Recursive self-improvement in AI is beginning. The path to AGI requires exponentially more compute, which requires exponentially more power. Every link in the energy supply chain — from fuel extraction to electron delivery — faces a demand shock unlike anything since industrialization. Key dynamics:
| # | Narrow Sector | AGI Demand Impact | Supply Constrained? | # Companies | Verdict |
|---|---|---|---|---|---|
| 1 | Natural Gas Producers | Strong | Moderate | 8 | HIGH |
| 2 | Natural Gas Pipelines & Midstream | Strong | High (permitting) | 8 | HIGH |
| 3 | Nuclear — Uranium Miners | Very Strong | Very High | 6 | HIGH |
| 4 | Nuclear — Reactors, SMR & Fuel | Very Strong | Extreme | 7 | HIGH |
| 5 | Grid Equipment (Transformers, Switchgear) | Very Strong | Extreme | 7 | HIGH |
| 6 | Electrical Utilities | Strong | Moderate | 10 | HIGH |
| 7 | Independent Power Producers (IPPs) | Very Strong | High | 6 | HIGH |
| 8 | Diesel/Gas Generators & Backup Power | Strong | Moderate | 4 | HIGH |
| 9 | Power Electronics & Power Conversion | Strong | Moderate | 5 | HIGH |
| 10 | Battery & Energy Storage | Moderate-Strong | Moderate | 5 | MEDIUM |
| 11 | Solar & Wind / Renewables | Moderate | Low-Moderate | 7 | MEDIUM |
| 12 | Fuel Cells | Moderate | Low | 4 | MEDIUM |
| 13 | Geothermal | Moderate | High (geology) | 2 | MEDIUM |
| 14 | Cooling & Thermal Management | Very Strong | High | 4 | HIGH |
| 15 | Electrical Infrastructure Services (EPC/Construction) | Strong | High (labor) | 5 | HIGH |
E&P companies extract natural gas (and associated liquids) from shale basins and conventional reservoirs. They sell gas into wholesale markets and via long-term contracts. The US is the world's largest producer, dominated by Appalachian (Marcellus/Utica), Permian, and Haynesville basins.
Demand: Natural gas is the primary fuel for new gas-fired power plants being built to serve data centers. Every 1 GW of gas-fired capacity consumes roughly 150-180 MMcf/d. With 30-60 GW of new gas capacity needed by 2030, incremental gas demand could reach 5-10 Bcf/d — a 5-10% increase over current US production. This is a structural, multi-year demand tailwind.
Supply: US gas supply is abundant but not unconstrained. Pipeline takeaway from Appalachia remains tight. Capital discipline among producers (post-2020 cash flow focus) means supply will ramp slower than demand. LNG export growth adds competing demand. Net effect: sustained higher gas prices benefit producers with low-cost, high-volume positions.
Midstream companies gather, process, transport, and store natural gas via pipeline networks. They earn fee-based or commodity-linked revenues. Pipelines are natural monopolies with high barriers to entry — permitting a new interstate pipeline takes 3-5+ years.
Demand: Every molecule of incremental gas burned for AI-driven power needs a pipeline to carry it. Data centers are being sited in Virginia, Texas, and the Midwest — all requiring expanded pipeline capacity. Pipeline utilization rates are already at multi-year highs.
Supply: Severely constrained by permitting (NEPA, state-level opposition). The Mountain Valley Pipeline took a decade to build. New capacity additions will lag demand, supporting higher tariffs and utilization for incumbents. This is perhaps the most supply-constrained link in the gas value chain.
Uranium mining companies extract U3O8 from deposits via open-pit, underground, or in-situ leach (ISL) methods. The ore is processed into yellowcake, then sold to utilities and fuel fabricators for enrichment and use in nuclear reactors. Uranium is a tiny fraction of nuclear fuel cost but 100% essential — there is no substitute.
Demand: Nuclear is the premier 24/7 zero-carbon baseload source, making it ideal for AI data centers. Microsoft, Google, Amazon, and Oracle have all signed nuclear PPAs or announced nuclear-powered data center plans. Global uranium demand is ~180M lbs/yr but new reactor builds and restarts (Three Mile Island, Palisades) are adding incremental demand. AGI drives this further as every hyperscaler seeks firm, carbon-free power.
