AI power constraint Binding ◆

Rasgon: assuming AI demand holds, what stops Jensen's $3T-by-2030 build-out isn't chips — it's power. The US grid can't add the needed capacity, so on-site generation is booming: gas turbines (3-year lead times), small modular reactors, even Three Mile Island restarting. His framing: "The US has chips but no power. China has power but no chips" — China can throw up coal plants at will, but semicap sanctions (no ASML) keep its chips uncompetitive; Huawei is forced into known-but-early techniques "because they have no choice." App Economy (NVIDIA, May 22): the bottleneck is visibly shifting from GPUs to power — NVIDIA's up-to-5GW DSX partnership with IREN (plus a 5-year right to buy 30M IREN shares at $70) is the AI-factory thesis in action; the constraint is now "power, land, cooling, networking, deployment speed, and operating know-how." Prins (Jun 9): the "electrification of everything" is the demand side — BloombergNEF shows data centers and high-heat industry (not EVs/heat-pumps) driving an unprecedented power surge, bullish grid reliability plus nuclear and geothermal generation. Polomny (Jun 13): the binding constraint is now showing up as local backlash — towns slapping moratoria on data centers over water and power (his own county, plus Pennsylvania's surging power prices) — so own the land/royalty layer where the build-out is unopposed: Texas Pacific Land (TPL) and LandBridge (LB) in empty West Texas (stranded gas + water; "nobody lives out there"), where data-center demand has lifted his local power price ~50% in five years. Salzman (Jun 10): a pure-play on the workaround emerges — ERock (EROC) IPOs on the NYSE selling on-site natural-gas "rockblock" gensets that give data centers speed-to-power around the multi-year grid-interconnection queues in Texas/Pennsylvania (customers Meta/Microsoft/Foxconn, $1.3B backlog; peers Caterpillar, Generac, Cummins). The queue length is the moat — but down 5% on debut and at ~32× 2025 sales (vs Caterpillar's 6×) the stock prices in a steep ramp. Hay (Jul 2): the electricity bottleneck — not chips — is the binding constraint on the AI data-center build-out, and it is a US-vs-China race the US is losing: America leads ~10:1 on data centers (4,423 vs the UK's ~470, China's 369, with 700 under construction and ~1,500 planned) but China's electricity output is growing far faster (much of it coal, plus more installed renewables than the rest of the world combined and the largest new-nuclear program) and already generates ~2× US power. Hay's standing fix: combined-cycle natural-gas plants are the main answer, but the turbines at their core are in severe short supply, making small modular reactors essential — at least 10 companies (unnamed) have hit or are near criticality, some targeting meaningful output in 2027. Hay (Jul 8): the microreactor/SMR race has entered the data-center power stack — Aalo Atomics reached test-reactor criticality over the July-4 weekend (one of only four companies to do it; truck-deliverable, mass-producible microreactors) while rival Valar demonstrated full temperature and power conversion AND signed a 30MW contract to power an Nvidia data center — nuclear increasingly supplementing/replacing natural gas as the data-center fuel, tempered by years-long NRC licensing that could erase any first-mover's lead; China's deployment lead makes the race a national imperative. Murti (Jul 11): the "Power Surge" super-cycle passed its first war test — the S&P and the AI trade shrugged off the Hormuz crisis (the April-7 cease-fire the turning point, with strong Q1 tech earnings), which he reads as supportive of the multi-year power/electricity demand build; no bubble call either way. Barron's (Serwer, Jul 10): WATER is the other physical constraint — Veolia (the largest private US water operator) calls data-center water a "rising opportunity" and is already the water-services layer at TSMC's Phoenix chip complex and Amazon's Mississippi data centers; the same community-pushback-over-water dynamic Polomny flagged makes the treatment/recycling vendors the toll collectors. Currie (Jul 17): the AI/electrification/data-center build-out cannot reach equilibrium without much higher commodity prices — FIDs are needed upstream (oil & gas), downstream (refining) and in metals/mining to feed it. The hyperscalers "are commodity guys… in a cyclical business" needing a re-rating; energy is ~3% of the S&P vs ~18% in 2014, so generalist capital only returns once prices force the re-rating. Smith (Chronometer, Jul 21): reframes the power constraint as a gas constraint — every BYOG data-center genset/fuel-cell press release "means more gas," and energy could rise from ~10% to 20-40% of hyperscaler compute cost by 2029 ("this sounds like DRAM two years ago"); distributed gen-sets are inefficient, high-heat-rate assets that should only