, and it compounds. The longer it runs, the more expensive leaving becomes.
The prices being paid for this are not subtle. Harvey was valued at US$11 billion in March 2026, and the pitch under that number is context: it works from each law firm's own precedents and templates, so its output differs firm to firm. EvenUp crossed US$2 billion in October 2025 pricing injury claims off the millions of medical records it says it has read. Two different corners of law, one pattern: the product that holds the context commands the price.
The vertical AI companies Bessemer backs, software built deep into one industry, show it in aggregate. They already charge around 80% of what the incumbent software charges and are growing about 400% a year on Bessemer's early numbers. They win on encoded know-how. Even the money is saying it now: Bessemer writes that context and memory may be the new moats.
Then there is the founder signal. Bryant Chou co-founded Webflow, the website builder behind 0.9% of all websites on W3Techs' July 2026 count. He came back to Y Combinator in its Spring 2026 batch with Ploy, which looks like a website builder for about a minute; then it plugs into your analytics, search console and CRM, and works on your marketing while you sleep. On YC's own podcast he signed off on the hosts' summary: "a company brain for your marketing." The man who made building websites easy is betting his second act on the context around them, and in June 2026 he announced a US$27 million seed to build it.
Content is the scarce thing now
You can generate a thousand apps in a year. You cannot generate a video strangers choose to watch, or a guide they trust enough to follow. That kind of content stays scarce, because it comes from having done the thing.
I train jiu-jitsu. A model can write a technically perfect guide to escaping a choke, better organised than anything I would write. Mine is still the one worth reading, because I have been in the choke, and the reader can tell. As machine-made writing and video flood every feed, the stuff a real person made gets scarcer, and worth more. Audiences are already turning on generic AI output: enthusiasm for AI-made creator content fell from 60% in 2023 to 26% in 2025 in a Billion Dollar Boy survey.
This is why software companies are buying audiences outright. HubSpot has spent years acquiring newsletters and media brands, most recently Starter Story in February 2026, because attention has become harder to manufacture than software. When anyone can make the software, the scarce assets are the content people actually want and the trust that makes them show up for it. If you are a founder now, your real work in public is the top of the funnel: being the first thing a stranger finds, long before they would ever pay you.
Where Malaysia stands
I build from Kuala Lumpur, so the question I care about is where Malaysia stands on each of the five.
The result. As far as I can tell, nobody in Malaysia charges for the result yet. Every AI product I checked still bills by the seat or by the meter. The first company here to price like Intercom starts with the field to itself.
The ground. I went looking for the Malaysian company selling the layer that makes AI-built software safe to run. I could not find one. The vibe-coded apps landing in production here are all going to need it, and someone will get to sell it first.
The last mile. The one position with proof it can be done from here. Wise AI, a Malaysian identity-verification company, raised an eight-figure Series A in July 2024 for one boring job done properly: checking Southeast Asian identity documents and faces so a bank can onboard a customer remotely. Global vendors rarely walk that mile well for a market this size. Wise AI walked it and got funded for it.
The context. The biggest gap of the five, and the position I just called the strongest. The closest thing here is Cleve, a small KL-born team building what it calls a memory layer for human creativity. That is personal context, for creators, and it is still tiny. A product that holds a Malaysian company's context, I cannot find at all.
The content. The only position with no gatekeeper and no funding round required. By the Digital Ministry's count, Malaysia's digital creative industry has exported RM12.1 billion to date, and a creator in KL competes with the world the day they publish.
Whether Malaysia ends up owning any of the intelligence itself is a longer argument, and I make it in a separate essay: Can Malaysia be more than an AI landlord?.
Where the moat went
- Pricing leaves the seat and does not come back. The first billing disputes will test outcome pricing, and the vendors who can prove a result in the invoice will own their categories.
- The ground outlasts the apps. The apps are disposable. The ground they run on has to hold up for years, and somebody is going to own that layer.
- Context beats the buzzword. Memory is about to be on every vendor's slide. The ones that hold a customer's context become too expensive to leave, and the rest get found out.
- Malaysia's openings get taken. Three of the five positions stand empty here as far as I can tell. Either somebody local claims them, or foreign vendors will.
I build in this market. These are the bets I am making with my own time, and in two years we will know which were right.
There is a second half to this question: which of today's software companies does the value leave first? I have been mapping that, name by name, and it will be its own piece.
Personal view, from building owned-code software with AI. Not investment advice; companies named as examples, not endorsements.