The Grid Is the Binding Constraint

The AI capex programme has committed a category error: it has treated electrical power as a software-speed input — abundant, low-marginal-cost, scaleable on the chip timeline — when power is a physics-speed input built on 50-to-70-year infrastructure cycles. That mismatch is not a future risk; NERC's January 2026 Long-Term Reliability Assessment, PJM's capacity auction clearing at the FERC ceiling for two consecutive years, and the hyperscaler nuclear arithmetic together confirm the collision is binding now. The investment implication follows directly: value is migrating from the platform layer to the physical-input layer, and the price signals are already in the data.

The Demand Shock

The North American Electric Reliability Corporation is the federally designated reliability authority for the bulk power system, and it is constitutionally conservative. Its job is to be conservative because when its forecasts are wrong on the low side, the lights go out. When NERC's January 2026 Long-Term Reliability Assessment revised the ten-year US summer peak demand forecast upward by 224 gigawatts — a 69% increase in a single annual cycle, the largest in the assessment's history — that number had to be read as an understatement. NERC attributed roughly 70% of the revision to large industrial loads, the category now dominated by hyperscaler data centres; it designated MISO and PJM elevated-risk from 2026 to 2028, escalating to high-risk by 2028-2029; and it warned that large industrial loads alone could add up to 300 GW of new demand between 2028 and 2030. PJM alone expects summer peak demand to grow by 56 GW to 210 GW by 2035.

The corroborating evidence converges from four independent institutions using four separate methodologies. The EIA's Annual Energy Outlook 2026 projects data centres and EVs together accounting for 50-80% of all incremental US electricity demand growth to 2050, with data-centre server electricity use reaching 818 billion kWh in the high-demand case — sixteen times the 2020 figure. EPRI, working bottom-up from utility interconnection requests, projects data centres consuming 9-17% of US electricity by 2030, a 60% upward revision from its prior estimate. The IEA's Electricity 2025 adds the global frame: data-centre electricity demand rose 17% in 2025, AI-specific consumption rose 50%, and the US accounted for roughly 50% of global incremental electricity demand growth in 2025, almost entirely data centres. NERC, EIA, EPRI, and the IEA coordinate nothing with one another. The lowest-case consensus in 2026 now exceeds what was the bull case eighteen months earlier.

The insider call preceded all of it. Leopold Aschenbrenner's June 2024 Situational Awareness projected data-centre electricity reaching approximately 5% of US production by 2026 and 20% by 2028, and stated plainly: "probably the single biggest constraint on the supply-side will be power." The first of those three projections has materialised on his calendar. The second is what NERC's January 2026 LTRA is now describing.

The Supply Reality

Against that demand curve, the grid's supply reality is a widening dispatchable shortfall. Meeting the revised outlook while maintaining historical reserve margins requires roughly 30-40 GW per year of new dispatchable generation through 2030 — capacity callable at any time, regardless of weather or sunlight: gas, nuclear, hydro, battery-firmed renewables. Actual net new dispatchable capacity additions are running at 5-8 GW per year. Solar and wind are being added at 30-40 GW per year nameplate, but they are not one-for-one substitutes for firm capacity; a megawatt of solar does not deliver at 7 PM on a still summer evening at peak data-centre load. Coal retirements are simultaneously removing 10-15 GW per year of dispatchable capacity through 2030. The shortfall is not static — it widens as the grid retires firm capacity and adds intermittent capacity in the same motion, and grid inertia declines as the synchronous generation required for frequency stability is progressively replaced by intermittent sources.

The interconnection queue is where the build-out either happens or fails. FERC reported approximately 2,600 GW of generation in national queues as of mid-2025 — more than twice total installed US generating capacity. The historical clearance rate is the binding constraint: roughly 14% of queued projects ever reach commercial operation, and the average queue time in PJM is now 5-7 years, up from 2-3 years before 2020. The queue is full; the grid is short; the binding constraint is not capital allocation but permitting, transmission, supply chain, and skilled labour, in roughly that order.

Equipment lead times compound the problem. Large power transformers extended from 6-9 months in 2015 to 18-24 months in 2025. Generator step-up transformers — the substation-grade units without which no new generating station can connect to the high-voltage grid — averaged 144 weeks per Wood Mackenzie's Q2 2025 survey, with some high-capacity units extending to four years. Transmission build is roughly twice as slow as generation because every long-distance line crosses jurisdictional and political boundaries negotiated separately. Microsoft's CFO has disclosed that approximately $80 billion of contracted Azure demand is currently blocked by power availability rather than capital availability. That sentence is the constraint stated in financial terms.

