The product here is the most physical thing in the entire AI stack: firm electricity — actual electrons, sold by the megawatt (MW; one MW is roughly the power draw of about 800 average US homes, or a small slice of a large data center). Artificial-intelligence data centers are warehouses full of chips that run around the clock at near-full power, and they cannot run on anything else. The companies in this group own the things that make and deliver that power: nuclear and gas power plants (Constellation/CEG, Vistra/VST, Talen/TLN, NRG), a large regulated wires-and-poles utility (NextEra/NEE), and the factory that builds the turbines, generators and grid gear that plants are made of (GE Vernova/GEV). They earn cash by selling electricity, by selling the machines that make electricity, or by earning a regulated return on the grid they build. Given that AGI is arriving — meaning compute demand keeps compounding rather than levelling off — the questions for an owner are: how much firm power will the world want, how fast can it actually be built, and what does the market charge today to own the gap between the two. This sheet lays out those facts; it does not tell you whether to buy.
In plain terms: forecasts point to demand for firm electricity outrunning the ability to build it for several years, because the binding limits are physical and slow — power plants take roughly 3–7 years, transformers and large turbines are booked years ahead, and grid hookups are backed up. By those forecasts the product is short (more wanted than can be made), not long, for a multi-year window. The prices the market puts on the existing, already-built power assets (especially nuclear) have risen substantially, so an owner today is paying up-front for power that is in several cases contracted but not yet flowing as cash. This is a description of the trade-off, stated neutrally — the reader judges whether the price is worth it.
The product is electricity that is firm (available essentially all the time, on demand), sold as megawatts of capacity and megawatt-hours of energy (a megawatt-hour is one MW delivered for one hour). A data center signs up for, say, 500 MW and expects it 24 hours a day, every day — which is why intermittent solar and wind, on their own, do not qualify, and why nuclear and natural-gas plants are the assets being competed over.
There are three different money machines inside this one group, and they earn cash in different ways:
Translation to cash for an owner: with the generators you are buying a claim on the price of power and on long contracts being signed; with the utility you are buying a claim on a steady regulated return that grows as it spends; with the equipment maker you are buying a claim on the build-out itself — it gets paid whether the eventual electron is nuclear, gas or otherwise.
Where demand is today (grounded in the scan): US data centers drew roughly 20 GW of power in 2024. For about 15 years before AI, total US electricity demand was essentially flat — utilities planned around little-to-no growth. AI changed that. Some utilities now project 5–10% annual load growth in their territories, a pace the scan notes has not been seen since the 1960s.
Where demand is going (forecast — labelled as forecast): The scan projects US data-center draw rising to 50–80+ GW by 2028–2030, "driven almost entirely by AI training and inference clusters." That is roughly a tripling-to-quadrupling of this specific load in about five years. Separately, roughly 2,500 GW of total capacity sits in interconnection queues (the waiting list to plug new projects into the grid) — far more than will actually get built, but a signal of how much wants in. Treat the GW endpoints and the queue figure as approximate est..
The AGI lens: Reasoning from the premise that AGI is arriving, the demand line bends up rather than plateaus. Each capability jump pulls in more training compute, and — more importantly for power — inference (running the trained models for users and agents) scales with usage, runs continuously, and does not stop when training does. A later, additional wave — physical AI / humanoid robots electrifying factories and logistics — is not in these numbers; the scan explicitly notes "AGI robots are NOT yet deployed… the demand driver is pure compute." So the 50–80 GW forecast is a compute-only number; any robot wave would be incremental on top of it.
Who the buyers are: the hyperscalers — Microsoft, Google/Alphabet, Amazon, Meta, plus AI-natives such as OpenAI/Anthropic and smaller GPU-cloud providers ("neoclouds") — buying through 10–20 year PPAs, increasingly for carbon-free baseload (always-on power). The scan states they are willing to pay premium prices for firm, around-the-clock power, and several are contracting nuclear output directly (for example, Talen's Susquehanna nuclear plant has a direct data-center PPA; Constellation is recontracting nuclear capacity for hyperscaler load).
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Current capacity & expansion (grounded in the scan): The named generators already own large fleets. Vistra (VST) owns the largest competitive generation fleet in the US, including the Comanche Peak nuclear plant. Constellation (CEG) is the largest US nuclear-fleet operator. Talen (TLN) owns the Susquehanna nuclear plant, which has a direct data-center PPA. NRG runs a large Texas fleet. NextEra (NEE) pairs regulated Florida utility scale with the largest US renewables portfolio and is adding gas to serve data-center load. GE Vernova (GEV) is described as the dominant maker of heavy-duty gas turbines with the largest installed base globally, and its turbine order backlog is described as at multi-decade highs.
The main bottleneck: the scan frames supply as limited by physics and lead time, not by willingness to spend. Specifically:
Market-share structure (who controls supply): firm baseload power, especially nuclear, sits in relatively few hands. CEG (largest US nuclear operator), VST (largest competitive fleet), TLN and NRG together hold scarce, hard-to-replicate generating assets in the corridors where data centers are being sited. On the equipment side, GEV is one of three global heavy-turbine makers — a concentrated market (an oligopoly) on machines in high demand. NEE controls the largest US renewables capacity and a large regulated footprint. The structural feature the scan emphasizes: the assets that matter most cannot be quickly copied, which is the source of the scarcity. The exact market-share percentages are not in the provided files and are not asserted here.