ever be backup, large-scale AP1000 nuclear (~2033-34) is the only durable fix, and SMRs are still "science experiments" not set up to manufacture at the tens-of-GW scale required. Chanos (Risk Reversal, Jul 17, the dissent): power is NOT the bottleneck — it's ~5–6% of data-center revenues, the smallest cost component, so power-scarcity stories are "silly" ("we have permission to connect to the Texas grid" — great, at wholesale prices; "read the fine print"); transmission/regulatory bottlenecks get engineered around within 2–3 years (Texas, the biggest new-construction market, already is), and Microsoft's Chevron direct-power deal shows the workaround. The bigger data-center risk is political backlash, not electrons — a direct counter to both the scarcity thesis and the space-data-centers-save-power pitch. Hay (Jul 26): the US has the data centers but not the electricity; China the reverse. The grid "is already so taxed" that new gigawatt reactors couldn't be wheeled to load anyway — hence behind-the-meter micro reactors; gas is the bridge but turbine capacity is unobtainable, one of enough physical bottlenecks (turbines, capacitors, transformers, permits, land, water) that the buildout "probably [isn't] going to get built at the rate that Wall Street is currently expecting" — and AI capex has been ~70% of GDP growth in some quarters, so a slowdown bleeds into the whole economy. Singh (Jul 26): the OpenAI build Nvidia is in talks to backstop (~$250B guarantee) is a 10-gigawatt SoftBank project in Southern Ohio — single-project scale is now measured in gigawatts. Huhn (Jul 23, Uranium Insider): the buildout's binding constraint is physics, not capital — Jefferies finds ~50% of 2026 and ~80% of 2027–28 data centres had not started construction as of Q2 2026, JPMorgan >60% of 2027 capacity not broken ground, Sightline only ~5GW of the 12–16GW announced for 2026 under active construction; HV transformer lead times have gone 12–18 months (2022) → 30–36 months (up to five years), with interconnection queues 3–5 years in Texas/Virginia/the Carolinas. A second reset risk is pricing — Z.AI's open-weight GLM 5.2 scores 51 on the artificial-analysis index at ~85% lower run cost than ChatGPT 5.5, raising ROI questions over frontier-lab capex. His read: an "AI scare" transmits to uranium mechanically via Wall Street's AI factor baskets but is short-term — the Jan-2025 DeepSeek −10% day was "a false flag," slipped gigawatts make surviving capacity more valuable, Jevons Paradox applies to compute, and "the uranium supply/demand story is not reliant on AI growth." App Economy (Aug 1): a rare second-order read — Hilton traced its biggest RevPAR flip (US mid-scale and upper mid-scale, from negative last year to strong growth) directly to the AI data-center buildout, because the contractors and engineers doing that work stay in mid-range hotels, not luxury. "If data center spending is really driving mid-scale growth, that demand shouldn't fade until the buildout does." 2026-AUG-03 (Paulo Macro): the power-buildout equities are cracking — "the power gen/components/buildout stories like Caterpillar, Eaton, Vernova, etc have started to break down amidst the violent rotation away from semiconductors and a general stagnation in the AI dreamscape." Per GS the AI Data Center factor has become a primary driver of copper prices YTD, so the breakdown is "a real risk" to copper too; the rhyme is tech investors crowding into uranium in 2020-21 as a "value" energy hedge. "The trade is not 'sleep at night'… do the hard trade." 2026-AUG-08 — App Economy (PRO): DigitalOcean has ~155 MW of power secured, most landing through 2027, and cut FY26 FCF-margin guidance to 11–13% to fund it — demand is explicitly "accelerating faster than capacity." Pinterest committed $4B to AWS through 2031. 2026-AUG-11 — Hay (Haymaker Daily): the constraint stated as a physical disqualification — data centers "require uninterrupted power to properly function. They can't operate in accordance with daylight hours and wind patterns," and "wind and solar are unequal to the task of providing reliable baseload power at scale." The US already has 10× China's installed data centers, before the hyperscalers' $5.8T through 2030 — "without electricity these facilities will be nothing but inert shells." (Hay's quarrel with the buildout stays with its financing and accounting, never its electricity bill.) 2026-AUG-11 — Oakley (David Lin Report): in Texas the binding constraint is water, not power — "we have the energy… in West Texas and South Texas you don't have enough water"; the state's data-center construction moratorium "probably all changes after the election" 2½ months out. 2026-AUG-07 — Salzman (Barron's): Texas is building more natural-gas plants than the next seven states combined and leads in renewables, yet the binding constraint is turning out to be water. TWDB: a severe drought leaves the state ~20% short of water demand by 2030; a University of Texas study puts data centers at <1% of state water use last year → 3.7% by 2030 → up to 9.1% by 2040 (more than manufacturing uses today), clustered in West Texas "where energy is abundant but water isn't." Corpus Christi is the live case: Choke Canyon Reservoir 8% full, Stage-3 residential restrictions since Dec 2024, residents cutting use 19% (2023–25) vs 3% for large users, industry taking ~60% of city water (Exxon ~13m gal/day). Gov. Abbott has halted new data-center approvals pending a water audit — directly corroborating Oakley's Aug-11 "in Texas the binding constraint is water, not power." 