PAZ Kaffi Review of the Week W25: The Week AEC Learned to Draw Its Own Single Points of Failure
The week AEC learned to map its single points of failure — grid as the new AI bottleneck, vendor lock-in stress-tests, auditable BIM, and Archicad AI.
Twenty stories crossed the desk this week, and one shape kept reappearing under all of them: the dependency graph. Power, models, fire code, smart-building subsystems — story after story asked the same Swiss-procurement question I have asked for eight years. Where does this thing break, and who is holding the line when it does?
←TODAY: In June 2026, AEC’s scarcest resource quietly shifted from GPUs to megawatts, and a fake Mistral meme exposed how fragile our AI supply chains already are. →3012: By the Zurich-3012 horizon, the studios that survived were the ones that drew their own articulation points instead of trusting a vendor’s. Fulcrum: The insight only lands if you can see both at once — the future bottleneck is grid and governance, and the present fix is a graph you draw this week.
Top stories:
The week’s loudest signal was infrastructural. “Power, Not Chips” reported a $27M seed betting the real AI constraint is the grid — 110 of 123 simulated futures agreed. For any office touching data-centre work, this reframes power, batteries, and acoustics as first-class program, not an afterthought tucked behind the render. As Forbes laid it out, the megawatt is the new bottleneck, and that is a siting-and-procurement problem before it is an engineering one.
Second, vendor sovereignty stopped being theoretical. “The Fat Kitten That Wasn’t” — a fake “Le Chaton Fat” Mistral model that fooled half the timeline, per Business Insider — is a free stress-test for your real dependencies. The lesson is the one three Kantone learned the hard way about data residency: if you cannot run it yourself when the vendor vanishes, it isn’t yours. Qwen3, one year on quietly proved the antidote: Apache-2.0 weights now run 235B-class capability on a sub-CHF-4,000 desk, keeping client data in-house.
Third, the Archicad thread tightened. “AI Inside Archicad” mapped the four layers any AI feature rides — model, API, codegen, inference — and told you to mark where client data crosses each arrow, pulling AI in through the Graphisoft-native Python API and PAZGPT so suggestions land as defensible native elements. It is the homework for the PAZ-Next Powersession on KI (24 June) where you put AI to work inside your own model. Bring a real project.
Fourth, accountability in code-checking. “Auditable BIM Compliance” showed SGR-BIM’s graph reasoning beating rule templates on geometry-heavy fire-code checks — but 84.3% means the human signer stays mandatory. Treat it as triage, demand a traversable audit trail, and negotiate an exit clause before you commit.
Signal vs. noise: The genuine signal was the convergence on auditability and ownership — articulation-point analysis (the symplectic “Cotangent Buildings” preprint), traversable compliance trails, open-weight fallbacks, build telemetry on Swiss metal via OpenTelemetry. The noise was agent-tool churn: Grok Build is one more capable terminal agent, but the durable asset is the plain-text convention (AGENTS.md, MCP) and the versioned git diff, not the model name on the box.
Hidden gem: Gerrymandle looked like a civics toy and was actually the week’s sharpest design lesson: gerrymandering and space planning are the same problem — partition a grid into equal, connected regions under a fairness rule. The same is_connected contiguity check validates electoral districts and your fire compartments. Quiet story, load-bearing idea.
Looking ahead: Watch the 24 June Powersession for how PAZGPT and the PAZ-BOX add-ons sit inside a live model, and watch whether the grid-as-bottleneck thesis starts showing up in cantonal data-centre procurement. The sovereignty stories are not slowing down.
Hack: This Hack teaches you to find the single point of failure in any project graph — the motif under half this week’s stories. The medium is runnable Python; the domain is graph analysis. Model each subsystem as a node and each dependency as an edge, then ask NetworkX for the articulation points: cut any one of them and the graph splits.
import networkx as nx
G = nx.Graph()
G.add_edges_from([("grid","cooling"),("cooling","racks"),
("racks","model_api"),("model_api","clash_check")])
print(sorted(nx.articulation_points(G))) # -> the nodes that, removed, break the building
Run it on one live project’s real dependencies. The node it returns that surprises you is the one the project was hiding — and likely the one that says “grid” or “single vendor.”
A federal closing thought from the late-2070s: federalism is a slowness budget, and Switzerland kept its sovereignty by spending it — three Kantone insisting on local data residency when nobody was watching. When your municipality procures software this year, read the data-residency clause yourself. Draw your own graph before you trust someone else’s.
Sources & Further Reading
PAZ Kaffi · multidisciplinary editorial, led by PAZ Academy