A line in AEC Magazine‘s “BIMnopoly” stopped us mid-coffee this week: the claim that “physical AI” — large multimodal models that learn the syntax and semantics of building data across heterogeneous sources — could “decouple meaning from format.” For a fourteen-person office that has paid the format tax for two decades, that sentence is not hype. It is a budget line.

Here is the lineage, because the lock-in was engineered, not accidental. As the piece traces it, control started at the file boundary — proprietary formats, object libraries, application-specific parametric behaviours — then migrated to the cloud, where the gate moved from “who can open the file” to “who can access the data.” IFC was supposed to be the escape hatch; in practice, as the article bluntly notes, it reads cleanly but writing leads you straight back into the authoring tool. We learned that the expensive way on a Generalplaner mandate two years ago, when our “open” round-trip quietly dropped half our custom property sets.

The research-side reason this might finally shift sits next to the format war, not inside it. The same geometric-reasoning and graph-learning methods that power neural radiance fields — the NeRF-in-the-Wild work from Martin-Brualla and colleagues at CVPR 2021, now extended to aerial city-scale benchmarks like the TU Delft survey — treat a building as learned structure, not as a vendor’s internal schema. When meaning lives in a model that read a thousand heterogeneous sources, the strategic value of a closed format genuinely declines. That is the AEC Magazine thesis, and the science under it is real.

←TODAY: In 2026 our IFC still round-trips one direction; semantics we author in Archicad don’t survive a return trip from a partner’s platform. →3012: The model becomes the asset and the format becomes a render target — owning the meaning, not the file, is what an office sells. Fulcrum: Only an office that already keeps disciplined, machine-readable semantics today can hand them to an AI interpreter tomorrow; garbage in still means garbage out, with or without the model.

That last point is the trap. “BIMnopoly” calls the BIM model an “empty canvas” whose informational quality depends entirely on manual input — naming conventions, classification, assembly logic. A semantic AI does not rescue a sloppy model; it inherits the sloppiness at scale. The honest trade-off: physical AI lowers the cost of escaping a vendor, but it raises the cost of bad data governance, because now your mess is legible to everyone downstream.

Atelier: On our current execution model we treat the BEP and a written LOIN as the real product — the IFC is just where it gets serialised. The week we wrote down which party owns each property set, our quantity take-off stopped disagreeing with the engineer’s, and that discipline is exactly what an AI interpreter will need to read us correctly.

Hack: This Hack teaches you to audit, in one move, whether your IFC export actually carries the semantics you think it does — before you trust any “open” round-trip. The medium is runnable code; the domain is IFC. Walk the file with IfcOpenShell and count what survived:

import ifcopenshell
model = ifcopenshell.open("competition_export.ifc")
walls = model.by_type("IfcWall")
typed = [w for w in walls if w.IsTypedBy]
print(f"{len(walls)} walls, {len(typed)} carry a type/property set")
# If typed << walls, your 'open' export is geometry, not information.

Run it against last week’s export. The gap between the two numbers is the size of the format tax you are still paying.

PAZ has covered the underlying competence before — our “Are you BIMlingual?” piece argued that an office fluent across Archicad, Revit and Grasshopper isn’t chasing tools, it is refusing to be monolingual. Physical AI is that argument’s next chapter: fluency moves from the operator to the model, but only if the office fed it clean grammar.

PAZ Takeaway: The capability that answers this story is the PAZ Grasshopper↔Archicad Library — it personalises the seam between the two platforms and fills the gaps the native connection leaves open, so the semantics your office authors are the ones that round-trip. That is the asset an AI interpreter will read; documenting it is cheaper than escaping the next vendor.

Don’t wait for the format war to end. Run the IfcOpenShell audit on your last export this week, and write down who owns each property set — that ledger is the only thing that will still be yours when meaning stops living in the file.

Source: AEC Magazine — BIM / computational / AI tooling