2026-AUG-02 (Gromen): US electricity generation in 2024 was essentially flat vs 2004 despite "massive wealth growth on paper" — since power use and real GDP are tightly correlated, "the US didn't really grow on a real basis for 20 years." AI first, then reshoring, now reverses it; the grid/equipment names are "the people selling picks and shovels to the mining boom" (PAVE/GRID as the holdings screen, Japanese industrials the reshoring workhorse). 2026-AUG-14 (Haymaker): "At its core, the AI story is fundamentally an energy story… secondary to the fact that AI is going to require an enormous amount of reliable power." New evidence standard: PJM and ERCOT set all-time peak loads of 168GW and 91GW in July 2026 — "the structural AI demand proof the thesis requires, visible in real-time load data rather than forward projections" — while hyperscaler capital partners directly with generation (Helix = KKR + NVIDIA + Kuwait Investment Authority with VST as preferred power partner; Meta ~2,600MW PJM nuclear PPAs; AWS ~3,800MW at Comanche Peak). 2026-AUG-14 (App Economy): conversion, not demand, is the binding limit — CoreWeave targeting >1.85GW of active power by year-end against a $104B backlog (+246%), Cerebras >600MW live/contracted through 2027 against $25.4B RPO; backlog is a conversion problem measured in energized megawatts. AUG-16 (Jay Singh, SSR): three fresh datapoints on power as the binding constraint, each expressed as a position. Riot Platforms is renting 109 MW (to 286 MW) of Rockdale, Texas capacity to Anthropic for $9.1B (to $16.1B) through 2048 at $2.4M per MW of annual rent on an 80-90% NOI margin — real-estate economics on an electricity connection. CoreWeave's active power more than tripled to 1.5 GW (target now >1.85 GW) against a $104B backlog. And Babcock & Wilcox's backlog jumped 533% to $2.6B with a >$14B pipeline and another 1 GW of Siemens Energy steam turbines secured for anticipated data-centre projects, the CEO citing AI/utility/industrial demand for generation capacity. 2026-AUG-17 — Jay Singh (David Lin Report): his top 2H theme is the AI pivot from chip hype to power and ROI: grid bottlenecks, battery infrastructure, co-location, and power names like VST — "they may not need to buy chips every single year, but they will need to buy power every year." Prinsights (Nomi Prins), 2026-AUG-19: data-center power consumption is projected to climb toward ~600 TWh by 2035 (Investor's Business Daily) — natural gas anchors the near-term additions while nuclear plus renewables take the lion's share of the long-term structural additions. Her framing of the constraint: it is not raw generation but round-the-clock reliability, which is why hyperscalers (Microsoft, Amazon, Alphabet, Meta as the counterparty class) will sign long-term, high-premium power purchase agreements direct with nuclear operators — and why she says to look past traditional utility valuations and co-located assets toward off-grid / behind-the-meter baseload, where the firmness premium accrues to the owner rather than the ratepayer. 2026-AUG-19 — Salzman (Barron's): the constraint acquires a political layer, and it is bipartisan — a one-year data-center moratorium in New York, an industry "audit" in Texas pausing grid connections, and Pennsylvania Gov. Shapiro's executive order forcing data centers to bring their own power and win community buy-in before the DEP will consider an application — applying "whether or not they connect to the larger electric grid," and stripping every data-center project out of the state's grid "fast track." Johns Hopkins' Abe Silverman: "We are a split country, except on this one issue." Pennsylvania has no AI data centers yet but 100+ proposals, surplus generation and abundant gas. The effect is delay, not derailment (no ban; BTIG's Kania sees no cancellations) — but it makes dedicated new generation the mandatory complement, which reprices turbine slots rather than megawatts. (Avi Salzman / Barron's, 2026-aug-19.) 