Skip to content
Fritz Gerald ZephirinMenu

Social

Models Are Commodities. Workflows Are Moats.

4 min read

In January 2025, a Chinese lab released an open-source reasoning model that matched the market leader’s flagship, and $600 billion came off NVIDIA in a single day. Eighteen months later, the most-downloaded model family on earth is Alibaba’s Qwen, approaching a billion downloads, and the price of frontier-adjacent intelligence has fallen so far that it is effectively a utility rate. Every frontier capability spread of the past four years — GPT-4’s, o1’s, each successive flagship’s — has compressed to parity within roughly a year and a half.

That is what commoditization looks like while it’s happening. And yet, over the same period, a code editor wrapped around other people’s models went from zero to $500 million in revenue and a $9.9 billion valuation, and the most important infrastructure story of 2025 wasn’t a model at all — it was a protocol for connecting models to tools, adopted by every rival lab at once. The value didn’t vanish when the models commoditized. It moved one layer up.

Clay Christensen had a name for this: the law of conservation of attractive profits. When one layer of a stack becomes modular and interchangeable, the profits don’t disappear — they migrate to the adjacent layer where integration is still hard. PC hardware commoditized; Microsoft and Intel collected. Airlines commoditized; reservation systems and loyalty programs collected. The pattern is one of the most reliable in business history, and the AI stack is running it in fast-forward.

So what is the adjacent layer where integration is still hard? I run AI in production — support floors, payroll operations — and I can tell you precisely, because I pay for it monthly. It is not the model. Swapping models is a config change; we have done it more than once, without ceremony. What took months, and what no vendor could sell us, was everything around the model: which knowledge base counts as policy when two articles disagree. Which conversations the AI may finish and which it must hand to a human, and how that handoff carries context. What the agent is allowed to touch, per client, per country, per regulation. The eval set built from our actual tickets that tells us whether this week’s model is safe to ship. The exception list — grown entry by entry, year over year — that encodes every way reality has disagreed with the documentation.

That layer is the workflow, and it has the two properties moats need. It appreciates with use — every exception handled makes it more complete, every audit makes it more trusted, which is why it behaves like a compounding asset rather than a depreciating one. And it is illegible from outside — a competitor can see our results and copy our tools, but the thousand encoded decisions are exactly the part that can’t be read off the surface. The model is the same for everyone. The workflow is the company, in executable form.

The objection with real teeth: won’t the labs integrate upward and take the workflow layer too? They’re visibly trying — agents, browsers, enterprise suites. Watch where they succeed: broad, horizontal, consumer-shaped workflows. The browser is a workflow everyone shares; the labs will own it. But enterprise workflows fragment by domain, regulation, and history — there is no general model of your escalation tree, your compliance posture, your client’s definition of resolved. Fragmented, domain-deep territory is where incumbents and focused builders win, which is exactly what the buy-versus-build evidence shows: purchased generic tools plus deep workflow embedding beats both extremes. A second objection — that some new Mythos-class model will restore a durable capability spread — runs against four years of compression history; even the most dramatic frontier launches now price as temporary leads, not moats.

One honest caveat, consistent with this site’s own rules: moats move up the stack — that is the general law — and they don’t stop moving on your account. Agent frameworks are already trying to make workflows portable, the way MCP made tools portable. If they succeed, today’s workflow moat becomes tomorrow’s commodity, and the durable layer moves again — to the data exhaust, the trust, the accountable humans. Strategy in a stack like this isn’t a position; it’s a heading.

For builders, the heading is clear enough to act on. Treat the model bill like the electric bill. Put your engineering where the integration is hard and yours: context, permissions, exceptions, evals, handoffs. And when a new model drops and the feeds light up, ask the only strategic question in the launch: does this change my workflow, or just my unit costs? Usually it’s the second. The second is good news that requires nothing from you. The first is the moat moving — and you want to feel it before your competitors believe it.

Sources & further reading