AI data centres are being held up not by chips, and not by making electricity, but by the unglamorous kit that moves power around: transformers, switchgear and grid connections.
68%likely
Well sourcedNext check
Watching
Why it is different
The forecast isolates grid gear and interconnection from the broader, already familiar claim that AI is power-constrained.
Why it matters
Long lead times for transformers, switchgear and connections can cap how quickly new AI capacity reaches service.
Next check · 31 Oct 2026
On its next earnings call, does Microsoft again say AI capacity is limited by power and buildings rather than chip supply?
How we scored it ·
from how similar cases went
If right, how big5/5How overlooked3/5Upside vs downside2/5Source trust4/5Independent sources4/5
The exact wording we score
At 31 Dec 2029, at least 33% of US AI data-center capacity that was publicly announced for energisation in 2026–2029 has either been cancelled or missed its original energisation date by more than 12 months, and contemporaneous developer/utility filings attribute the primary cause for a majority of that delayed capacity to transformers, switchgear or interconnection rather than chip supply. Resolves YES only if both legs hold; NO otherwise. The announced-capacity denominator and each project's original date are frozen from public records available by 31 Dec 2026; later announcements cannot dilute the denominator.
PROOF
Evidence and provenance
The links below are the evidence recorded when this forecast was scored. Each note states source quality or bias; a citation is not an endorsement.
Recorded sources · 4 citations / 4 independent origins
2026-07-28 | sell-side: generic AI-power constraint is widely owned; mainstream press: power scarcity is consensus; priced market: electrical-equipment rerating already reflects part of the story; specialist: grid and interconnection evidence still isolates transformers and switchgear as the less-priced bottleneck | E3
Revision history4 entries
Probabilities and interpretations are never silently overwritten. This is the append-only record of what changed and why.
Resolution rule repaired. The prior wording said “roughly a third to a half” but did not freeze a denominator and called the signal falsified only above 75% on-time delivery, leaving a 25–33% dead zone. The claim now uses a frozen announced-capacity cohort, an exact 33% delay/cancellation threshold, a separate cause-attribution leg, and exhaustive YES/NO outcomes. claimtype is corrected from atomic to conjunction; P=68 is now the explicit 0.80 × 0.85 product.
(2nd pass) — Impact 4→5. Against the domain-neutral anchors, 'roughly a third to a half of announced US AI data-center capacity 2026-2029 slips or is cancelled' is unambiguously the $1T+ capital-pool tier. It had been scored below china-electron-gap (I=5) for a strictly larger consequence. Inherited the NERC LTRA and PJM BRA checkpoints from the china signal's split. Two coupling warnings added to the indicators: (a) the Microsoft power-vs-chips sentence resolves gpu-collateralized-debt in the OPPOSITE direction — they are anti-correlated, not correlated; (b) this signal's GE Vernova backlog checkpoint and turbine-chokepoint's resolve off the SAME filing on the same date, so they are one observation and must not be counted as two independent corroborations. corrgroup refined ai-capex → ai-capex-physical.
v2 rescore, P held at 68 on conflicting evidence. Supporting: GE Vernova Q2-2026 (2026-07-22) put Electrification equipment backlog at $41B, +69% YoY, orders +66% at 1.7x book-to-bill across substations, switchgear and transformers. Weakening: the BLS transformer PPI has been essentially flat since Oct-2025 (362.6 Nov-2025 → 365.3 Jun-2026, ~+4.2% YoY with no acceleration) and LBNL's Queued Up: 2026 Edition (Jun 2026) reports early improvement in interconnection queues. The price signal is the first crack in this thesis — hence the PPI indicator above.
created from research sweep; edge kept at 3 (generic power thesis now mainstream). Top source (IEA) spot-verified.
Method challenge
Adversarial review record
Conditionally admitted
Structured internal role review tied to this frozen claim. It is not independent human peer review. Independent specialist review not yet performed
technical-domain specialist
delivery and operations reviewer
policy or standards reviewer
forecasting-method reviewer
Strongest specialist challenge
The strongest objection is not that projects will avoid delay but that the forecast attributes a majority of delayed gigawatts to one equipment family. Grid connection, generation, land, financing, permitting, load contracts, and chip delivery are often jointly binding.
Outside view
Large infrastructure projects frequently slip, so the delay leg has a high outside-view probability. The narrower reference class is public megaproject delay with a reproducible primary-cause code; that second leg is materially less certain than the first.
Causal rival
Demand revision or financing can cancel a project that also lacks transformers, while behind-the-meter generation can bypass interconnection without fixing gear supply. A decline in AI economics would produce the same delayed-capacity observation through a different mechanism.
Measurement risk
There is no current public frozen list of the denominator, original dates, and capacity weights. Developer explanations can be strategic and multi-causal, so cause attribution needs a predeclared hierarchy and archived contemporaneous documents rather than retrospective narrative coding.
Residual risk
Even with a frozen cohort, private schedule changes and missing utility filings can create non-random evidence gaps. The condition is therefore release-critical: without the cohort manifest the exact 33 percent and majority-attribution legs are not independently reproducible.
Admission condition
Publish and archive the project-level 2026 announced-capacity cohort, original energisation dates, capacity denominator, and attribution coding rules by 2026-12-31; otherwise void the parent call.
COLDTELL