Putting the two sides together: demand for firm power is forecast to roughly triple this decade while the supply of new firm power is gated by 3–7 year plant timelines and multi-year equipment lead times. If those forecasts hold, that mismatch is the definition of a short market — more wanted than can be made for a multi-year window. The demand side is a forecast; the supply-side lead times are grounded in the scan.
The signals the scan points to are prices, contracts and lead times:
| Measure | Demand side | Supply side | Net read |
|---|---|---|---|
| US data-center power | ~20 GW (2024) → 50–80+ GW (2028–30, forecast) est. | New firm plants ~3–7 yrs to build | Demand forecast to ramp faster than supply can be added |
| Firm / baseload (nuclear) | Hyperscalers want carbon-free around-the-clock power | Existing fleet roughly fixed; ~no new builds this window | Forecast short |
| Grid & turbine equipment | Each GW needs transformers + turbines | Lead times ~3–4 yrs; 3 global turbine makers | Largely spoken-for near-term |
| Wholesale capacity price | Rising load | Tight supply | Record PJM clears (grounded) |
When could it flip to oversupply? A flip would require new firm supply to actually arrive in bulk — the gas plants now being permitted finishing (mid-to-late decade), SMRs (small modular reactors, factory-built units in the 50–300 MW range) reaching commercial scale, and equipment lead times normalizing. Note the scan flags that SMRs are still pre-commercial today (no completed US commercial reactor since Vogtle in 2023–2024, and the supply chain has atrophied). A rough timing estimate is 2029–2032+ est., and it could move further out if AGI-driven demand keeps compounding (or if the physical-AI/robot wave lands on top of compute). Stated neutrally: by the scan's own framing the shortage is a dated, multi-year window, not permanent — and not near-term either. The exact flip year is not knowable from the provided files.
Below, the scan grounds what each company is and its competitive position; the dollar sizes (revenue, market value) are general-knowledge approximations, are not in the provided files, and are not live-verified — every dollar figure carries an est. tag and is given as a range, not a precise number.
| Company (ticker) | What it makes / sells | Exposure to firm power | Rough size est. | Position (grounded in scan) |
|---|---|---|---|---|
| Constellation (CEG) | Electricity from nuclear + other plants | Generator; essentially all revenue is power | ~$80–110B mkt cap; ~$20B+ rev est. | Largest US nuclear-fleet operator; carbon-free baseload sought by hyperscalers |
| Vistra (VST) | Electricity (gas, nuclear, solar, retail) | Generator; most revenue is power | ~$45–65B mkt cap; ~$17B+ rev est. | Largest US competitive generation fleet; owns Comanche Peak nuclear |
| Talen (TLN) | Electricity from nuclear + gas | Generator; essentially all revenue is power | ~$12–20B mkt cap; ~$2–3B rev est. | Susquehanna nuclear plant with a direct data-center PPA (a nuclear-to-AI contract) |
| NRG (NRG) | Electricity + retail energy (large Texas fleet) | Generator + retail; most revenue power-linked | ~$20–35B mkt cap; ~$28B rev est. | Large Texas (ERCOT) fleet; retail customer book on top |
| GE Vernova (GEV) | Gas turbines, grid gear, wind, SMR (in development) | Equipment, not electrons — exposed via the build-out, not the power price | ~$90–140B mkt cap; ~$35B rev est. | One of 3 global heavy-turbine makers; largest installed base; backlog at multi-decade highs |
| NextEra (NEE) | Regulated FL utility (FP&L) + largest US renewables | Diversified utility; data-center load is one slice of a much bigger base | ~$140–170B mkt cap; ~$25B+ rev est. | Largest US renewables portfolio; regulated rate-base growth model; adding gas for data-center load |
Reading the table as an owner, stated neutrally: CEG, VST and TLN derive essentially all revenue from the price of firm power; NRG is similar but with a retail business attached; GEV is a claim on the act of building (it is paid regardless of which fuel wins); NEE is the most diversified, where firm-power-for-AI is one part of a large, regulated, renewables-heavy business. Which of these shapes an owner prefers is a judgment, not stated here.
Two different money shapes sit in this group, and they are priced differently. All multiples below are approximate and not live-verified est. — they are rough orders of magnitude shown as ranges to illustrate shape, not exact readings. They are not in the provided files.
What you pay per $1 of this-year revenue (price-to-sales, a crude "what does a dollar of sales cost" gauge — market value divided by annual revenue):
The money-in / money-out shape:
Neutral arithmetic, no verdict: per the scan, the firm-power names are repricing from "bond-proxy utilities" toward "AI infrastructure," and the scan calls this repricing "still early." Factually, that means the market price of exposure has already moved up to reflect the expected demand-supply gap; an owner buying today is paying for power that is largely scarce-and-contracted but in several cases not yet fully flowing as cash. Whether that price is high or low relative to what the assets ultimately earn is left entirely to the reader.
A map of where a company-level deep-dive would add the most information, grouped by type. This is a map of where to look, not a recommendation of what to buy.
What was used:
/Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/01-energy-power.html — source of the company lists, sub-sector qualitative supply/demand notes, the ~20 GW → 50–80+ GW data-center forecast, lead-time figures, interconnection-queue size, the PJM record-capacity-price note, the "repricing still early" framing, and each company's described position. These are treated as grounded.Hard vs approximate (plain restatement):
Source: 500-stocks AGI Energy & Power scan (May 28, 2026); general knowledge (cutoff ~early 2026, not live-verified). No live web/quote retrieval was available when this fact sheet was built.