the seam

1 · Pre-Development / Feasibility

Pricing the risk behind the data-center boom, before the ground is bought

Azraq, an Abu Dhabi startup founded by former Arup engineer Alexandra Coleman, closed an oversubscribed pre-seed led by A-typical Ventures to underwrite the risk of large infrastructure projects — data centers first. Its platform runs Monte Carlo simulations across six dimensions (market demand, environmental exposure, infrastructure reliability, labor constraints, regulatory complexity, covenant performance) to produce lender-grade outputs — value-at-risk, debt-service-coverage distributions, covenant-breach probability — in place of consultant reports that can run to $1 million. The backdrop: U.S. data-center credit deals hit $182 billion in 2025, and 57% of projects slip three months or more, at roughly $14.2 million a month on a 60 MW build. For AEC firms, it’s a reminder that the feasibility conversation increasingly happens inside a financial model before a designer is in the room. FWDstart, Jul 10 →

2 · Pre-Design

What developers are actually testing: 123,000 feasibility runs, read as a market signal

The real-estate newsletter Thesis Driven analyzed roughly 123,000 site feasibility studies run on TestFit last quarter — not a product pitch, but the aggregate read as a signal of what gets tested before design begins. Retail’s share of feasibility runs climbed from 33.5% (Q3 2024) to 42.9%, overtaking higher-density multifamily, which fell from 37.3% to 28.5%. Single-family subdivision studies nearly tripled, to 8.6% of residential runs, concentrated in Phoenix and Washington, D.C. Median multifamily density dropped from 90.2 to 75.9 dwelling units per acre. And the clock stretched: a feasibility run now spans 56–59 days start to finish, against under three in late 2024 — deals aren’t dying so much as being shelved and re-penciled, over and over. Thesis Driven, Jul 9 →

3 · Design

AI meets the door schedule — specification’s least glamorous, most exposed corner

A trade piece in PBC Today puts AI where specification actually hurts: architectural ironmongery — the door hardware, closers, and locks whose schedules carry dense performance data and fire, accessibility, and security certifications. Consort’s Daniel May argues the clearest value is right here, rapidly processing certification data and flagging missing or inconsistent compliance information across schedules — the document-heavy work that eats specifiers’ hours. Context: RIBA’s 2025 report finds 59% of practices now use AI on at least some projects, up from 41% a year earlier. But the caution is pointed and post-Grenfell: AI can process the data, yet it isn’t accountable for the decision — traceability and human sign-off stay central where a mis-specified fire door has consequences. PBC Today, Jul 13 →

4 · Construction

The money says: put the budget in agents, not robots

Writing in Construction Executive, Grace Calengor reads a venture-funding chart (F-Prime Capital, surfaced by Social Capital) worth every contractor’s attention: in 2025, applied robotics — humanoids and jobsite machines — drew roughly $18–21 billion, while total AI funding topped $211 billion. The point isn’t that robots don’t matter; it’s that the deployable capacity is in the back office now — agents that read, route, approve, and reconcile pay apps, change orders, and submittals. Her reframing: don’t ask whether an agent replaces an estimator, ask whether it frees nine hours a week per person. She leans on the same March 2026 finding the Anthropic index made famous — large theoretical capability, thin observed use. The gap isn’t the model. It’s deployment. Construction Executive, Jul 13 →

5 · Post-Occupancy

Retro-commissioning gets an assistant that reads a million trend-log rows

In retrofit, commissioning engineers Mark Gelfo and Gary Haydu make the operations case with unusual specificity. In existing buildings — “massive repositories of unstructured data” — AI can ingest multi-year building-automation trend logs to isolate performance drift, catch the inefficiencies that never trip an alarm (simultaneous heating and cooling, hunting VFD loops), and cut documentation from days to minutes. Their sharpest example: during a chiller-plant retrofit, an AI review flagged a mismatch between a specified 4–20 mA control signal and a submitted 0–10 VDC one, heading off field rework. The frame is ‘trust but verify’ — the tool parses the data in milliseconds, but the human still walks the plant and carries the liability for the sign-off. retrofit, Jul 13 →