The Intelligent Building Brief — Issue №3
The Intelligent Building Brief
What “Verified” Actually Looks Like
Issue №3 · June 14, 2026 · ~10-minute read
Intelligence for the people who own, operate, finance, and build commercial real estate. Capital markets, the energy/AI shift underneath all of it, and what’s signal versus noise. Every number sourced.
From the Editor
Last issue we made the uncomfortable case that the energy-savings numbers being sold to you are systematically overstated. That was only half a thesis. This week the other half arrived — and it’s the constructive half. A peer-reviewed, ten-building, three-year study, independently validated by Singapore’s Building & Construction Authority, shows what rigorous verification actually looks like: whole-building, continuously-updated dynamic baselines, prediction error under 5%. Put the two papers side by side and you get the sentence the whole industry is circling: zone-scale demos overstate; whole-building, dynamic-baseline, third-party-audited M&V is how you actually verify. Meanwhile the money got bigger and the power got tighter — four hyperscalers now guide to $725B of 2026 capex. In a market where the megawatt is the scarce asset, “verified” stops being an ESG nicety and becomes the asset.
— The Editor · AISB Intelligence Desk

1. The Lead — we finally have a picture of “verified”
The most useful finding this week is the constructive answer to last week’s warning. A new peer-reviewed study — “AI-enabled energy baselines for verified building decarbonization” (Scientific Reports, Nature portfolio, Jan 2026) — pairs an LSTM + XGBoost model with Singapore’s Energy Conservation Calculation method to build continuously-updated dynamic baselines across ten commercial, residential and mixed-use buildings over three years (2021–2024). The results that matter: baseline prediction error generally under 5% (avg 4.9%), 3,221 tCO₂e of cumulative reductions, a retrofitted hotel running 56–61% better than its 2005 benchmark — and, the part that makes it credible, outcomes independently confirmed by the Singapore Building & Construction Authority and cross-checked against utility-meter data. Read it next to February’s overestimation paper (arXiv:2602.22499) and the thesis is complete: stop buying the headline percentage; buy the method — whole-building, dynamic-baseline, third-party-auditable.
2. The Numbers
$725B — combined 2026 hyperscaler capex, +77% YoY (Amazon ~$200B · Alphabet ~$190B · Microsoft ~$190B · Meta $115–135B); ~75% AI ≈ $450B. (Tom’s Hardware; CreditSights; CNBC)
$80B — Microsoft’s Azure backlog it cannot fulfil for lack of power — the cleanest proof that the binding constraint is megawatts, not demand. (Microsoft; CNBC)
<5% — baseline-prediction error (avg 4.9%) in the BCA-validated 10-building, 3-year whole-building M&V study. (Scientific Reports, PMC12894995)
43,010 MW — Taipower’s 2026 peak-load projection; new demand to exceed 5 GW by 2030 (~1 GW/yr) from foundry + AI-server build-out. (Digitimes; Taipower)
76.1M sq ft — projected 2026 US data-center completions, more than 2025’s 33.5M — supply chasing demand into a power-limited grid. (CoStar)
T-17 days — to Germany’s EnEfG ERF ≥10% waste-heat-reuse gate (Jul 1 2026, data centers ≥300 kW). (White & Case)
📊 The Two-Sided M&V Thesis
Two 2026 papers, one conclusion: how you measure decides whether the number is real
Zone-only demos → OVERSTATE
Control a few zones, report savings for those zones; interior-only demos can show savings that are ~100% fictitious.
arXiv:2602.22499 (Feb 2026)
Whole-building dynamic baseline → VERIFIED
10 buildings, 3 yrs, error <5%, BCA-validated against meters. This is what auditable looks like.
Scientific Reports (Jan 2026)
