Datacenter power stocks: the second-derivative AI trade

When people talk about the "AI trade," they almost always mean compute — $NVDA, $AMD, the chip stack. That trade is crowded, beta-heavy, and concentrated in single-name idiosyncratic risk. The cleaner expression sits one layer downstream of the silicon: the electricity that runs the silicon.

The QuantAbundancia taxonomy calls this bubble Datacenter Power. Constituents: $VST, $CEG, $NRG, $GEV, $ETR, $SO, $AEP. Residualized intra-cluster correlation: ~0.55 over a 252-day window. Validated.

That number is lower than Quantum's 0.76 or Semi Equipment's 0.82 — but the bloc is broader (7+ liquid names instead of 4-5), the demand signal is verifiable in publicly-disclosed PPA tonnage, and the constituents are cash-generative incumbents rather than pre-revenue speculations. Different risk profile, same thesis.

Why the bloc clusters at 0.55 and not higher

The structural reason: AI campuses are local. A 1 GW hyperscaler campus in Texas is served by ERCOT generators (mostly $VST, $NRG). The same campus design in PJM is served by $CEG, $VST, the eastern utilities. In the Southeast it's $SO + $ETR. In MISO/SPP it's $AEP.

Each region has its own capacity auction, its own regulator, its own queue interconnection rules, and its own coal-retirement schedule. The demand signal is uniform — hyperscalers signing 10-15 year PPAs at unprecedented GW scale — but the price response varies by region.

This is why the residualized correlation is 0.55 instead of 0.80. The bloc shares one demand signal split across roughly five regulatory regimes. Compare with Semi Equipment at 0.82 where four companies sell to the same handful of foundries on the same global cycle — single regime, perfect cluster.

The demand signal — what's actually driving the bloc

Three numbers anchor the thesis:

1. PJM 2025/26 capacity auction cleared at $269.92/MW-day. That's 9× the prior year's $28.92. The auction prices forward capacity; this print is the market's first honest reaction to AI-driven load growth in the largest US grid. Existing generators with PJM exposure ($VST, $CEG, $NRG) reprice their capacity revenue stream by roughly the same 9× multiplier on the affected fleet share.

2. EIA's 2024 datacenter electricity demand projection went from 4.4% of US grid load to 9-12% by 2028. Roughly a doubling in four years. That is not a forecast that historically plays out gracefully through a grid that takes 5-7 years to permit new transmission lines.

3. Aggregate hyperscaler capex commitments for 2026 alone are ~$685B (per Hyperscalers article). A non-trivial fraction of that flows into multi-year PPAs with the bloc, locked at premiums to spot to secure capacity ahead of competing buyers.

The bloc isn't expressing "AI is a story." It's expressing a measurable transfer of capital from hyperscaler balance sheets to generator P&Ls, locked over decade-plus contracts. That's why the residualized correlation holds even when the broader compute trade rotates.

The constituents, by exposure

Not every name in the bloc has the same exposure profile. The cleanest mental model is to bucket them by what the AI demand specifically does to their P&L.

The IPPs — pure capacity-price beta

$VST and $NRG are independent power producers. They sell wholesale into competitive markets (mostly ERCOT and PJM). When capacity auctions clear higher, their revenue rises by roughly the auction multiplier × their cleared MW. There's no ratepayer-protection mechanism that forces them to share the upside with consumers. They're the highest-beta expression of the AI-load thesis.

$VST's 2024-2025 chart is the cleanest single-name expression of the entire datacenter-power thesis: from ~$30 in early 2023 to a high north of $200 by mid-2025, driven primarily by its Texas capacity contracted to AI campuses. The narrative was correct early; the price caught up violently.

The nukes — pure baseload + sentiment

$CEG is the only US-listed pure-play nuclear operator. Its 2024 deal to restart Three Mile Island Unit 1 under a 20-year Microsoft PPA — at premium pricing, with $MSFT as a credit-quality counterparty — is the single most concrete data point in the entire AI-power thesis.

The PPA template that deal established (large hyperscaler + restart-or-new nuclear + 20-year fixed pricing + premium-to-spot) is now being copied. $CEG's remaining fleet has roughly 5x the Three Mile Island capacity available for similar contracting. Each subsequent announcement of this template repeats marks the bloc higher.

