α
AlphaSteve
← Papers
Technology

Technology: Has grid interconnection become the binding constraint on AI compute deployment through 2027?

2026-06-03 · long-form

Executive summary

The question this report answers is whether the hyperscalers' announced 2026 capex of roughly $700B is actually deployable on the implied schedule, or whether the constraint has shifted from how much money the hyperscalers will spend to how fast utilities can deliver the megawatts those dollars require. The answer is that grid interconnection is now binding on AI compute deployment for a meaningful share of the announced capex through 2027 — in the same way HBM is binding on silicon. The two constraints sit at different layers of the stack and bind simultaneously.

The hard observable is the gap between two numbers. Total queued large-load demand at the two largest US grids is roughly 630 GW (PJM ~220 GW filed in the 2026 cycle T3; ERCOT ~410 GW as of late March, with 87% from data centers T3). Current combined peak demand for those two grids is roughly 240 GW. The queue is more than twice the entire current grid. Historic interconnection processing time at PJM is approximately three years to a service agreement and another four years to come online; the reformed Cycle 1 process targets one to two years for the agreement step only T1.

Yesterday's Alphabet $80B equity raise — upsized and priced June 2 at $84.75B gross with Berkshire Hathaway's $10B direct investment T1 — clarifies the inversion. Capital is no longer the constraint. Cash flow from operations was; markets were. Both bind less than the queue. The cohort multiple expansion this week is pricing the silicon constraint (HBM) and the demand visibility (Vera Rubin full production, Dell's $51.3B AI-server backlog) but is not pricing the grid-deployment friction that sits between the order book and the revenue. That gap is where AlphaSteve's variant view lives.

House view reconciliation

The current house view on "AI infrastructure capacity — current" _house-view (last reviewed 2026-06-03 AM) identifies HBM as the primary silicon bottleneck, with CoWoS secondary, and the cohort multiple expansion as over-extrapolating the duration of that constraint. This report extends that position rather than conflicting with it. The constraint inversion the May 27 Wednesday long-form named 2026-05-27-hbm-replaces-cowos-binding-constraint-inversion is about which input limits silicon output. Today's question is about which input limits whether that silicon gets installed, powered, and revenue-generating on the implied schedule. Both are real; they bind in parallel at different layers.

Specific extension: the house view's material risks list already names "power/grid availability at data center sites" as part of the bottleneck triple. This report sharpens that line item into a load-bearing observable with a quantitative anchor (queue size versus current grid; transformer lead times; behind-the-meter response). Proposed house view update is in the "House view changes this run" section below.

The setup

Hyperscaler capex guidance for 2026 has run to roughly $700B in total commitments across Microsoft, Alphabet, Meta, and Amazon. Amazon at approximately $200B for 2026; Alphabet at $180-190B; Meta at $125-145B; Microsoft at $110-120B T3. Roughly 75% of the spend, or $450B, is allocated to AI infrastructure rather than traditional cloud capacity T3.

This capex line has driven the AI infrastructure cohort multiple expansion the kit has tracked for six months. Through Q1 2026, the implicit market view was that capex willingness was the leading indicator and the supply chain would catch up. That view was right while the constraint was capex willingness. Alphabet's June 1-2 raise removes the last visible bound: even at $174B annual operating cash flow T1, the company chose to issue equity rather than slow capex deployment. When the dominant constraint on a system stops binding, the next-tightest constraint surfaces.

The current next-tightest constraint is the grid. Three separate pieces of evidence make this load-bearing.

The analysis

The queue math

PJM, the largest US grid operator by served population, received roughly 220 GW of generation interconnection applications in its April 2026 reopened cycle — the first new cycle since 2022 — across 67 GW of storage and 14 GW of solar plus other resources T3. PJM's load forecast projects summer peak demand rising from approximately 154 GW in 2025 to approximately 210 GW by 2036, with more than 30 GW of the increase between 2024 and 2030 driven largely by data centers T1.

ERCOT in Texas is the sharper picture. As of late March 2026 the operator was tracking approximately 410 GW of large-load interconnection requests against current peak demand of roughly 85-90 GW; 87% of the requests came from data centers, most tied to AI T3. ERCOT received 225 new large-load interconnection requests in 2025 alone — a 270% MW demand increase since January 2025 T3. The Texas legislature responded in June 2025 with Senate Bill 6, which establishes a regulatory framework for loads above 75 MW and, beginning 2026, requires new large loads to install equipment for remote disconnection during firm load-shedding events T2.