2026-AUG-22: TIP: the under-discussed constraint is load shape, not load level — tens of thousands of GPUs compute in parallel, pause to synchronize, and fire again, swinging hundreds of megawatts in seconds; batteries sit between the building and the grid and absorb the shocks (Brodersen's stadium analogy: an AI data center is 70,000 people all cheering at once, every second or two, for weeks). Sizing: data centers need 200 GW of continuous power by 2030 — roughly the state of California. Storage carries higher margins than EV cells; CATL's H1-2026 storage sales grew ~88% Y/Y. (The Investor's Podcast, Bull vs. Bear 2026-AUG-22) 2026-AUG-27: Every: the political backlash to data centers is building — rising electricity bills, “fields turned into these giant metal boxes,” compute that may take jobs — and statecraft needs consent: “the man and woman in the street absolutely [must feel] that they're getting something from it.” Expects a political pivot toward local benefit-sharing. (Michael Every, Thoughtful Money 2026-AUG-27) 2026-AUG-24: The backlash inverts: because state pauses act on new interconnection requests, not operating assets, an already-connected site is the asset that gets marked up. Morgan Stanley's Stephen Byrd calls Bitcoin miners (CIFR, HUT, GLXY, MARA, RIOT) “a safe harbor in the selloff” — they queued for grid access years ago and are converting warehouses to AI hosting — while power-plant owners needing new approvals (CEG, NRG) are the casualties: “a lot of data center projects have run into cancellations, delays, lack of approval, and that just makes these companies' sites more valuable.” WGMI −16% in a month priced the group as exposed; Byrd says it inherited the theme's beta, not its risk. (Avi Salzman, Barron's 2026-AUG-24, quoting Morgan Stanley's Stephen Byrd) 2026-AUG-28 — the constraint reaches the equipment layer: every large-scale data center "requires multi-megawatt backup generation capable of carrying the full facility load through grid outages," and the order book is now measurable — Generac (GNRC) took ~$1B of new data-center orders in the 90 days to Jun 30 ("roughly $11 million per day"), carries a ~$1.6B data-center backlog with $1.35B scheduled for 2027 (excluding a second hyperscaler's committed volume), raised FY data-center revenue guidance to $450M, and signed ~$700M of committed 2027 volume with one hyperscaler plus a global supply agreement with a second. C&I sales +29% in Q2 after +28% in Q1. Haymaker's framing inverts the capex worry: "while others are worrying about CapEx spend on the hyperscalers, Generac has been, and will continue to be, a net benefactor of said spend." (Haymaker/Bustamante POW 2026-AUG-28) 2026-AUG-30 (Jay Singh, SSR call): the bottleneck moved from the grid connection down to the last few centimetres of the board. Navitas (NVTS) agreed to buy Claros for $232.8M for vertical power delivery and integrated voltage regulator technology, addressing "a major growing challenge for AI infrastructure — delivering thousands of amps to increasingly power-hungry processors at extremely high speed. Traditional voltage regulators distribute power across circuit boards, creating losses… Claros's technology moves the power conversion closer to the processor" — more than doubling Navitas' 2030 serviceable addressable market to over $8B and, per its CEO, helping "break the AI infrastructure power wall." "On pullbacks, this NVTS could be interesting." Two related datapoints: Goldman's desk names power bottlenecks and bipartisan policy pushback on data centres among the three reasons the semi multiple is compressing, and SpaceX plans Nvidia Vera Rubin NVL72 racks in orbit via the Starmind satellite — "much earlier than anticipated" — where cooling and solar power are free. 2026-AUG-24 — John Polomny: the constraint has turned political and therefore durable — "in Texas they've banned data centers," Pennsylvania power prices are up, and "even Mr. Trump said if you want to build a data center you have to have a behind-the-meter power plant… it becomes a populist issue with the voters." His expression is the second derivative: own the land/water/gas layer — LandBridge and TPL, empty West Texas counties ("300 people living in it — no one's going to complain"), owned aquifer rights plus produced-water recycling, and associated gas so abundant "they have to pay people to take it away." Sourced to Murray Stahl / Horizon Kinetics' free research on how much water a megawatt needs. 