Data centers consume 4% of US electricity today. By 2030, that figure reaches 9-12%. A single NVIDIA Blackwell rack draws 120kW. No Electrons = No Intelligence.
US electricity demand was essentially flat for two decades — a plateau that led to chronic underinvestment in generation and grid infrastructure. That era is over. The convergence of AI data centers, electric vehicles, industrial reshoring, and electrification of heating is creating what the North American Electric Reliability Corporation (NERC) calls the "most significant load growth in a generation."
Data centers are the primary demand driver. US data center power consumption reached approximately 35 GW in 2024, representing roughly 4% of total US electricity generation. The Electric Power Research Institute (EPRI) projects this will reach 80-100 GW by 2030, consuming 9-12% of US electricity. To contextualize: 65 GW of incremental demand is equivalent to adding the entire electricity consumption of France to the US grid in six years.
The power density of AI computing is staggering. A traditional cloud data center rack consumes 5-10 kW. An NVIDIA Blackwell GB200 NVL72 rack — the standard for GPT-5 class training — draws 120 kW. That is a 12-24x increase in power density. A single AI training run for a frontier model now consumes 10+ GWh of electricity, equivalent to powering 1,000 US homes for a year. Meta's planned 2 GW data center campus in Louisiana would consume more electricity than some small US states.
Source: EPRI, IEA, Goldman Sachs Research, NERC Long-Term Reliability Assessment, Market Watch estimates.
A traditional Google search query accesses indexed data and returns results — consuming roughly 0.3 Wh of electricity. A ChatGPT query runs a large language model with hundreds of billions of parameters through multiple transformer layers, performing trillions of floating-point operations — consuming roughly 3-10 Wh, or 10-30x more. But the real energy hog is training, not inference. Training GPT-4 reportedly consumed 50-100 GWh. Training GPT-5 or equivalent frontier models is estimated at 200-500 GWh. As models scale toward 10 trillion parameters and multimodal architectures (video, audio, robotics), energy consumption scales superlinearly. This is why every major AI lab is now hiring energy procurement teams — the bottleneck has shifted from talent and chips to raw electricity.
The AI industry's insatiable demand for 24/7 carbon-free baseload power has triggered a nuclear renaissance that seemed unthinkable five years ago. Solar and wind are cheap but intermittent (capacity factors of 25-35%). Natural gas is reliable but emits CO2, conflicting with hyperscaler "net zero" commitments. Nuclear is the only energy source that is simultaneously carbon-free, baseload (90%+ capacity factor), and scalable. The market has noticed: Constellation Energy (CEG), the largest US nuclear fleet operator, has risen from $80 to $300+ since 2023.
The deals are coming fast. Microsoft signed a 20-year PPA with Constellation Energy to restart Three Mile Island Unit 1 (the unit that did NOT melt down) — a $1.6 billion restart project for 837 MW of carbon-free power. Amazon acquired a nuclear-powered data center campus from Talen Energy near the Susquehanna nuclear plant in Pennsylvania (960 MW). Google signed the first-ever corporate PPA for Small Modular Reactor (SMR) power with Kairos Power, for 500 MW by 2035. Oracle announced plans for data centers powered by three SMRs totaling 1 GW+. France, never having abandoned nuclear, is building 14 new EPR2 reactors in addition to its existing 56 reactors.
| Deal | Buyer | Supplier / Asset | Capacity | Timeline | Status |
|---|---|---|---|---|---|
| Three Mile Island Restart | Microsoft | Constellation Energy (CEG) | 837 MW | 2028 | In Progress |
| Susquehanna DC Campus | Amazon (AWS) | Talen Energy | 960 MW | 2025-2026 | Active |
| Kairos SMR PPA | Kairos Power | 500 MW | 2030-2035 | Development | |
| SMR Data Centers | Oracle | Multiple SMR vendors | 1 GW+ | 2030+ | Announced |
| NuScale VOYGR | UAMPS (Utah) | NuScale Power (SMR) | 462 MW (6 modules) | 2029-2030 | NRC Approved Design |
| X-energy Xe-100 | Dow Chemical | X-energy (HTGR) | 320 MW (4 modules) | 2030 | Development |
| France EPR2 Program | EDF / French State | Framatome | 24.7 GW (14 reactors) | 2035-2050 | Planning |
| Holtec Palisades Restart | Holtec International | Palisades Nuclear (MI) | 800 MW | 2025-2026 | In Progress |
Source: Company announcements, NRC filings, EDF, World Nuclear Association, Market Watch compilation.