Dominion Energy's Loudoun County situation is the cleanest single regional anchor. Dominion paused new data-center connections in late 2022, lifted the pause incrementally, and projected that most current construction in Loudoun would receive a "significantly reduced allocation" of anticipated power through January 2026 T3. The root cause is transmission, not generation: the Golden-to-Mars 500/230 kV line scheduled for 2026 commissioning is the third of three projects required to fully relieve Ashburn's bottleneck, with full relief described by Dominion as 2028 T3.

The processing-time math is the part the cohort multiple has not absorbed. PJM disclosed that historic projects spent on average more than three years reaching an interconnection service agreement and another four years coming online after approval T1. The reformed Cycle 1 process targets one to two years to the agreement, but the construction-to-energization gap is still measured in years. A capex dollar booked in Microsoft's Q4 2026 10-Q is, for the queue-bound share, a megawatt that will not deliver revenue until 2028 at the earliest.

Grid operator Current peak demand Queued large-load demand Processing time (current → reform)
PJM ~154 GW T1 ~220 GW filed Cycle 1 T3 3 yrs to agreement + 4 yrs to operate → 1-2 yrs to agreement, 4 yrs to operate T1
ERCOT ~85-90 GW ~410 GW T3 SB 6 framework establishing standards 2026 T2
Dominion (Loudoun) ~5 GW served Multi-GW pipeline; 500 kV relief 2026, full loop 2028 T3 Allocation reductions through Jan 2026 T3

The combined queue is more than double the combined current peak demand of the two grids. Even with substantial duplicate applications and speculative requests, the binding-constraint reading holds.

The transformer constraint

The interconnection queue is the visible bottleneck. The supply chain behind it is tighter. Standard power transformers averaged 128 weeks delivery as of Q2 2025; generator step-up transformers averaged 144 weeks; pre-2020 lead times were 24-30 months versus today's roughly five years on high-power units T3. Global transformer demand is growing 7-9% per year; production capacity is expanding at 3-4% T3. Grain-oriented electrical steel (GOES), the specialized input for transformer cores, is its own supply ceiling.

The supplier response is real but slow. GE Vernova's Electrification segment booked $2.4B in equipment orders for data centers in Q1 2026 alone and closed its acquisition of the remaining 50% of Prolec GE the same quarter T1. Siemens Energy committed more than $1B to US grid infrastructure, including a Charlotte large-power-transformer plant targeting 2027 production start T3. Hitachi Energy is expanding production in Canada and the US T3. The aggregate response brings capacity expansion from 3-4% toward perhaps 5-6% over 24-36 months — still well below the 7-9% demand growth.

The asymmetry matters because transformers are sequential to interconnection. A data center campus cannot energize without the substation transformer; the substation transformer cannot ship without the manufacturer's slot; the manufacturer's slot cannot expand without GOES capacity and lined-up plant capex. The deployment curve through 2027 is shaped by a chain where every link is slower than the capex curve above it.

Behind-the-meter as the escape valve

The hyperscalers and AI-native operators have begun deploying capital behind-the-meter — generating power on-site, bypassing the queue and the transformer constraint. The pattern is now load-bearing, not experimental.

Project Stargate's Abilene campus runs on GE Vernova LM2500XPRESS aeroderivative gas turbines installed by Crusoe, scaling to 1.2 GW T3. The second Stargate site in Shackelford County, also Texas, runs on a behind-the-meter gas microgrid from Voltagrid in partnership with Vantage T3. Voltagrid's separate agreement with Oracle for the broader Stargate footprint covers 2.3 GW of behind-the-meter gas generation T3.

Meta's Hyperion campus in Richland Parish, Louisiana — a 2 GW first phase scaling to 5 GW — pairs the buildout with reported plans for on-site gas generation; the campus targets a 1.5 GW interim milestone by end-2027 with full first phase opening 2028 T3. Meta's Indiana site in Lebanon targets ~1 GW with first buildings operational end-2027 T3.

Behind-the-meter solves the queue problem at the cost of a different problem. Natural gas turbines need gas-pipeline interconnection, water for cooling, and air permits. They also embed the data center in a long-duration fossil exposure precisely when the cohort's enterprise customers — including Microsoft and Alphabet themselves — have public net-zero targets. The escape valve works for first-mover campuses; it does not scale as the default path for the full $450B AI capex envelope.

Where the rent migrates

The bottleneck framework predicts that when a constraint binds, the rent migrates upstream. On the silicon side, the rent has visibly moved to the HBM suppliers (SK Hynix Q1 2026 operating margin 72%, Micron and SK Hynix joining the $1T market-cap club within 24 hours on UBS's structural-rerate framing) T1. On the power side, the rent migration is less mature but is starting to surface in three places.