The Price Signal

The PJM capacity market is the cleanest available read on whether the grid has enough firm capacity to meet projected demand. The sequence demands no commentary:

PJM procured 134,311 MW of unforced capacity in 2026/27 — below reliability-target requirements in some zones — per the 2026/27 BRA report. In a normal market, a near-elevenfold price increase in twenty-four months would attract new capacity inside two years and normalise prices. The spike persisted at the regulatory ceiling for two consecutive auctions. The second-order problem is equally important: even at $329.17/MW-day, the economics of building a new 500 MW combined-cycle gas plant are marginal — roughly 6-8% return against a 10-12% hurdle rate. The price signal is screaming and the supply response is muted because the cap suppresses it and construction economics are damaged simultaneously. A capacity auction clearing at the regulatory price ceiling for two consecutive years is the textbook empirical signature of a binding physical constraint. There is no plausible reason to expect the 2027/28 auction to clear lower; the queue of new dispatchable capacity that could clear it is empty.

The Nuclear Arithmetic

The standard reply from every hyperscaler earnings call since late 2024 is that nuclear procurement solves the constraint. The arithmetic deserves direct examination. Microsoft's Constellation partnership restarts Three Mile Island Unit 1 at 835 MW, targeted 2028. Amazon's expanded Talen Energy deal at Susquehanna covers 1,920 MW, ramping 2029-2032. Google has contracted with Kairos Power for 500 MW of SMR capacity, targeted 2035. Meta announced 6.6 GW of aggregate commitments across Vistra, TerraPower, and Oklo in January 2026.

The number that matters is net new dispatchable nuclear by 2030 that would not otherwise exist on the grid. Three Mile Island counts — the reactor was decommissioned. Susquehanna is existing capacity; the Amazon deal is long-term offtake, with perhaps 300-500 MW of net new ramp by 2030. Meta's Vistra deal covers 2.1 GW of existing PJM nuclear plants — reallocation, not new megawatts. TerraPower, Oklo, and Kairos do not deliver commercial megawatts before the 2030s. Brownfield restarts add roughly 800 MW to 1.5 GW of true net new by 2030. Total hyperscaler-driven net-new nuclear by 2030 lands at approximately 1.8 GW against 80-100 GW of hyperscaler-driven demand growth over the same period — a coverage ratio under 2%.

Vogtle Units 3 and 4 are the structural benchmark for what new nuclear actually costs: original 2009 estimate $14 billion; final 2024 cost $36.8 billion; time elapsed fifteen years, six years late. NuScale cancelled its flagship Idaho SMR project in November 2023 after costs escalated. The hyperscaler nuclear deals are real and useful. They are best understood not as the answer to the constraint but as evidence the hyperscalers are aware of it — buying optionality and hedging against PJM-signalled price increases.

The Physical-Layer Chokepoints

The capacity-constrained equipment and network holders define who captures the spread the price signal is already generating.

GE Vernova carries a 100 GW gas turbine backlog as of Q1 2026, delivery slots sold out through 2030, new orders pricing 10-20% above the prior-year backlog on a $/kW basis, and CEO Scott Strazik has stated the company expects to be fully booked through 2030 by end-2026 against annual production capacity of around 20 GW. (Utility Dive) Siemens Energy sold 194 gas turbines in FY2025 against 100 in FY2024, reported a record order backlog of €138 billion, faces 5-7 year lead times for new heavy-duty units per S&P Global's May 2025 survey, and has a €2 billion manufacturing-capacity expansion not online until 2028. (Siemens Energy FY2025) Together they form a tight oligopoly in which every delivery slot inside the decade is effectively allocated.

