Bundling and unbundling in the AI era
Software keeps swinging between bundling and unbundling, and people usually explain each swing after the fact with a story about vision. But I think the real driver is duller than that. It's the cost of running several tools that have to talk to each other. When that cost is high, the all-in-one wins. When the cost falls, specialists win again.
Agents make that cost fall a lot. An agent can read from one tool, write to another, reconcile them, and hand you the answer, and you never learn a second interface or maintain an integration. So specialists win. That seems like the whole argument, and for a while I thought it was.
It isn't, though, because the people buying aren't one group. Someone who wants one platform that does everything is telling you something about themselves: they don't want to think about tools. And people who don't want to think about tools also don't want to pay for five of them. That's a real market. It's just a commodity one. High volume, thin margins, and the competition is on price.
The other end has different arithmetic. A working podcaster doesn't want the podcast feature a newsletter company shipped last quarter. He wants the thing made by people who think about podcasting all day, and he'll pay more for it, because the tool sits upstream of the work. Price isn't what binds. Capability is.
Which turns the bundle's own pitch against it. The pitch was always: everything in one place, so it works together. Take away the second half, because agents now supply it, and what's left is "everything in one place." That's a description of the product, not a reason to buy it. And the specialist's old handicap — you'll have to wire this into your stack — quietly stops being one.
What's left worth owning is being the system of record. The canonical data for some domain, plus being genuinely good at that domain, plus making it easy for everything else to read and write. That works from either side. You can be the thing others plug into, or the thing that plugs in cleanly. The position that doesn't survive is the middle: passable at many things, canonical for none.
The bundlers are betting convenience compounds. I think convenience was a moat only while connecting things was hard. When it stops being hard, what you're left holding is whether the software is actually any good.
