All Investigations / AI-GRID-CRUNCH-2026
PEAK
FILE · AI-GRID-CRUNCH-2026
75% · B+
AI FRONTIER · POWER GRID · 7 PRIMARY SOURCES · 1 CONTRADICTIONS
DEEPWIRE INVESTIGATION

The Gigawatt Bottleneck

AI's 2026 hard constraint isn't compute — it's the seven-year wait between gigawatt queue entry and gigawatt-on-the-bus. A forensic accounting of the load forecasts, the supply backlogs, and the workarounds.

The race to build frontier AI is no longer rate-limited by chips, capital, or talent. It is rate-limited by transformers, gas turbines, and the seven-year average it takes a queued project to actually energize. ERCOT just revised its 2030 data-center demand projection from 29 GW to 77 GW in a single planning cycle. Approval rate on the queue: under 2%. The Stargate, Colossus 2, and Anthropic-Google compute build-outs all sit on the same constraint set.

The queue numbers above carry the historical shadow-ratio caveat: only 15–25% of queued projects energize. Applied uniformly, the "effective" PJM data-center queue is closer to 33–55 GW than 220 GW. But hyperscaler-backed queue applications have a higher conversion rate than the historical average — these are not speculative renewables developers filing to flip permits.

PJM has 220 gigawatts of proposed projects waiting on a grid whose summer peak demand is 154.

Even if every queued data-center project had unlimited capital, three physical-supply constraints would still cap the rate of energization:

1. Heavy-duty gas turbines. GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries collectively produce ~80 large-frame gas turbines per year. 2027 delivery slots are largely sold; 2028 is the first cycle with material availability. Backlog clearance time: 24–36 months from order.

2. Large-power transformers (LPTs). Global LPT manufacturing capacity (Hitachi Energy, Siemens, GE, Hyundai) is tight. 765 kV and 500 kV class transformers carry lead times of 30–60 months. The US imports ~80% of its LPTs; tariff and supply-chain risk is non-trivial.

PJM, ERCOT, and the National Picture

3. Skilled trades. Lineworker, substation technician, and PE workforce contracted during 2007–2019 flat-load period. Re-staffing to 2025-2030 build requirements takes a full apprenticeship cycle (4 years minimum). This is the constraint nobody talks about loudly.

The dossier's load-bearing claim is "Grid energization constrains frontier AI training compute deployment through 2028." Three competing hypotheses; observables; diagnosticity scored 1 (consistent) to 5 (decisive against).

Reading: H2 (bottleneck moves) is the most-consistent hypothesis. H1 (binds) is close behind. H3 (dissolves) is essentially refuted by the LPT and turbine supply-side constraint. The headline forecast favors H2 — the bottleneck does not so much bind as redirect: AI compute deployment continues via BTM and nuclear restart, the public grid queue stays constipated, and the apparent "queue size" overstates the effective constraint on frontier compute.

Three Scenarios For The 2026-2029 Window

The US grid is unique in its degree of state-by-state regulatory fragmentation and queue-based interconnection process. Other major economies have different constraints and different policy levers — and the gaps between them are themselves strategic information.

The dossier's headline claim resolves on quarterly PJM/ERCOT planning reports. But six secondary questions require regulator filings, manufacturer orderbook data, and Federal Register publications the public discourse hasn't yet surfaced.

Falsifiable watchlist

PJM queue approval throughput remains below 5%/yr through 2027.
As of 2026-05-14
FALSIFIED IF
PJM's 2027 annual queue report shows approval throughput ≥5% of queued volume, OR cumulative 2026-27 GW energized exceeds 25 GW.
At least 3 additional nuclear restart PPAs (beyond TMI 1) signed with hyperscalers by end-2027.
As of 2026-05-14
FALSIFIED IF
No additional restart PPAs publicly signed by end-2027, OR existing restart sites (Palisades, Duane Arnold) cancel feasibility studies.
At least one frontier AI lab energizes ≥2 GW of dedicated compute via BTM generation by end-2027.
As of 2026-05-14
FALSIFIED IF
No frontier lab announces ≥2 GW BTM energization by end-2027, OR Stargate Abilene + Colossus 2 + Anthropic-Google TPU sum to less than 4 GW operational by Jan 2028.
Primary sources
01ferc.govwww.ferc.gov02ferc.govwww.ferc.gov03datacenterfrontier.comwww.datacenterfrontier.com04datacenterknowledge.comwww.datacenterknowledge.com05brief.bismarckanalysis.combrief.bismarckanalysis.com06enkiai.comenkiai.com07whitecase.comwww.whitecase.com08interestingengineering.cominterestingengineering.com
091Federal Energy Regulatory Commission — Fact Sheet: PJM Co-Located Large Load Directive T1 primary — ferc.govcited source
102FERC — PJM Innovation and Consumer Protection Order T1 primary — ferc.govcited source
113Data Center Frontier — "The Gigawatt Bottleneck: Power Constraints Define AI Data Center Growth," May 2026 T2 — datacenterfrontier.comcited source
124Data Center Knowledge — "Why AI Data Center Projects Face Years of Delays After Approval," 2026 T2 — datacenterknowledge.comcited source
135Bismarck Analysis Brief — "AI 2026: Data Centers Restart Growth of a Stagnant US Electrical Grid" T2 — brief.bismarckanalysis.comcited source
146EnkiAI — "US Grid Congestion 2026: PJM's Data Center Crisis" T2 — enkiai.comcited source
SEE THIS ON THE NETWORK
See what other investigations share NVIDIA · OpenAI · xAI. Substrate Finder will surface every dossier these entities also appear in.