Editor’s Take  “BCA-validated” is the phrase to steal. When a vendor quotes a savings number, ask who audited the baseline. If the answer is “we did,” it’s marketing.
4. Capital Markets — the money got bigger; the asset class shifted
Four hyperscalers now guide to $725B of 2026 capex (+77% YoY), roughly $450B of it AI-related. Microsoft’s CFO attributes ~$25B of its own increase to rising memory and component costs — even the inflation is partly scarcity. The capital-markets story underneath: data centers are becoming a financing asset class. US completions are set to more than double to 76.1M sq ft in 2026 (CoStar), and stabilized hyperscale facilities are increasingly funded through single-asset, single-borrower CMBS. The legacy fracture line is unchanged — office still carries a record 12.34% CMBS delinquency against multifamily’s 0.5% — but the new money is being built, and financed, around power.
📊 The Two-Speed Market — one market, opposite outcomes
CMBS delinquency on 2026 maturities — office vs multifamily
Office
12.34%
Multifamily
0.5%
A ~25× gap — bars drawn to scale. (S&P Global)
📊 The $725B Year — 2026 hyperscaler capex
Up 77% YoY — ~75% AI-related (~$450B). Bars to scale.
Amazon
~$200B
Alphabet
~$190B
Microsoft
~$190B
Meta
$115–135B
Full-year company guidance + analyst aggregation. (Tom’s Hardware; Statista; CreditSights)
Editor’s Take  $725B is a real-estate-and-power number wearing a tech headline. A large share lands as buildings, substations, and cooling — the stack our industry owns.
6. Sector Watch
Data centers stay the asset class of the cycle — but the binding constraint is electricity, not chips, and completions are now outrunning the grid’s ability to power them. Office remains two assets sharing a name: trophy tightening, commodity in runoff at 12.34% delinquency. Quiet winners: multifamily (0.5%), industrial, medical office — rolling over with a haircut, not a default.
📊 Sector Scorecard — where the cycle stands
Data centers HOT power-constrained, asset of the cycle
Multifamily FIRM 0.5% delinquency — quiet winner
Industrial FIRM rolling over with a haircut, not a default
Office — trophy MIXED tightening — a buy at a reset basis
Office — commodity RUNOFF 12.34% delinquency — repricing
Taipei 101 at night — Taiwan's grid is the world's most legible 'megawatt is the scarce asset' case
Taipei after dark — Taiwan’s 2026 peak-load projection (43,010 MW) is climbing ~1 GW/yr against a thin reserve margin.
7. The Energy Constraint — the megawatt is the scarce asset
Microsoft’s $80B of Azure demand it cannot serve for lack of power is the cycle’s defining sentence. The slow lever — new generation — doesn’t clear fast: gas-turbine lead times run 7–8 years. The fast lever is verified demand-side efficiency, which is exactly why Section 1 matters. Nowhere is this sharper than Taiwan: Taipower projects a 43,010 MW 2026 peak and ~1 GW/yr of new demand to 2030 from foundry, memory and AI-server build-out, with summer tariffs already live since June 1. In a grid that tight, every rigorously-verified saved kWh is a load-forecast-integrity issue for the operator — not an ESG line item.
📊 US Data-Center Completions — supply more than doubles
2025
33.5M sq ft
2026
76.1M sq ft
2027
73.4M sq ft
Projected US completions; bars to scale. (CoStar)
Editor’s Take  You can finance and pour 76M sq ft. You cannot pour a megawatt. Completions doubling into a power-limited grid is why the demand-side lever is the one that clears.
9. The Regulatory Clock
🇩🇪 Germany EnEfG — T-17 days to the Jul 1 2026 ERF ≥10% gate (data centers ≥300 kW), stepping to 15% in 2027 and 20% in 2028; fines reach €100k. Verified energy reuse becomes a build/permit prerequisite, not a virtue.

