Three data-center decisions in a single week are turning community entitlement — not power, water, or capital — into the binding constraint on where AI physically gets built. In Palm Beach County, commissioners were scheduled to vote this morning on “Project Tango,” a roughly one-million-square-foot hyperscale campus (opponents estimate up to ~600 MW) on a 202-acre Loxahatchee site; the county's own zoning commission had voted 6–0 to urge denial, breaking with planning staff, and a judge on Tuesday refused to block the hearing. No decision was public at press time. Near Atlanta, Coweta County approved Prologis's 900 MW, 4.34-million-square-foot “Project Sail” the same week. And in Birmingham, the Nebius / Oxmoor Valley injunction hearing ran without a ruling and resumes Thursday, the defense now calling the city's own permitting director to the stand. For the firms who design and build these campuses, entitlement risk — not feasibility — is becoming the critical path.

A peer-reviewed study from Poznań University of Technology — published in late June and, as far as we can tell, unnoticed by the trade press — put the commercial space-planning tool PlanFinder against nine real, built, inhabited Polish apartments. The finding is a clean split: the AI hit 100% on macro spatial logic (daylight, minimum room sizes, sensible public/private zoning) but showed “zero compliance” on a raft of mandatory details — no space designated for washing machines or storage, a missing bathroom sink, walk-through “enfilade” rooms, and no awareness of internal load-bearing walls or MEP. It is the same shape we keep seeing this month: fluent at the legible, blind at the code-bound. Source material for a capacity study, not a set of drawings.

Nothing in the BIM, drawing-production, or specification space cleared the bar today: the freshest substantive items are a month old, and everything fresh was vendor advertorial. Specification writing — the task most exposed to language models — still has no dedicated tool or benchmark to point to. An honest gap beats filler. The real design-side movement remains the agentic-BIM push we've tracked all week, and it is still waiting on clean, standardized model data to be worth much.

Guthrie AI raised a $4 million seed led by Chicago Ventures this week to scale what it calls “Virtual Bid Assistants” for trade subcontractors, starting with glaziers. The pitch is deliberately unglamorous: software that gathers drawings, routes scope, chases vendor pricing and assembles proposals, paired with a human bid coordinator — supplementing the estimator, not replacing them. The company says 52 glaziers and vendors have signed on, cutting bid turnaround from weeks to days and bidding ~70% more work without adding staff, at $4,125/month. The most telling detail is the restraint. As one skeptical write-up put it, this is “outsourced estimating support with an AI wrapper” — and its refusal to promise an autonomous end-to-end agent is exactly what makes it read as credible rather than mythological.

Noda, a Washington, D.C. startup led by CEO Kate Henningsen, launched what it bills as the first agentic AI for commercial building operations — a system that doesn't just flag problems for staff to chase, but finds, fixes and verifies operational issues across a portfolio around the clock. Noda claims coverage of 1,200 buildings and 350 million square feet, and reports — self-reported, against Salesforce baselines — that it eliminates ~80% of manual workflows, cuts issue-to-resolution time ~40%, and drives 0.5–2% NOI growth. Take the performance figures with the usual vendor-metric caution. The more interesting number is the labor one: with roughly 15% of building-operations roles getting filled, Henningsen frames this less as AI displacing engineers than as software covering a gap the workforce no longer can.