2026 06 25 Feed Semianalysis US Grid Constraints 40gw Behind THE Meter Datacenter
The binding constraint on AI datacenter buildout is the grid, not chips or capital — interconnection headroom 'turns negative by 2027', so behind-the-meter (BTM) generation powers 'well over half of new US datacenters' by 2028, a ~50GW/yr TAM routing capex to gas-turbine, fuel-cell, RICE and nuclear-co-location suppliers (GE Vernova, Bloom Energy, Vistra, Constellation).
view source ↗Summary
SemiAnalysis argues the binding constraint on the AI datacenter buildout is US grid interconnection capacity — not chip supply or financing — and that available grid headroom "is already approaching zero and turns negative by 2027." Because gas turbines and step-up transformers now carry 3–4 year lead times (vs. an 18-month historical norm) and interconnection queues stall, datacenter operators are routing around the grid with behind-the-meter (BTM) generation: BTM "will power well over half of new US datacenters" by 2028, a ~50GW/year equipment TAM by 2029. The capex flows to gas-turbine OEMs (GE Vernova dominant), fuel cells (Bloom Energy), reciprocating engines/RICE (INNIO, Wärtsilä, Caterpillar/Cummins), and nuclear co-location (Vistra Comanche Peak + AWS; Constellation/Calpine Freestone). This directly strengthens the ai-capex-to-power-and-materials-cascade chain (power is the binding constraint) and nuclear-baseload-for-ai-data-centers, and opens a 2nd-order BTM-equipment chain (GE Vernova / Bloom Energy) the wiki has not yet traced.
Article
Readable extraction of the SemiAnalysis post; quoted passages preserved verbatim, attributed to SemiAnalysis's Energy Model. Cite: source URL above.
Core argument — the grid is the binding constraint. Per SemiAnalysis's Energy Model, the grid faces a structural shortfall: "available headroom is already approaching zero and turns negative by 2027." Three bottlenecks prevent grid-connected growth:
- Generation delays — "Gas turbine and generator step-up transformer lead times have each extended to three to four years, versus a historical norm of roughly 18 months." "The bulk of gas capacity ordered by utilities is made up of CCGTs and combustion turbines," where "construction and time from planning to COD run 4–6 years."
- Interconnection queues — PJM has "57GW cleared studies" but "24GW of projects with fully executed agreements terminated before reaching commercial operation" since 2020.
- Transmission constraints — utilities cannot deliver promised load timelines; datacenter operators post "substantial letters of credit" and "take-or-pay commitments" with minimal penalties for utility delays.
Projected behind-the-meter growth. "BTM will power well over half of new US datacenters" by 2028; the 2029 BTM-equipment TAM is "50GW/year." Datacenter demand grows from "+21GW in 2026 to +84GW by 2030," but the grid supplies only "15GW of net-new ELCC capacity annually" (rising "towards 20GW+ by end of decade") — a gap widening ~20–30GW/year by 2028–2030.
Generation technologies filling the gap:
- Gas turbines — GE Vernova turbines dominate announced projects. Example: "Crusoe also has a TCEQ air-permit filing for up to 933 MW of gross nameplate gas generation, of which we expect ~665 MW is the 19x GE Vernova LM2500 turbines."
- Faster alternatives — "Faster technologies like Fuel Cells and RICE are being aggressively secured directly by datacenter operators for onsite generation" (reciprocating internal-combustion engines).
- Nuclear co-location — AWS signed a 1,200 MW net-metering arrangement at Vistra's Comanche Peak nuclear plant.
- Solar + storage limitations — renewables suffer "declining, weather-dependent value"; solar ELCC credited "essentially zero" in congested ERCOT zones; 4-hour battery storage has "little marginal ELCC" as grid risk shifts beyond 4-hour durations.
Named beneficiaries. Equipment: GE Vernova (gas turbines), Bloom Energy (fuel cells), Bergen Engines / Wärtsilä (alternate suppliers), INNIO (reciprocating engines), transformer/breaker OEMs (3–4yr lead times). Power operators: Vistra (Comanche Peak nuclear co-location), Calpine (Thad Hill, Freestone), Constellation Energy (Freestone partner), NRG (IPP sector). Datacenter developers: Crusoe Energy, AWS, CyrusOne, Meta (self-built, skipping grid backup).
Supply constraints. Turbine + step-up transformer lead times at 3–4 years. SemiAnalysis notes a transient relief: "we were the first to flag the risk of temporary peak gas turbine orders," and "surging turbine availability" in secondary markets — but frames it as temporary before the 2028+ demand acceleration. Main power transformers (MPTs) and HV breakers are critical-path items utilities cannot expedite.
Regulatory framework (ERCOT "Batch Zero", effective July 2026). Texas codified hybrid co-location structures: NMA (Net-Metering Arrangements — existing generators co-locate with datacenters), BYOG (Bring Your Own Generation — new generation built alongside load), WLPUN (Withdrawal-Limited Private Use Network — oversized load caps grid withdrawal, e.g. 1,000 MW load capped to 100 MW grid draw). Early co-location announced: ~2,885 MW across Crusoe, AWS, CyrusOne.
Key metrics. PJM 2027/2028 auction cleared "134,478MW" against a "20% target," a "6,517MW UCAP deficit." Early co-location ~2,885 MW across the named developers.