"Is software dead?" is the question doing the rounds at every SaaS conference in 2026. The veteran investor Rory O'Driscoll gave the honest answer at SaaStr: no — but it just got a lot harder to win. That's not a comforting answer, and it shouldn't be. It's a useful one, because it tells you exactly what to build.
The argument isn't that AI kills software. It's that AI removes the moats software used to hide behind. When a capable model can replicate a competent feature in an afternoon, "we built a nice CRUD app for X" stops being defensible. O'Driscoll's framing was memorable: of the software companies that existed when ChatGPT launched, roughly 10% were "dead on arrival," and the remaining 90% split into rough thirds — insulated, additive, and threatened.
He also named the traps. Repackaging a pre-2022 product as "AI-native" without genuine architectural change. Mistaking panic-driven corporate mandates for durable demand. Spending 40% of revenue on model inference while building only marginal value. All three are ways to look busy while the moat drains.
The valuable half of the talk was the list of what still holds. Defensibility now comes from things a foundation model can't simply absorb: proprietary data the model never trained on, network effects from multi-sided marketplaces, full-stack business models where you become the service rather than sell a tool, and data flywheels that compound as your product runs. Notice the through-line: every durable moat on that list is about owning something — data, a network, a transaction relationship — rather than merely coding something.
There's a second, quieter shift underneath. The "harness" — the software orchestrating models through context, filtering, logging, and escalation — is becoming the new technical foundation, the way the LAMP stack once was. Everyone builds on a similar substrate now. Which means the differentiation moves up, into the business model and the data, and away from the plumbing everyone shares.
Here's the strategic bind O'Driscoll drew, and it's sharp: "you get compute that sells itself, or people who sell it. You can't fund both." Model-heavy companies run 70–80% compute costs; application companies spend around 10% of revenue on inference. Try to carry a heavy enterprise sales team and a heavy compute bill at the same time and the maths simply doesn't close. The companies that win pick a lane — and the application lane, where inference is a small fraction of cost, is where most durable software businesses will actually live.
Read that list of surviving moats — proprietary data, marketplaces, full-stack ownership, low inference cost, ship-fast-to-find-the-flywheel — and it reads almost like a spec sheet. That's the case for building on VBWD, a self-hosted, source-available platform for subscriptions, marketplaces, and AI apps.
Take the moats one by one. Proprietary data: VBWD keeps your customer and transaction data in your own database, so the flywheel compounds for you, not a vendor. Network effects: the platform ships a marketplace with creator payouts out of the box, so a two-sided model is configuration rather than a year of build. Full-stack ownership: billing, subscriptions, tax, invoicing, and payments (including non-custodial crypto) are already there, so you can be the service, not just tool it. Low inference cost: the AI features resolve a model connection you control, so you decide where inference runs and what it costs. And ship-fast: because it's a modern plugin architecture that ships with AI coding-agent instructions in the repo, standing up the first version of an idea is closer to days than quarters — which is exactly how you find the data flywheel before your runway ends.
The point isn't that a platform hands you a moat. It doesn't — the moat is your data, your network, your judgement about what to build. The point is that VBWD hands you the foundation those moats sit on, so your scarce time goes into the defensible part instead of rebuilding billing for the thousandth time.
None of this is a shortcut past the hard part. "Software isn't dead, it's harder to win" means the winning is still on you — VBWD gives you a fast, own-your-stack starting line, not a finish line. It's self-hosted, so you run it. And it's one approach among several; if you're a model company, your problems live in a different cost structure entirely.
But if you're building an application business — the lane where most software will keep being built and won — the O'Driscoll checklist and the VBWD design point the same direction: own your data, own the transaction, keep inference cheap, and ship fast enough to find the flywheel. Software isn't dead. It just rewards ownership now, and punishes renting.
VBWD is free for commercial use below 6.7 BTC a year in attributable sales. See the plugins, the architecture, and the docs, or browse the source on GitHub.