Supply: Extremely constrained. Primary mine production covers only ~75% of reactor demand; the deficit has been filled by secondary supplies (government stockpiles, underfeeding) that are depleting. New mines take 7-10 years to permit and build. Kazakhstan (40% of global supply) faces geopolitical risk. Western uranium production is minimal. The supply-demand gap is structural and widening.
This segment includes companies that design/build nuclear reactors (large-scale and small modular reactors), provide nuclear fuel fabrication and enrichment, and offer reactor services (maintenance, decommissioning, engineering). SMRs (Small Modular Reactors) are factory-built reactors in the 50-300 MW range designed for faster deployment and siting flexibility.
Demand: Hyperscalers are desperate for firm, carbon-free power. SMRs are the technology most frequently cited in data center power announcements. Kairos Power, X-energy, NuScale, and others have signed agreements with Google, Dow, and utilities. The US DOE is actively funding advanced reactor deployment. Demand for nuclear services at existing plants (life extensions, uprates) is also surging.
Supply: Extremely constrained. No new commercial reactor has been completed in the US since Vogtle (2023-2024). SMRs are still pre-commercial — NuScale received NRC design certification but canceled its first project on cost overruns. The nuclear supply chain (forgings, qualified welders, ASME-certified components) has atrophied over decades. Rebuilding it will take years. First-mover advantage is enormous.
Grid equipment manufacturers build the hardware that steps voltage up/down (transformers), routes power (switchgear), and protects circuits (breakers). Large power transformers (LPTs) are custom-built, weigh 100-400 tons, and take 12-18 months to manufacture. They are the physical bottleneck of the grid.
Demand: Every new data center needs substation transformers, medium-voltage switchgear, and circuit protection. A 100 MW data center campus requires multiple large transformers. With 30-60 GW of data center capacity being added, transformer demand is surging. Utilities are also upgrading aging grid infrastructure simultaneously.
Supply: This is the single most supply-constrained sector in the entire energy value chain. Lead times for large power transformers have stretched from 12 months to 3-4+ years. There are only a handful of global manufacturers. Grain-oriented electrical steel (GOES), the key input, has limited global capacity (mostly in Japan and South Korea). New transformer factories take 2-3 years to build. Hyperscalers are now pre-ordering transformers and even acquiring transformer companies.
Regulated and semi-regulated utilities generate, transmit, and distribute electricity to end customers. They earn returns on invested capital (rate base) approved by state regulators. Utilities with large data center customers in their service territories are seeing unprecedented load growth after two decades of flat demand.
Demand: US electricity demand had been flat for 15 years. AI data centers have broken that trend. Dominion Energy (Virginia), AEP (Ohio/Texas), and Duke (Carolinas) are seeing the largest data center interconnection requests. Some utilities are projecting 5-10% annual load growth in their service territories — a rate not seen since the 1960s. This drives massive capital investment in generation, transmission, and distribution.
Supply: Utilities can build to meet demand, but regulatory approval, siting, and construction take 3-7 years. The key question is whether regulators will allow utilities to socialize data center infrastructure costs across all ratepayers, or require hyperscalers to pay directly. Either way, the rate base growth is real and substantial.
IPPs own and operate power plants that sell electricity into wholesale markets or via bilateral contracts (PPAs). Unlike regulated utilities, IPPs are fully exposed to merchant power prices and contract negotiations. Their assets include gas-fired plants, nuclear plants, and renewables.
Demand: IPPs are the most direct beneficiaries of rising wholesale power prices driven by AI load growth. PJM (the Mid-Atlantic grid operator) capacity auction prices have surged to record levels. Hyperscalers are signing 10-20 year PPAs at premium prices for firm power. Nuclear IPPs are especially sought after for carbon-free baseload.
Supply: Existing fleet is finite and takes years to expand. Many gas plants are being fully contracted. Nuclear plants are unique, irreplaceable assets. New combined-cycle gas plants take 3-4 years to build. IPPs with existing capacity in data center corridors hold scarce, appreciating assets.
These companies manufacture reciprocating engine generators, gas turbines, and uninterruptible power supplies (UPS) used for backup, peaking, and increasingly primary power at data centers. Products range from small diesel gensets to large natural gas reciprocating engines and industrial gas turbines in the 5-500 MW range.
Demand: Every data center needs backup power (N+1 or 2N redundancy), which means diesel or gas generators for every MW of IT load. But increasingly, hyperscalers are deploying gas turbines and reciprocating engines as primary, behind-the-meter generation to bypass the grid entirely. This converts generators from backup equipment to baseload infrastructure. Demand for large gas turbines has surged — GE Vernova's heavy-duty gas turbine orders are at multi-decade highs.