Hitachi Energy sits at the centre of the transformer constraint: demand for generator step-up transformers is up 274% since 2019, a $1 billion South Boston plant investment is scheduled for 2028, and the 144-week average GSU lead time means a hyperscaler that cannot get a transformer cannot commission the substation that would connect its campus. The broader transformer market context — transformer prices up approximately 80% over five years, with Cleveland-Cliffs as the sole domestic US producer of grain-oriented electrical steel, the core input for power transformers — makes the constraint upstream as well as at the equipment level; Cleveland-Cliffs' single-source status creates a hard ceiling on domestic transformer production expansion that no amount of capital can immediately remove. Constellation Energy has locked in 20-year PPAs with Microsoft for the 835 MW TMI-1 restart and with Meta for 1,121 MW at Clinton Clean Energy Center, and closed its $16.4 billion Calpine acquisition in March 2026, adding gas dispatch capability across a 5,650+ MW fleet. Vistra Corp signed a 20-year PPA with Meta for 2,609 MW of PJM nuclear capacity, cleared 10,255 MW in the 2026/27 PJM auction at a $273.45/MW-day weighted average, runs 100% hedged for 2026 and 84% for 2027, and extended its dispatchable portfolio via the $4 billion Cogentrix acquisition. (Vistra 10-K) These two entities are capturing the bilateral PPA premium that the PJM price signal is generating.

The second-order chokepoints are the upstream gates that determine whether the first-order layer can scale. Prysmian, Nexans, and NKT dominate HVDC and high-voltage cable manufacturing: Prysmian's transmission backlog ran approximately €17 billion at FY2025 close; NKT's backlog rose from €10.2 billion at end-2025 to €13.5 billion in 2026; DC cable lead times now exceed five years. Eaton sits at the medium-voltage switchgear layer, with its Electrical Americas backlog 4× the 2019 level at roughly $10 billion and AI-related backlog up over 30% year-on-year in 2024. BWX Technologies is the only currently operating US manufacturer of TRISO nuclear fuel for advanced reactors, having delivered the first full core for the Pentagon's Project Pele microreactor in December 2025. (BWX Technologies) Centrus Energy is the only current US HALEU producer at 900 kg/year at Piketon, Ohio — pilot-scale, not commercial — with a $900 million DOE task order to expand to commercial scale by 2030; until that facility comes online, HALEU is a pre-commercial bottleneck for the SMR fleet the hyperscaler PPAs are funded to build. (Centrus DOE contract)

The desk currently tracks AECOM (ACM) against this bottleneck thesis. AECOM disclosed in May 2026 its membership in the UK Infinity Fusion Consortium alongside Tokamak Energy and Type One Energy — making it the first publicly listed EPC firm with a named fusion-programme engagement. Its broader grid-EPC backlog in transmission and data-centre site civils underwrites the equity independent of fusion-timeline risk; fusion is optionality on top of a picks-and-shovels grid position the constraint thesis already supports. Posture remains monitor.

The Scholarly Foundation

The structural anchor is Vaclav Smil's iron law. Energy and Civilization (2017), How the World Really Works (2022), and Speed (April 2025) all rest on the same empirical observation: every major historical energy transition has taken 50-70 years from first commercial introduction to majority share — coal displacing wood roughly 80 years, oil displacing coal 75 years, natural gas from the 1970s still incomplete at roughly 80 years. The governing dynamic is addition, not substitution: global coal generation rose from 9,500 TWh in 2015 to 10,200 TWh in 2025; natural gas rose 25.9%; renewable generation doubled, but as supplement, not substitute, with the global renewable share of primary energy creeping from 11% to 14% while absolute fossil use increased. (Smil transition paper via EnergySkeptic) The iron law is thermodynamic and industrial. Capital cycles in the technology sector run on five-to-ten-year horizons; energy infrastructure transitions run on fifty-to-seventy-year horizons. The mismatch is not a forecasting error — it is structural.

Goldman Sachs's Generational Growth (Murti and Schneider, April 2024) quantifies the capex gap: required cumulative power-infrastructure investment over the decade is roughly $1.4 trillion; current trajectory is roughly $700 billion; structural under-investment of $70-100 billion per year sustained over a decade. Mark Mills at the Manhattan Institute arrives at the same figure from a different methodology: an additional $70-100 billion per year of energy capex above current trajectory required to keep pace with AI-driven demand growth.