2026-AUG-31 — CNBC Halftime committee: the backlash goes electoral "the minute we turn the calendar page after Labor Day" (Saccocia); a ~200,000-account Chinese bot swarm pushing anti-data-center misinformation was found on X (with Gavin Baker's long pro-data-center post, quote-tweeted by Jensen Huang, as the Valley's rebuttal); CenterPoint (TX, $5B) and PG&E (CA) both announced ratepayer givebacks funded by data-center load; and 60% of data centers approved for 2027 haven't broken ground. Seasonal overlay: Sept→Nov of midterm years ≈ a 10% drawdown since the 1960s, strong returns after. Pernas (Monetary Matters, 2026-AUG-07): the arithmetic behind bring-your-own-power. US grid demand has been flat at roughly 4,000 TWh for 20 years — appliance efficiency cancelling population growth — so nothing was built for growth; AI is now "poised to take 20% of the entire US energy grid demand by 2030, roughly 100 gigawatts." Two hard blockers: transmission is "akin to a two-lane highway which should be an eight lane highway," and data-centre load fluctuation exceeds what the lines can carry — leaving 18–24 month interconnection queues. So hyperscalers must self-generate, and the US "is very fortunate that they have natural gas pipelines almost everywhere" to plug turbines into. Dean Pernas gives the theme 3–5 years "until more alternatives can be brought to the market," and says open-weight models only accelerate it. Second-order: 1–3 GW monolithic campuses are "topping out soon" — you cannot get more power to one site and towns are imposing moratoriums over emissions and water — so the build shifts to ten 1 GW sites that must operate as one machine, and inter-site bandwidth needs rise ~15×. Utility-scale turbine lead times of 18–24 months are the gap microturbines (1–3 month delivery) fill. (Pernas, Monetary Matters, 2026-AUG-07.) Slegers 2026-AUG-13: the constraint expressed as an equipment supplier rather than a utility. Schneider Electric (Paris, SU) is picked for a twenty-year hold on three compounding drivers — permanent electrification, efficiency mandates on buildings and factories, and data centres with “a bottomless thirst for electricity” — supplying the electrical infrastructure, hardware, software and services to run homes, factories and cities across 100+ countries. The pitch captures the build-out without owning generating assets or carrying regulated returns. Greg Abel (CNBC) 2026-SEP-02: the operator's correction to the framing — energy is the constraint, but not the electrons. "I've sort of always had a strong view that energy would be the constraint… we can produce the energy. It's… how long it would take to get the sites prepared and being in a position they could serve the data centres. And I continue to see that as a big constraint." So the binding item is site readiness and interconnection lead time, not generation capacity — which points the trade at transmission, interconnection and construction rather than fuel. Hard number from the utility that serves it: in Iowa, ~8% of Berkshire Hathaway Energy's load came from data centres last year, with "incremental load coming on, both customers requesting it and what we can serve" — the gap between requested and servable being the quantity to track. Slegers 2026-MAY-07: puts Brookfield forward as the physical-layer vehicle — "they own the land, the power, the buildings and the cables" — with three signed commitments rather than ambitions: a $100bn global AI infrastructure programme; a $5bn agreement with Bloom Energy to install up to 1 GW of behind-the-meter power for data centres and AI factories; and a Swedish land allocation of ~350,000 sqm letting one site "more than double its capacity from 300MW to 750MW." Behind-the-meter generation as the answer to grid interconnection queues is the operative detail. Pieter Slegers (Compounding Quality), 2026-JAN-22 reports a new entrant attacking the constraint from the asset side: Brookfield is launching a cloud service called Radiant "to compete with AWS and IREN," planning to "undercut competitors on price because they already own the necessary power, real estate, and renewable energy infrastructure," aimed at "AI Factories for large corporations and entire countries that want sovereign AI infrastructure," and starting from internal use before selling the surplus — "similar to what Amazon did with AWS." Read as a marker that the scarce input has moved from chips to electricity and land, and that the owners of those are now integrating forward into compute rather than renting to it. Pieter Slegers (Compounding Quality), 2026-MAR-08: inverts the AI question into an ownership question — "AI can write code and analyze spreadsheets, but it cannot replace physical infrastructure… Brookfield owns the physical stuff that makes AI actually work," from renewable generation to pipelines and ports. The barrier is permission and time rather than capital: "these assets take billions of dollars, decades of permits, and years of construction to create," so "the more AI grows, the more valuable Brookfield's assets become." The confirming datapoint is contractual rather than projected: "Microsoft and Amazon are already signing deals with Brookfield just to lock in the energy their AI data centers need." 2026-AUG-26 (Dowd/Phinance, WTFinance): power is one of the two hard stops on the build-out — "there's not enough power to plug all these data center build-outs that have been hyped into the grid," which is why he doubts "a lot of this stuff's ever going to get deployed." 