🌎 SBTi V2.0 remains unpublished and the date guidance has widened again — advisories now span “early” to “late 2026.” V1.3 stays valid for new targets through Dec 31 2027; V2.0 becomes mandatory Jan 1 2028. Lowered urgency, live watch.
10. The PropTech Lens — the twins are getting good (inside the wall)
On June 2, Vertiv shipped the first converged physical-infrastructure digital twin for NVIDIA’s Omniverse DSX (its SmartRun system, demoed at Computex Taipei) — power, thermal and operational infrastructure designed and validated as one model before build-out. It joins a deepening roster (Schneider, Siemens, Eaton, Cadence, PTC, et al.). Real capability — but read where it points: every public scope marker stays inside a single AI factory, one owner, one plant. The integrated-FM consensus (“AI as operator, not assistant”) is a tailwind; whole-building, auditable M&V is still the differentiator riding it.
11. The Playbook — how to spec “verified”
1. Require a whole-building baseline, weather-normalized — not a zone-level extrapolation.
2. Demand a dynamic, continuously-updated baseline (the Scientific Reports method), not a static pre/post snapshot.
3. Ask who audited it. Third-party / authority validation (BCA-grade) beats a vendor’s own dashboard, every time.
4. Hold the model to a stated accuracy bar (the study ran <5% baseline error). No stated error = no claim.
5. Capitalize the verified savings only — underwrite the proof, not the pitch. (Robin — CV(RMSE)/NMBE thresholds flagged for your sign-off.)
Editor’s Take  Notice the APAC tell: the reference design for “verified” came out of Singapore, audited by a national authority. The M&V standard-setter this cycle isn’t US or EU.
12. Contrarian Corner — the best CRE-tech ground is still above the wall
Every new DSX twin — Vertiv this month — makes the same bet: orchestrate brilliantly inside one gigawatt-scale, single-owner AI factory. None of them crosses the wall. The defensible altitude no one is taking is the cross-owner, cross-vendor decision layer above N independently-owned, heterogeneous, mostly-legacy buildings — the commercial stock, not the new-build GPU barns. The more the in-factory layer commoditizes, the more valuable the above-the-wall layer becomes. The smartest 2026 bet isn’t inside the AI factory; it’s the portfolio layer over the buildings that already exist.
Multi-level highway interchange at night — the layer above the independent roadways is where the value routes
The value isn’t in any one roadway — it’s in the layer that routes across all of them.
13. Tools & The Signal
IPMVP (EVO) — the M&V framework behind the Playbook, free. Pair it with a dynamic-baseline method and a third-party audit and you have the Section 1 reference design.

Singapore BCA Green Mark — worth reading even outside APAC; it’s where the auditable-baseline bar is being set this cycle.

The Signal (track, not yet urgent): “verified” is hardening from adjective into artifact — the thing you must produce to transact, permit, or refinance.
📅 The Lookahead — dates that move the market
Jul 1, 2026Germany EnEfG ERF ≥10% gate enters force (≥300 kW DCs).
Summer 2026Taiwan peak-load test vs the 43,010 MW projection (tariffs live since Jun 1).
Aug 29, 2026NYC LL97 final filing deadline — first penalty exposure crystallizes.
2027Debt-maturity wall peaks at $1.26T; EnEfG ERF steps to 15%.
Jan 1, 2028SBTi V2.0 becomes mandatory (3,000+ companies); EnEfG ERF →20%.
Our Conviction Calls
Four calls we’d put our name on this week. These are opinions, not facts. Hold us to them.
[HIGH] “Verified” is becoming the asset. Whole-building, dynamic-baseline, third-party-audited M&V is now an evidenced standard, not a pitch — and the spread between owners who can prove performance and those who only claim it is where 2026’s value sits.

[HIGH] Power access beats location in data-center-adjacent CRE. Completions doubling into a power-limited grid says underwrite the interconnection, not the dirt.

[MED-HIGH] The cross-owner layer stays open. Every new in-factory DSX twin makes the above-the-wall, legacy-stock decision layer more valuable, not less.

[MEDIUM] APAC is setting the M&V bar. The reference design for auditable savings came out of Singapore this cycle — watch the standard travel.
Where we might be wrong: the Scientific Reports study validates baseline-prediction accuracy + authority-confirmed decarbonization — not a randomized savings-attribution. If an independent randomized trial fails to reproduce the savings at whole-building scale, our “verified” conviction softens. We’re watching for exactly that.
The Intelligent Building Brief is free, and that’s the point. If it earned ten minutes of your week, forward it to one colleague who’d say the same. That’s the entire growth plan.
— ai-smart-buildings.com