50-300 MW per module. Factory-built, truck-deliverable. NuScale first NRC-approved design. Kairos (molten salt), X-energy (HTGR), TerraPower (sodium-cooled). Modular = faster, cheaper, lower risk.
Three Mile Island (837 MW), Palisades (800 MW) being restarted. US fleet of 93 reactors running at 93% capacity factor. License extensions to 80 years becoming standard. Cheapest GWh on a LCOE basis.
Commonwealth Fusion Systems targeting 2030s demo. Helion (backed by Altman/MSFT) targeting 2028. TAE Technologies. Not investable yet, but the endgame for limitless clean energy.
Even if we build enough generation capacity, the electricity must reach the data centers. The US grid is a 70-year-old patchwork of aging infrastructure. The average large power transformer in the US is over 40 years old, well past its 30-year design life. There are approximately 700,000 miles of high-voltage transmission lines, many operating at or near capacity. The American Society of Civil Engineers gives the US energy grid a grade of C-minus.
The investment required is staggering. The Department of Energy estimates $2+ trillion in grid investment is needed by 2035 to meet rising demand and integrate renewables. McKinsey estimates that the US needs to add 47,000+ miles of new high-voltage transmission lines by 2035 — a 60% increase over the current system. Yet the average time to permit and build a new transmission line is 10-15 years. The bottleneck is not technology. It is regulatory permitting, NIMBYism, and the fragmented ownership of the US grid across 3,000+ utilities.
Large power transformers (LPTs) are the critical chokepoint. A single LPT costs $5-10 million, weighs 200-400 tons, and takes 18-36 months to manufacture. The US has only 2-3 domestic manufacturers (e.g., SPX Transformer Solutions). Most LPTs are imported from South Korea (Hyundai Electric), Germany (Siemens), and Japan (Hitachi). Current lead times exceed 3 years for some transformer models. If a transformer fails at a critical substation, the replacement timeline can cause months of power disruption.
Imagine building a new highway system. You can manufacture all the cars you want (generation), but if the roads (transmission lines) and interchanges (transformers/substations) cannot handle the traffic, everything grinds to a halt. The US grid is like a highway system built in the 1960s trying to handle 2030 traffic volumes. Generation is relatively easy to build. Transmission is the hard part. A new natural gas plant takes 2-4 years. A new nuclear plant takes 7-12 years. But a new long-distance transmission line takes 10-15 years because it crosses multiple states, requires dozens of permits, and faces NIMBY opposition at every turn. This is why companies like Eaton (ETN), which makes the transformers, switchgear, and grid equipment that form the backbone of the grid, are among the best positioned in the entire AI investment chain. The grid equipment market is entering a structural supercycle that will last 15-20 years.
Battery energy storage systems (BESS) solve the intermittency problem of solar and wind, and provide critical grid services (frequency regulation, peak shaving, backup power). Lithium-ion battery pack costs have fallen from $1,100/kWh in 2010 to $139/kWh in 2024 (BloombergNEF). The trajectory points to sub-$80/kWh by 2028, at which point grid-scale storage becomes cheaper than peaking natural gas plants on a fully loaded cost basis.
Global grid-scale battery deployments reached 45 GW / 100+ GWh in 2024, up from 16 GW in 2022. The standard is evolving from 2-hour to 4-hour duration systems, enabling solar+storage to serve evening peak demand (the "duck curve" problem). California alone has deployed 10+ GW of grid-scale batteries, which are now routinely providing 20-25% of evening peak power.
Beyond lithium-ion, next-generation chemistries are emerging. Form Energy has developed iron-air batteries capable of 100+ hour duration at an estimated $20/kWh — potentially revolutionary for seasonal storage. Sodium-ion batteries (CATL, HiNa) use abundant materials (no lithium, no cobalt) and are already shipping for grid applications in China at 10-20% lower cost. Flow batteries (ESS Inc., Invinity) offer 10,000+ cycle lifetimes ideal for daily grid cycling.
Source: BloombergNEF, NREL, LDES Council, Market Watch estimates.