2026-SEP-02 (CNBC Halftime — Weiss): the backlash has moved from politics into prices — "it's one of the reasons why GE Vernova is down… why Vertiv hasn't done really much of anything. It's really hitting those stocks that power the data centers or sell into the data centers," with the transmission channel being retail bills: "all neighborhoods know is I don't want that near me… and that is driving my electricity prices through the roof." His counter-evidence is a private price signal from a board seat at a cold-storage cloud company: "our prices are going through the roof in terms of leasing out data center capacity" — so the constraint is siting and permitting, not demand, and the pause is local ("they'll put them elsewhere"). 2026-SEP-07 (Jay Singh, quoting Musk at the G20): “AI growth is now being constrained by power… I think power will be the biggest opportunity over the next few years. There's a potential 15 gigawatt shortfall just by next year alone, with AI chip production growing 40–50% annually while power outside China is only growing at 10–20%.” Corroborating: data-centre construction is the only accelerating building category, and Broadcom's guide is anchored on enabling 20 GW of compute. Vistra's CEO bought $270k of stock the same week. 2026-SEP-09 (David Hay / Haymaker): the shortfall stated as arithmetic, not forecast — "the U.S. is facing a mammoth deficit of electricity over the next decade," with the supply menu limited to "natural gas, nuclear, and coal." The EIA series (1950–2025) carries the point: total US generation plateaued near 4,100 bkWh around 2007, ran flat for roughly fifteen years and has only just inflected up to ~4,400 — so the fleet, workforce and permitting pipeline were all sized for a flat total. Underneath that flat aggregate the mix churned without adding capacity: coal ~2,000 → ~650, gas <500 → ~1,900, nuclear flat at ~800 since the early 1990s. Hay's read: gas share is "steadily increasing" while "nuclear and, particularly, coal have been shrinking," leaving nuclear as the underbuilt and "emission-free" answer. 2026-SEP-10 (Toby McKenna, CEO Rockpoint Gas Storage): an operator's discount on the headlines — "there's more discussion about the demand than there has actually been FID projects" (on Meta's $13bn Alberta announcement). What survives the discount is the structure: data centres want 99.9% redundancy, so gas peakers backstop the load even when renewables supply it — "the battery backup or the storage backup to back up the real battery which is the power peaker." ~10 GW in the Alberta queue; 12.5 GW arrived in California in a single quarter. His conservative underwriting still gets LNG 20→30 Bcf and Alberta power doubling in a decade. 2026-SEP-10 (Ciampaglia, Sprott): hyperscalers moved from renewables to nuclear after running into intermittency. They bring much-needed capital for nascent small reactors and sign long-term, high-value PPAs (Google with a Finnish utility). High electricity prices let utilities invest after two decades without load growth, with reshoring adding demand; base-load power is 'paramount' for AI data centers even if the OpenAI/Anthropic IPOs disappoint. 2026-SEP-15 (App Economy Insights, SB Energy S-1): PORTS-Pike (8 GW for OpenAI, first phases 2028) alone needs 9.2 GW of new gas generation; the filing warns gas-turbine lead times can stretch to seven years. 'Power is both the moat and the bottleneck.' Parag Sanghani (Westwood, 2026-Sep-17): US power demand grew 0.1%/yr 2000–2020 and now ~5%/yr on 1,300 GW installed (data centers + reshored factories) — >$1T to be spent over the next decade across fuel, transportation, generation and energy tech; data-center siting shifting from Virginia to Texas/Ohio. Energy (~3–3.5%) and utilities (<3%) are barely in the S&P, so the power chain is the under-owned way to play AI. Halftime (2026-SEP-17), Flatt: power demand rose first on electrification, then digitalization, only lately AI factories — "the biggest impediment to the data centres being built... is power" (nuclear 5–7 years, gas 4, solar 2.5, wind 3–4); grids constrained almost everywhere, a slowdown would just let supply catch up; DOE granted Brookfield a 5 GW site in Kentucky. Chips are becoming "an investment asset class" financeable like real estate via the $500B Nvidia plan, with concrete deals now. 2026-SEP-21 Hay: the AI-power trade gets a contract. Generac's 8-K Amazon agreement covers up to $8B of data-center backup generators through 2033 ($2.4B in 2027-28), with Amazon able to buy up to 1.7M GNRC shares at $200.93 only as it places orders. Generac's data-center backlog is $1.6B and it is tripling large-megawatt capacity. Hay also calls cheap U.S. gas (EQT) "among the best AI plays" for those averse to paying lofty tech multiples.