Jonathan Koomey is the empirical sceptic, and the sceptic's concession is important. The 2020 Masanet et al. paper in Science established that prior-decade data-centre energy projections were systematically overstated three to four times because efficiency gains absorbed most projected growth. Koomey's 2025 position is more specific: AI demand is growing at 20-25% annually against dispatchable capacity growth of 3-5% annually; efficiency gains will partially compress the trajectory, perhaps by 15-20% per year. The bottleneck thesis does not require the alarmist case. It requires only Koomey's own central estimate — 20-25% demand against 3-5% capacity — and the arithmetic still produces a binding shortfall by the late 2020s. The sceptic case narrows the gap. It does not close it.

The Investment Expression

Thesis claim C-26 holds that as AI inference commoditises and platform-layer margins compress, value migrates to the physical-input layer — energy, materials, grid — where picks-and-shovels suppliers earn the revenue the software layer fails to translate. The commoditisation mechanism is already operational: DeepSeek R1 in January 2025 erased over $500 billion of NVIDIA market capitalisation in a single trading day by demonstrating frontier-grade performance at a fraction of the compute budget. Jevons's paradox ensures that as inference cost-per-token falls, demand for inference expands faster than per-query efficiency gains — the value of the physical inputs is unaffected by software-layer compression. The platform layer — the Mag 7 — requires 18-23% consensus EPS growth to justify current multiples. The physical-input layer earns revenue regardless of whether the platform layer earns its return.

PJM clearing at $329.17/MW-day for two consecutive years is the price signal the physical-input layer is capturing. GE Vernova's backlog priced 10-20% above prior-year on a $/kW basis; Siemens Energy at €138 billion record backlog; Constellation and Vistra with bilateral PPAs above market; the transformer market with prices up approximately 80% over five years and Cleveland-Cliffs as the sole domestic GOES supplier creating a hard ceiling on any rapid domestic expansion; Prysmian and NKT with DC cable backlogs compounding against lead times exceeding five years — these are the named holders of the spread the demand curve is generating, and none of their order books depends on whether the AI platform layer ever earns its return.

The geopolitical confirmation adds a further dimension. Aschenbrenner wrote in June 2024: "we're going to drive the AGI datacenters to the Middle East, under the thumb of brutal, capricious autocrats." In May 2025, OpenAI, Oracle, SoftBank, Nvidia, and G42 announced Stargate UAE — a 1 GW Abu Dhabi cluster with 200 MW expected operational by 2026. The warning materialised eleven months later. When the supply constraint binds domestically, capital does not wait for the constraint to be solved — it relocates to jurisdictions where it does not bind. The domestic analogue is equally telling: approximately 30% of newly planned data-centre capacity is designed for on-site generation as of early 2026, up from effectively zero in 2024, with one-third of hyperscalers targeting full on-site power by 2030. (Troutman/Tamarindo Off-Grid Data Centers, February 2026) The constraint is not being solved — it is being routed around.

What Would Change My Mind

Thesis claim C-05 — that the bottleneck binds on the 2026-2027 timescale — is falsified if the PJM Base Residual Auction clears below $200/MW-day for two consecutive auctions and NERC's LTRA peak demand growth revision falls below 30 GW. Either condition alone narrows the thesis; both together reverse it. Thesis claim C-26 — that value migrates to the physical-input layer — is falsified if PJM clears below $100/MW-day within two auctions, indicating the constraint has been resolved and physical-layer pricing power has collapsed; or if Mag 7 EBIT margins expand more than 300 basis points per year for two consecutive years, indicating the platform layer is successfully translating capex into earnings and the migration thesis is wrong. Three additional thresholds: if GE Vernova or Siemens Energy announces available heavy-duty gas turbine delivery slots before 2029 for orders placed in 2026, the manufacturing lead-time bottleneck has dissolved; if off-grid share of newly planned data-centre capacity falls back below 15% in 2026-2027 announcements, grid interconnection timelines have shortened materially; and the 2027/28 PJM BRA report, published December 2025 for the next clearing cycle, is the primary near-term read on which direction the arithmetic is moving. What does not falsify the view: further DeepSeek-class efficiency gains — Jevons's paradox and the absolute demand base already established by NERC mean the arithmetic survives Koomey's own optimistic efficiency assumptions; additional hyperscaler nuclear PPA announcements — additional deals do not change the sub-2% net-new coverage ratio by 2030 unless SMR and restart timelines compress materially; or the Anthropic round closing at a terminal-stage valuation — that confirms the platform-layer analogue but does not resolve the physical constraint the platform layer sits on top of.