The Death of SaaS, and What Comes After
I've been sitting with this thought for a few months now and I think the conclusion is hard to avoid: SaaS is dying. Not in a slow decline way. In a "the ground is shifting under the business model" way.
Here's what I mean.
The cost of software is going to zero
Jia Yuan Zhang wrote a piece recently that I keep coming back to. The historical parallel that hit hardest: when Gutenberg invented the printing press, the cost of copying collapsed. Europe went from 30,000 books to 20 million in fifty years. Scribes vanished. But the demand for judgement about what to publish only grew.
Same thing is happening with code right now. Andrej Karpathy went from 80% hand-written code to 80% agent-delegated in a single month. That's not incremental improvement. That's a phase change.
When I built a Telegram-to-Notion bot last year, I wrote about how Lovable compressed what used to be a multi-day build into a single afternoon. That already felt fast. Now, with coding agents like Claude Code, the same build would take maybe an hour. And it keeps getting faster.
The marginal cost of producing software is heading to zero. So if you're selling software subscriptions, what exactly are you selling?
Tools vs. work
Julien Bek at Sequoia nailed this: "If you sell the tool, you're in a race against the model. But if you sell the work, every improvement in the model makes your service faster, cheaper, and harder to compete with."
His example is sharp. A company spends $10K/year on QuickBooks and $120K on an accountant to close the books. The next big company will just close the books. No tool. Just the outcome.
For every dollar spent on software, six are spent on services. That's the real TAM.
Y Combinator's latest Requests for Startups now explicitly asks for AI-powered agencies. Their argument: AI flips the agency model on its head. Agencies used to mean low margins and scaling by adding people. Now you can deliver finished work at software margins. Bek calls these "autopilots" (as opposed to "copilots" that just assist). Copilots sell the tool. Autopilots sell the work. And autopilots capture the work budget, which is always bigger than the tool budget.
The electricity trap
Here's the thing most people get wrong though.
Zhang's article digs up a stat that I think about a lot: when electricity arrived in factories in the 1880s, it took forty years before productivity gains actually showed up in the data. Early factories just swapped their steam engine for an electric motor and changed nothing else. Same layout, same processes, same org chart.
The real gains came when people designed entirely new factories around what electricity made possible. That's when Ford's assembly line happened.
We're in the steam-engine-swapping phase right now. Studies show most orgs using AI coding tools see no measurable productivity gains. Makes sense if you think about it. If you just give your existing dev team Copilot and change nothing else, you speed up one part of the pipeline and create a bigger pile-up in review and testing. Faster typing, same bottlenecks.
I keep a running list of concepts and mental models I come back to, and one I've been thinking about recently is the idea of a false peak in gradient ascent. That's what "AI-augmented same-old-process" feels like. Local maximum. The real gains require you to blow up the existing structure and rebuild.
What we actually did
At Voltade, we didn't bolt AI onto the existing setup. We tore up the org chart.
We bought everyone a Mac Mini and put them on Claude Max. Not as a perk. As infrastructure. The Mac Mini is the workstation, Claude Max is the engine, and each person is now their own vertical. There are no "developers" or "PMs" or "designers" any more. Everyone is an agent orchestrator who handles their work end-to-end: scoping, building, shipping, supporting.
I have a backend engineering background (I built internal tooling and automation systems at Ninja Van before moving into product), so the idea that a single person can now handle what used to require a cross-functional team isn't abstract to me. I've seen both sides. And the difference in velocity is real.
We stopped building new products entirely. Every engagement is a custom agent build, deployed on managed infra. Each person owns their client relationship and automates as much of the execution as they can.
The role boundaries dissolved because AI made them artificial. When an agent handles the implementation, the bottleneck is judgement: knowing what to build, for whom, and why. That's what we optimise for now.
Is it working? Too early to say definitively. But the velocity difference is noticeable. And the structure feels right in a way the old one stopped feeling about six months ago.
First movers
Most companies are still in the "bolt AI onto existing roles" phase. Give the devs Copilot, give the PMs some chatbot, call it transformation. The productivity numbers will disappoint them, for the same reasons electricity didn't help factories that just swapped the motor.
The ones who redesign the whole system around what AI actually makes possible, those are the ones who'll be hard to catch.
SaaS had a good run. But when the cost of building software approaches zero, selling software subscriptions stops making sense. What comes next looks more like a new kind of firm. One that sells the work, not the tool, and compounds with every model improvement.
We're trying to be that at Voltade. Early to it, probably wrong about some of the specifics. But the direction feels clear.
If you want to see what this looks like in practice, I wrote about building two AI agents for Voltade using OpenClaw. One manages our customers, the other runs internal ops. That post is the concrete version of the ideas here.