Supply: Gas turbine manufacturing has limited global capacity (GE Vernova, Siemens Energy, Mitsubishi). Lead times for heavy-duty gas turbines are 2-3 years. Reciprocating engine supply (Caterpillar, Cummins, Generac) is less constrained but still tight due to simultaneous demand from data centers, grid peaking, and backup.
Power electronics companies make the components and systems that convert, regulate, and manage electrical power: UPS systems, power distribution units (PDUs), voltage regulators, power semiconductors (SiC, GaN), DC-DC converters, and busbar systems. Every watt consumed by an AI chip must first pass through power conversion and distribution equipment.
Demand: AI GPU racks consume 40-120 kW per rack (vs. 5-10 kW for traditional servers). This 10x increase in power density requires completely re-architected power distribution — higher-voltage distribution, more efficient conversion (SiC/GaN switching), liquid-cooled PDUs, and 48V rack architectures. Every dollar spent on GPUs requires a proportional increase in power electronics. The shift from 12V to 48V and the adoption of silicon carbide (SiC) MOSFETs is a secular growth driver.
Supply: SiC wafer capacity is constrained (dominated by Wolfspeed, II-VI/Coherent, and international players). High-end UPS and PDU manufacturing has moderate lead times. The shift to higher power densities is creating design wins for companies with next-gen power conversion technology.
Energy storage companies design and deploy utility-scale battery systems (primarily lithium-ion, increasingly iron-air and other chemistries) that store electricity for dispatch during peak demand, provide grid ancillary services (frequency regulation), and enable renewable integration. Systems range from 1 MWh to 1+ GWh.
Demand: Battery storage is needed to firm up renewable generation for data centers and to provide grid stability as loads grow. Utilities and IPPs are deploying storage at record rates. However, batteries alone cannot provide 24/7 baseload power for data centers — they are complements to, not substitutes for, gas and nuclear. Battery demand is strong but the AI-specific pull is indirect.
Supply: Lithium-ion cell manufacturing is scaling rapidly (mostly in China), and battery costs continue to decline. US-based manufacturing is growing via IRA subsidies. Supply is less constrained than most other energy sectors. Competition is intense, especially from low-cost Chinese manufacturers.
Renewable energy companies develop, manufacture, and operate solar panels, wind turbines, and associated balance-of-system equipment. Solar includes utility-scale ground-mount, rooftop, and tracker systems. Wind includes onshore and offshore turbines. Revenue comes from equipment sales, project development, and long-term PPAs.
Demand: Hyperscalers have enormous renewable energy procurement targets (Google, Microsoft, Amazon are the largest corporate renewable buyers globally). Solar and wind PPAs are being signed at unprecedented volumes. However, renewables face a fundamental mismatch with AI workloads: intermittency. A data center needs power 24/7/365 at near-100% uptime — solar produces ~25% of the time, wind ~35%. Renewables will be part of the mix but cannot be the primary power source for AI without massive storage or gas backup.
Supply: Solar panel manufacturing is massively oversupplied globally (Chinese overcapacity). Module prices have crashed. Wind turbine manufacturers (Vestas, GE Vernova, Siemens Gamesa) have struggled with profitability. Supply is generally not constrained, which limits pricing power. The exception is US-manufactured solar (IRA domestic content bonuses) and solar trackers.
Fuel cell companies make electrochemical devices that convert hydrogen (or natural gas) directly into electricity without combustion. Technologies include PEM (proton exchange membrane), SOFC (solid oxide), PAFC (phosphoric acid), and MCFC (molten carbonate). Applications range from backup power to stationary baseload generation and transportation.
Demand: Fuel cells are being considered for data center backup and primary power, especially in locations where grid power is unavailable or unreliable. Microsoft has tested hydrogen fuel cells for data center backup. Natural gas fuel cells (like Bloom Energy's SOFCs) can provide on-site baseload power. However, hydrogen infrastructure is nascent, and natural gas fuel cells compete with cheaper gas turbines and reciprocating engines. Adoption is growing but from a very small base.
Supply: Not significantly constrained. Fuel cell companies have excess manufacturing capacity and are scaling up. The challenge is cost competitiveness, not supply. Most fuel cell companies remain unprofitable.
