The Project Is the Ecosystem

Companion to the essay applying Satya Nadella’s “human capital and token capital” argument to AEC. (Original Article)

What is Nadella actually arguing, in one line?

That in an AI economy the model is a commodity, and the durable asset is the learning loop you build on top of it—the system that encodes your institutional knowledge and compounds it with every use. He names two forms of capital every firm now has to build: human capital (judgment, relationships, pattern recognition) and token capital (the AI capability you own). His warning is that a frontier model without a frontier ecosystem is unstable, because value gets hoarded by a few models and the political economy won’t tolerate it.

And what’s the AEC-specific twist?

Nadella assumes the firm is the unit that learns and owns the loop. In construction, the firm has never been the unit that delivers the outcome—the project is. A building is delivered by a temporary coalition that assembles and dissolves. So the learning loop in AEC has to be owned by the project, not the firm and not the model. The knowledge that drives success lives in the connective tissue between parties, and today it leaks out at every handoff.

Isn’t this just “AI for construction”? What’s new?

The opposite of the usual pitch. Most AI-for-construction stories are about doing existing tasks faster—reading drawings quicker, drafting RFIs automatically. That’s “compute running in circles,” to borrow Nadella’s phrase. The argument here is that speed on a task is not the asset. The asset is a loop that captures the tacit judgment of a coalition and keeps it after the coalition disbands, so the ecosystem stops forgetting everything it learns on every job.

Why do you say the frontier model “already exists”?

Because it does. Any GC, architect, or owner can buy a frontier model today. Nobody in AEC has a model advantage, and nobody will—it arrives commoditized. The competitive question is not which model you pick. It’s whether you’ve built the ecosystem on top of it that lets the project learn. Picking the best model is the argument the industry is about to waste two years on. It’s the wrong argument.

What do you mean by “token capital is project-scoped by physics”?

In most industries, token capital can accumulate inside a company because the company is what persists. In construction, the company is not where the work happens. Owner intent, design rationale, means and methods, and field reality each live with a different party, and each is lost at the handoff between them. The only place the full picture exists is the project. So a learning loop that’s going to be real in AEC has to live at the project level—that’s not a design preference, it’s a consequence of how the industry is actually structured.

Where exactly does knowledge “leak”?

Four main seams. Owner intent degrades the moment it’s handed to design. Design rationale—the why behind a decision—is rarely written down and walks out with the engineer. Means and methods stay siloed with the GC and subs by contract design. And field reality (as-built versus as-drawn) is captured nowhere durable, which is why closeout is a scramble and facilities teams inherit buildings they can’t operate as designed. These are the interstitial gaps—the spaces between the parties where information and accountability leak.

Why is construction’s productivity problem relevant here?

Because the standard explanation—that the people or the trades are the problem—is wrong. The people are extraordinary. The problem is that the ecosystem has no memory. A coalition learns brilliantly for one project and forgets it all at closeout, then rebuilds the same knowledge from scratch on the next job. A project-scoped learning loop is the first mechanism that turns that disposable knowledge into a durable asset.

Does this mean AI replaces construction jobs?

No—and that’s not soft positioning, it’s structural. The work that drives a project is judgment: reading a site, anticipating a clash, knowing which sub to trust, understanding what the owner actually meant. You can’t offload judgment. What the loop does is make that judgment replicable across the coalition, so a veteran’s pattern recognition doesn’t disappear when the job ends. Expertise gets amplified and judgment gets made durable. The people become more valuable as the loop grows, because the loop is built out of what they know.

How does this connect to Nadella’s “societal permission” point?

He warns against an AI future that hollows out industries the way early globalization gutted manufacturing. Construction is where that risk is lowest, because the value is in human judgment that can’t be commoditized—and where the point is most worth saying clearly. The honest framing is amplification, not efficiency: judgment made replicable, not headcount removed. When the idea of AI displacement is in the air, lead with what the loop adds, not what it cuts.

What’s the strategic takeaway for a firm or an owner?

The firms and projects that own their learning loop will out-deliver and out-margin the ones renting intelligence one model at a time. Not because of a better model—because they’ve built the connective tissue that compounds learning instead of losing it at every handoff. If you’re allocating capital or running projects in the built environment, the moat is not the model. It’s the loop, and it belongs at the project level.

What should I do with this if I’m building in the space?

Stop optimizing for model selection and start asking where your projects forget. Find the handoffs where intent, rationale, methods, and field reality leak, and build the loop that captures them at the project level—owned where the work happens, portable across whichever model is best this quarter. The ability to swap a generalist model without losing the “project veteran” expertise built into your system is the real test of whether you own your loop or just rent intelligence.