Amazon employs roughly 1.5 million people. Most of them work in fulfillment centers - picking, packing, moving physical goods. Politicians, journalists, and tech critics have used this as an attack for years: if Amazon is a tech company, why does it have more warehouse workers than almost any other employer on earth?
The question assumes the warehouse headcount is an embarrassment. It isn’t.
The half-finished construction
Software has always needed two things simultaneously to generate value: someone to use it, and someone to maintain it. That second dependency is the one most software narratives quietly skip.
When a product scaled - more users, bigger audience - the maintenance layer scaled with it. More infrastructure, more engineers, more support staff. The labor required to sustain the software grew roughly in proportion to the user base it served.
So the productivity math was always circular. You gained leverage by deploying software, then spent a significant portion of that leverage maintaining it. Real gains, but bounded.
Robert Solow named the gap in 1987: “You can see the computer age everywhere but in the productivity statistics.” Computing capacity had grown roughly a hundredfold over the previous two decades. Productivity growth had slowed. The paradox seemed to dissolve in the 1990s productivity boom - new markets, new leverage, before the maintenance overhead caught up. But the underlying structure returned.
In early 2026, Fortune reported that AI had so far had no measurable impact on employment or productivity across major organizations. The same paradox, thirty-nine years later, with a different technology.
Marc Andreessen wrote in 2011 that software was eating the world. He captured something real. But the framing slightly missed: software was amplifying the world, and the parts of the world worth amplifying were exactly as complicated as they’d always been.
The proof
Amazon’s structure is what the thesis looks like when it works.
The recommendation engine, the fulfillment routing, the inventory management systems - they amplify an operational infrastructure. The 1.5 million people in the warehouses are the product. Amazon’s competitive advantage is the integration: software that makes human-powered logistics run at a cost and scale no competitor has matched.
Bezos was never embarrassed by the warehouse headcount. He was building a system in which the software was the amplifier, the operations were the value, and the integration between them was the moat.
Every software business that has ever worked at scale has had this architecture, whether or not the founders described it that way. The software amplifies the human operation - that was always where the value was.
The “not a tech company” criticism assumes a large service workforce signals weakness. Amazon answered by becoming one of the most valuable companies in history.
The loop starts to change
Something shifted in early 2026.
Morgan Stanley published a “gut check” on enterprise SaaS. The argument: AI agents are starting to bypass GUI software entirely. When the user of a product is an AI agent calling an API directly, software designed for human operators becomes a layer you no longer need. Fortune ran with the story under the headline about Andreessen’s prediction coming true “in a way nobody imagined,” and called it the “SAAsocalypse.”
This is the first clear signal that the user side of the amplifier loop is changing. The “someone to use it” requirement - which had always been a human - is now starting to be satisfied by agents. The maintainer side is shifting too, more gradually: AI-assisted development has compressed the maintenance cost curve in ways that weren’t possible three years ago.
If AI takes both ends - user and maintainer - the circular dependency that capped software’s productivity contribution since the 1970s starts to unwind. That’s a different kind of change than any previous tooling wave. Better frameworks made maintenance cheaper, cloud made it more elastic, AI-assisted coding made it faster. None of them broke the fundamental structure: software still needed humans on both ends.
This might be the first thing that does.
For forty years, every “software company” was really a company using software to amplify something else - a logistics operation, a marketplace, a content library, a financial system. The people in the warehouses, on the support lines, in the operations centers weren’t overhead. They were what the software was amplifying.
Amazon has been showing this architecture since 1994. The warehouse headcount wasn’t accidental to its scale - it was evidence of what they actually built.
The half-finished construction might be getting finished. What that leaves behind is what it was amplifying all along.