California's electricity demand chart looks like a duck: midday demand appears low because solar panels are generating massively, pushing net demand (demand minus solar) to near zero. But at sunset (5-9 PM), solar generation drops to zero while demand peaks (everyone comes home, turns on AC, cooks dinner). This creates a "ramp" of 10-15 GW in 3 hours — the steepest daily load swing of any major grid. Without batteries, this ramp requires gas "peaker" plants that sit idle 80% of the time. With 4-hour batteries charged by midday solar, the duck curve is "eaten" — batteries discharge during evening peak, replacing peaker plants entirely. This is already happening: California's batteries regularly discharge 5+ GW during evening peaks, and the state has curtailed no gas peaker plants that have been displaced.
| Technology | Cost ($/kWh) | Duration | Cycle Life | Key Player(s) | Status |
|---|---|---|---|---|---|
| Lithium-ion (LFP) | $139 (2024) | 2-4 hours | 4,000-6,000 | CATL, BYD, Tesla Megapack | Dominant |
| Sodium-ion | $80-100 (2025E) | 2-4 hours | 3,000-5,000 | CATL, HiNa, Faradion (Reliance) | Shipping |
| Iron-air | ~$20 (target) | 100+ hours | 10,000+ | Form Energy | Pilot 2025 |
| Vanadium Flow | $300-400 | 4-12 hours | 20,000+ | Invinity, CellCube, Rongke | Niche Deployed |
| Iron Flow | $150-200 | 4-12 hours | 25,000+ | ESS Inc. | Commercial |
| Compressed Air (CAES) | $100-150 | 8-24 hours | 30+ years | Hydrostor, Apex CAES | Projects Planned |
Source: BloombergNEF, LDES Council, company disclosures, NREL Annual Technology Baseline, Market Watch estimates.
The AI energy thesis produces four distinct investment lanes: nuclear operators with contracted power, independent power producers in deregulated markets, grid equipment manufacturers riding the infrastructure supercycle, and solar+storage players capturing distributed generation. We identify the best-positioned company in each lane.
| Ticker | Company | Lane | Investment Thesis | Moat | Risk |
|---|---|---|---|---|---|
| CEG | Constellation Energy | Nuclear Operator | Largest US nuclear fleet (21 GW). Only carbon-free baseload at scale. Microsoft TMI restart deal. PPAs repricing higher as contracts roll. | Irreplaceable Assets | Low-Med |
| VST | Vistra Corp | IPP / Texas Power | Largest power producer in Texas (deregulated, highest AI DC growth). 6.4 GW nuclear (Comanche Peak). 41 GW total fleet. Aggressively signing hyperscaler contracts. | Texas Market Position | Medium |
| ETN | Eaton Corporation | Grid Equipment | Makes transformers, switchgear, UPS systems, power distribution for data centers and grid. $2T+ grid modernization TAM. Backlog at all-time highs. Order growth 20%+ YoY. | Critical Infrastructure | Low |
| FSLR | First Solar | Solar + Storage | Only US-manufactured utility-scale solar panels (thin-film CdTe). IRA beneficiary. 70+ GW backlog. Domestic manufacturing avoids tariff risk. Integrating storage solutions. | US Manufacturing | Medium |
| GEV | GE Vernova | Gas Turbines + Grid | Makes the HA-class gas turbines (most efficient globally) that back up renewable intermittency. Grid solutions division. Only 3 companies make large gas turbines globally. | Oligopoly | Medium |
Source: Company filings, EIA, market data, Market Watch analysis.
Energy infrastructure is one of the most durable AI investment themes because the demand is contractual (10-20 year PPAs), the supply is physically constrained (you cannot 3D-print a nuclear reactor), and the regulatory tailwinds (IRA, DOE loans, FERC reform) are bipartisan. We present four trade setups for a 12-24 month position trade horizon.
Constellation operates the largest US nuclear fleet with 21 GW of capacity across 12 states. Nuclear plants are irreplaceable assets — you cannot build new ones in under 10 years, and no one is decommissioning them anymore. The Microsoft TMI restart deal establishes a new pricing paradigm for nuclear power ($100+/MWh for clean energy attributes vs. $40-60 historical). As long-term PPAs roll off and reprice at higher rates, CEG's EBITDA could double from $3.5B to $7B+ by 2028. Entry at $280-320 captures the pullback from the $350+ highs and sits above the 200-day EMA.