First, the established power-equipment makers. GE Vernova's electrification booking pace ($2.4B in a single quarter for data centers) implies a backlog quality that has not historically existed in the segment T1. Siemens Energy, Hitachi Energy (a Hitachi subsidiary), Schneider Electric, and Eaton are the public names with the cleanest exposure.

Second, the behind-the-meter operators and gas-turbine suppliers. GE Vernova's aeroderivative turbine business sits inside the company's Power segment; Caterpillar and Cummins ship the reciprocating generator end of the same market. Voltagrid and similar private operators are aggregating the deployment.

Third, the engineering, procurement and construction firms with substation and transmission scope: Quanta Services and MasTec on the EPC side; the regulated utility names (Dominion, Southern Company, NextEra) where the rate-case mechanics let data-center load growth flow to ratebase. The rent migration here is conditioned on regulatory accommodation — and the Virginia State Corporation Commission's January 2026 decision to assign more cost to data centers under Dominion's rate structure T3 is the first piece of evidence that the regulatory direction may compress utility upside even where load growth is real.

The Lawrence Berkeley National Laboratory analysis frames the demand at the macro level: data centers consumed 4.4% of US electricity in 2023 and are projected to consume between 6.7% and 12% by 2028, with absolute consumption rising from 176 TWh to 325-580 TWh T1. The mid-point of that range is roughly 450 TWh of incremental demand inside five years — equivalent to adding the entire current electricity consumption of Germany to the US grid.

Variant perception

Consensus framing is that the hyperscaler 2026 capex commitments will deploy roughly on schedule and that the rate-limiter on AI infrastructure revenue is silicon supply (HBM, CoWoS, advanced node logic). The Alphabet equity raise is interpreted as a signal of demand strength; the cohort multiple expansion at Micron, SK Hynix, and Marvell is interpreted as the bottleneck rent the silicon supply chain is extracting. The implicit assumption is that the megawatts are available, or will be made available behind-the-meter at scale, and that the silicon supply chain is the binding step.

AlphaSteve's variant view is that the announced 2026 capex is partially undeployable on the implied 2026-2027 schedule, with the slip concentrated in queue-bound traditional interconnection paths. A rough sizing: of the $450B AI-allocated capex in 2026, roughly half is allocated to power-sensitive land, buildings, and electrical equipment (the other half is GPU and server hardware that ships to existing capacity or to behind-the-meter sites already under construction). Of that $225B power-sensitive share, roughly 30-50% faces material grid-driven slip — 12 to 24 months — unless behind-the-meter natural gas scales materially faster than current permit cadence supports. The mechanical consequence is that the AI training and inference revenue line for hyperscaler customers (Microsoft Azure, Google Cloud, AWS, plus Meta internal) grows at the deployment-curve rate, not the capex-commitment rate. Sell-side models that bridge capex commitments to revenue growth via fixed asset turnover assumptions are over-stating the bridge.

Three observables would falsify the variant view. First, a measurable acceleration in either PJM or ERCOT effective processing time below the 2-year-to-agreement reform target. Second, a step-change in transformer manufacturing capacity disclosure (GE Vernova, Siemens Energy, Hitachi Energy) that materially closes the 7-9% demand / 3-4% supply gap. Third, a behind-the-meter natural-gas-turbine deployment cadence at 20+ GW per year across the hyperscaler campuses — roughly five times current run-rate.

Three observables would confirm the variant view. First, a hyperscaler capex commentary disclosure that specifies a year-over-year deployment pace below the implicit guide (most cleanly visible in 10-K depreciation expense relative to gross capex). Second, a public delay on either Meta Hyperion's 1.5 GW end-2027 milestone or Stargate Abilene's 1.2 GW full ramp. Third, a sustained Brent-WTI-style spread between hyperscaler capex commitments and electrical equipment supplier backlog growth — i.e., capex booked but not flowing to GE Vernova / Siemens / Hitachi backlog at the implied ratio.

The variant view is not a call against AI infrastructure demand. It is a call that the deployment-rate constraint is a real second axis the cohort multiple is not yet pricing, and that the rent migration toward power equipment is the more durable trade than chasing the HBM cohort at current stretch.