Geothermal energy harnesses heat from the earth's interior to generate electricity or provide direct heating. Conventional geothermal taps naturally occurring hydrothermal reservoirs. Enhanced/advanced geothermal systems (EGS) use techniques borrowed from oil & gas fracking to create artificial reservoirs, dramatically expanding the addressable geography beyond volcanic regions.
Demand: Geothermal is the only renewable that provides true 24/7 baseload power with a capacity factor of 90%+, making it theoretically ideal for data centers. Google has signed a geothermal PPA with Fervo Energy (private). The technology is gaining attention as a complement to nuclear for always-on clean power.
Supply: Constrained by geology and technology maturity. Conventional geothermal is limited to specific regions (Western US, Iceland, etc.). EGS technology is still early-stage and expensive, though costs are declining rapidly as oil & gas drilling expertise is applied. Very few public pure-play geothermal companies exist.
Cooling companies manufacture the systems that remove heat from data centers: precision air conditioning (CRAC/CRAH units), chillers, cooling towers, rear-door heat exchangers, and increasingly liquid cooling systems (direct-to-chip and immersion cooling). As AI GPU power density has soared, traditional air cooling has become insufficient, driving rapid adoption of liquid cooling.
Demand: An NVIDIA GB200 NVL72 rack consumes ~120 kW and must be liquid-cooled — air cooling cannot handle the heat density. Every next-generation AI training cluster requires custom coolant distribution units (CDUs), cold plates, and heat rejection systems. The liquid cooling market for data centers is expected to grow 30-40% annually. This is not optional — without cooling, the GPUs cannot run.
Supply: Liquid cooling for data centers is a nascent market with limited qualified suppliers. Vertiv, Schneider (private), and Cooltera (acquired by Vertiv) have been scaling capacity. The CDU and cold plate supply chain is immature and constrained. Traditional HVAC companies (Trane, Carrier) are entering the market but lack data center-specific expertise.
Electrical construction and engineering firms design, build, and maintain power infrastructure: substations, transmission lines, data center electrical systems, and industrial power installations. They provide the labor and project management to turn equipment (transformers, switchgear, generators) into functioning infrastructure. Services include engineering, procurement, construction (EPC), and ongoing maintenance.
Demand: Every GW of new data center capacity requires massive electrical construction — substations, switchyards, distribution systems, backup power installation, and grid interconnection. Utility transmission and distribution capital spending is also surging. The total addressable market for electrical infrastructure services is growing 15-25% annually.
Supply: Severely constrained by skilled labor shortages. Electricians, linemen, and substation technicians take years to train. The workforce is aging (average lineman age ~45). Immigration restrictions limit labor supply growth. This labor bottleneck may be the most underappreciated constraint in the entire AI power build-out. Companies with large, trained workforces have a durable competitive advantage.
98 ticker mentions across 15 sectors. After deduplication (GEV, ETN, VRT, CEG, VST, LEU appear in multiple sectors): 74 unique companies.
| Natural Gas Producers | EQT, AR, RRC, SWN, CNX, CHK, CTRA, CRK |
| Gas Pipelines & Midstream | WMB, KMI, ET, EPD, MPLX, DTM, AM, TRGP |
| Uranium Miners | CCJ, UEC, UUUU, NXE, DNN, LEU |
| Nuclear Reactors/SMR/Fuel | SMR, OKLO, LEU*, BWX, GEV*, CW, FLR |
| Grid Equipment | ETN, GEV*, HUBB, SPXC, AIT, POWL, ATKR |
| Electrical Utilities | VST, CEG, NEE, D, DUK, SO, AEP, SRE, EXC, PEG |
| IPPs | VST*, CEG*, NRG, TALEN, CPN, RNW |
| Generators & Backup | GEV*, CAT, CMI, GNRC |
| Power Electronics | VRT, ETN*, WOLF, AEIS, VICR |
| Battery & Storage | FLNC, FLUX, STEM, ENVX, QS |
| Solar/Wind/Renewables | FSLR, ENPH, SEDG, NEP, ARRY, NOVA, SHLS |
| Fuel Cells | BE, PLUG, FCEL, BLDP |
| Geothermal | ORA, GTHM |
| Cooling & Thermal Mgmt | VRT*, TT, CARR, LII |
| Electrical Infra Services | PRIM, MTZ, EME, PWR, FIX |
* = appears in multiple sectors