Vistra is the largest power generator in Texas (ERCOT), which is the epicenter of US data center growth. Texas has more data center capacity under construction than any other state due to cheap land, low taxes, fast permitting, and abundant natural gas. ERCOT is a deregulated market — meaning Vistra sells power at market prices, not regulated rates. As AI demand tightens the Texas grid, power prices rise, and Vistra captures the upside directly. The 6.4 GW Comanche Peak nuclear plant provides carbon-free baseload. Entry at $140-160 captures the post-rally consolidation near the 50-day EMA.
Eaton is the "picks and shovels" play on the grid modernization supercycle. Every new data center needs switchgear, transformers, UPS systems, and power distribution units — Eaton makes all of them. The $2T+ grid investment required by 2035 provides multi-decade demand visibility. Eaton's electrical segment backlog is at all-time highs, with order growth exceeding 20% YoY. Margins are expanding as pricing power increases (supply-constrained market). This is an industrial compounder with 15%+ earnings growth visibility through 2030. Entry at $310-340 captures pullbacks to the rising 50-day EMA.
First Solar is the only US-headquartered, US-manufactured utility-scale solar panel maker. Its thin-film CdTe technology avoids the polysilicon supply chain dominated by China. With 70+ GW of contracted backlog and IRA manufacturing tax credits worth $10+/panel, FSLR has industry-leading visibility and margins. The data center buildout drives utility-scale solar demand as hyperscalers pair solar+storage to meet renewable energy commitments. Entry at $180-210 captures the lower bound of the 12-month range. The trade requires patience — FSLR is a 12-24 month hold as backlog converts to revenue.
Horizon: 12-24 months (position trade). Energy infrastructure is a slow-burn thesis — patience is required. Entry method: Scale in over 3-4 tranches on pullbacks to EMAs or horizontal support. Total energy allocation: Maximum 15-18% of portfolio. Individual position max: CEG 5%, VST 4%, ETN 4%, FSLR 4%. Key catalysts: Hyperscaler PPA announcements, FERC transmission reform, earnings beats on order growth, IRA funding disbursements. Beta awareness: CEG and VST trade at ~1.5x SPX beta (higher vol). ETN at ~1.1x. FSLR at ~1.3x. Correlation note: All four are positively correlated with the AI/data center capex cycle. For diversification, pair with defensive positions.
A nuclear incident anywhere in the world — even a minor one — could trigger regulatory backlash that delays or cancels new nuclear projects. Fukushima (2011) led Germany to shut its entire nuclear fleet. A similar reaction in the US would devastate CEG and the entire nuclear renaissance thesis. Probability: Low. Impact: Catastrophic.
If AI inference efficiency improves faster than expected (model distillation, quantization, neuromorphic chips), data center power demand could plateau below forecasts. DeepSeek's R1 model demonstrated that competitive AI performance is achievable with far fewer parameters and less compute. If this trend accelerates, the "power crisis" narrative unwinds. Probability: Medium. Impact: High (for VST, CEG valuations).
New transmission lines, substations, and power plants all require permits that can take 5-15 years to obtain. NIMBY opposition, environmental lawsuits, and regulatory disputes routinely delay or kill projects. If the grid cannot expand fast enough, data center buildouts stall regardless of demand. Probability: Very High. Impact: Medium (priced in somewhat).
The Inflation Reduction Act provides critical subsidies for nuclear (production tax credits), solar (manufacturing credits for FSLR), and batteries (ITC). A political shift could reduce or eliminate these incentives. However, IRA benefits disproportionately flow to Republican-leaning states, making full repeal politically difficult. Probability: Low-Medium. Impact: High (for FSLR especially).
Q1 2026: CEG, VST, ETN, FSLR earnings (order backlog updates). DOE loan program disbursements for nuclear. FERC transmission rule finalization. Q2 2026: Summer peak power season begins (VST earnings leverage). NuScale VOYGR construction milestones. TMI restart progress update. H2 2026: IRA manufacturing credit verification (FSLR). Hyperscaler capex budgets for 2027 announced (DC pipeline visibility). Grid-scale storage deployment data (annual). 2027-2028: First SMR power generation (NuScale). TMI restart completion (CEG). 50+ GW DC demand milestone (EPRI verification).