Implications for AlphaSteve

The deployment-rate constraint is a parallel binding observation to the HBM constraint and changes the kit's read of the AI infrastructure cohort in three specific ways. First, hyperscaler revenue growth through 2027 is bounded by power deployment, not capex commitment; the implicit fixed-asset-turnover bridge in sell-side hyperscaler models is too optimistic by the queue-driven slip share. Second, the rent migration on the power side is structurally durable — transformer and turbine lead times are physical, not financial, and reverse on a multi-year horizon. Third, the AI infrastructure variant on duration sharpens: the cohort multiple is paying for the silicon constraint as if the silicon constraint is the only binding step, while the next constraint behind it is paying first-mover rent to a different set of names.

  • Portfolio: no action; AI-infrastructure-upstream exposure remains zero on margin-of-safety discipline. The power-equipment cohort enters explicit watchlist consideration.
  • Watchlist: add GE Vernova (GEV) for thesis pass — pure-play exposure to data center electrification (Q1 2026 booked $2.4B in segment), transformer and aeroderivative turbine breadth, Prolec GE consolidation, US grid build-out trajectory. Quanta Services (PWR) for transmission EPC exposure. Siemens Energy (SMNEY ADR / ENR.DE) as European comparable. MasTec (MTZ) on the EPC side. Add Vistra (VST) and Constellation (CEG) on the deregulated-power-sale side as behind-the-meter PPA counterparties. MP Materials thesis pass remains the top operational priority.
  • Theses on the workbench: no new thesis from this report; the deployment-rate constraint is observational and informs the AI infrastructure variant view; GEV thesis pass should be sized for next-week priority queue alongside MP Materials.
  • Sectors: Industrials / Electrical Equipment subsector (GE Vernova, Eaton, Siemens Energy, Hitachi Energy) and Utilities subsector with data-center concentration (Dominion, Constellation, Vistra, Southern Company) — both elevated as kit-watch targets on the deployment-rate constraint reading. Information Technology / Semiconductors framing is unchanged.
  • House view updates: update "AI infrastructure capacity — current" position to add the deployment-rate constraint as a parallel binding observation; add a new sub-position on power equipment as the next-stage rent migration. Specific text in "House view changes this run" below.
  • Daily-scan adjustments: add to the Wednesday daily-scan the explicit observables: (i) PJM or ERCOT effective queue-processing time disclosure; (ii) GE Vernova / Siemens Energy / Hitachi Energy quarterly electrification or grid backlog growth rate; (iii) hyperscaler 10-Q depreciation growth as fraction of capex growth. Any of the three crossing the variant-view falsification thresholds would re-rate the read.

Charts / data

Table 1 (above) — Grid queue versus current peak demand and processing time at the two largest US grids.

Approximate sizing of the deployable-share question, AlphaSteve calibration:

Capex layer 2026 announced Power-sensitive share At-risk to grid slip (AS-cal)
Hyperscaler total (MSFT/GOOGL/META/AMZN) ~$700B ~$350B ~$100-175B
AI-allocated share ~$450B ~$225B ~$70-110B

AS-cal

Sources

See sources-policy for the citation discipline applied.

House view changes this run

  1. Update "AI infrastructure capacity — current" position language to add the deployment-rate constraint as a parallel binding observation alongside the HBM-primary silicon constraint. Specifically: add a new clause that "demand-side and capex commitments are no longer the binding constraint; the binding constraints are HBM at the silicon layer and grid interconnection / transformer supply at the deployment layer, and they bind in parallel." Add the queue-math and transformer lead-time anchors to the supporting evidence list.

  2. Add to the material risks list the explicit deployment-rate observable: "a hyperscaler 10-K disclosure where depreciation expense growth materially lags capex growth, indicating capex booked but not yet flowing into operating assets, would confirm the deployment-rate constraint." Reciprocally: "a step-change disclosure in PJM or ERCOT effective interconnection processing time below 18 months, or in transformer manufacturing capacity above 6% growth, would partially falsify the variant view at the deployment layer."

  3. Add new sub-position under Technology & sectors: "Power equipment as next-stage AI rent migration — provisional." Position language: "GE Vernova, Siemens Energy, Hitachi Energy, Quanta Services, and the broader transformer/turbine/EPC complex are the next-stage rent recipients as the AI infrastructure binding constraint migrates from silicon to deployment. Confidence: medium. Provisional pending GEV thesis pass." Add the GE Vernova Q1 2026 $2.4B electrification booking disclosure, the 128-week transformer lead time data point, and the Siemens Energy Charlotte plant 2027 start as supporting evidence.

  4. Add to the changes log entry "2026-06-03 Wednesday long-form (technology): deployment-rate constraint observation extends HBM-primary silicon-constraint observation as parallel binding constraint at the power/data-center layer; new sub-position on power equipment as next-stage rent migration."

Linked