The nuclear-restart story bleeds into the Nuclear / SMR bubble ($OKLO, $NNE, $SMR), but those are pre-revenue speculations on technology that doesn't exist yet. $CEG has the reactors today. Different risk class, same upstream demand.

The integrated utilities — slowest beta, longest duration

$SO, $ETR, $AEP. Regulated utilities with rate-base growth as their primary value driver. They benefit from AI load less directly than the IPPs because their returns are capped at the regulated allowed-ROE — but their capex deployment opportunities expand dramatically when load growth justifies new generation, transmission, and substations.

Translation: their earnings growth rate inflects from ~5% historical to potentially ~7-9% as the rate base expands to serve AI campuses. That doesn't move the stocks 50% in a year like $VST, but it durably re-rates them by 1-2 PE turns over the cycle.

The lower beta is exactly why these names cluster less tightly with the IPPs in the residualized matrix. They share the demand signal but absorb it through a different mechanism (regulated rate base, not market-cleared capacity prices).

The picks-and-shovels — $GEV

GE Vernova spun out of GE in 2024. It builds the gas turbines, grid equipment, and increasingly the SMR architectures that get deployed when the bloc expands its supply. It's the single cleanest pure-play on new capacity additions across the entire bloc, regardless of which generator buys the equipment.

Higher residualized correlation to the rest of the bloc than its standalone fundamentals would suggest, because every constituent of the bloc that adds capacity adds it through equipment that $GEV sells.

Why this is a second-derivative trade

The first derivative of "AI is real" is compute — $NVDA and the chip stack. It's already crowded. Position sizing in NVDA-as-AI-proxy requires a view on margin trajectory, Blackwell shipping cadence, and competitive dynamics with AMD/Intel — three additional forecasting problems beyond "is AI demand real."

The second derivative is power. It strips out:

• The semi-cycle margin debate (you don't care if NVDA's gross margin is 70% or 80% — the GPU still gets plugged in somewhere)
• The China export-control risk (US AI campuses get power from US generators regardless of who supplies the chips)
• The hyperscaler-capex-cut risk (PPAs are 10-15 year commitments; even a 30% capex pullback doesn't unwind contracted capacity)

What's left is closer to a pure read on AI campus deployment volume × electricity content per campus. Both numbers are growing. The bloc captures both.

Catalysts that move the bloc

Four input categories drive the bloc's residualized signal:

• Capacity auction prints. PJM (May/June), MISO (April), ERCOT and others. Each clearing reprices the future cash-flow stream of every IPP with exposure. Ahead of an auction, sentiment runs into the print; clearing prints either confirm or break the trend.
• PPA announcements. A new $MSFT or $GOOGL nuclear or natural-gas PPA at premium pricing marks the entire bloc up. The 2024 Three Mile Island deal is the canonical example.
• Federal policy on transmission and permitting. Anything that shortens the queue interconnection backlog (currently 3-7 years in many regions) is a positive shock for incumbents who are already on the grid. Anything that tightens emissions rules is a negative shock to coal-heavy fleets, but the bloc has retired most coal exposure already.
• Earnings revisions on contracted-load disclosures. Quarterly disclosures of "MW under contract for new datacenter loads" from $VST, $CEG, $SO, $AEP are the highest-signal data point each cycle. The number is small relative to total fleet today; its growth rate is the signal.

How to size the bet

A 0.55 residualized correlation has specific implications for sizing:

Treat the bloc as one position, but with internal structure. Unlike Quantum (0.76, where the names are interchangeable), Datacenter Power has meaningful within-bloc dispersion. $VST can be up 80% in a year while $SO is up 12%. Both are correct expressions of the thesis at different beta levels.

The clean construction:

• Higher-beta core (50-60% of allocation): $VST, $CEG. Pure-play IPP + pure-play nuclear. Captures most of the upside from PPA prints and capacity-auction clears.
• Mid-beta diversifier (20-30%): $NRG, $GEV. ERCOT-focused IPP + the capex-deployment picks-and-shovels.
• Lower-beta ballast (15-25%): $SO, $ETR, $AEP. Regulated utility exposure with longer-duration rate-base re-rating.

This construction concentrates exposure where the empirical residualized correlation is tightest (the IPPs cluster with each other above 0.65 within the bloc) while preserving downside structure through the lower-beta utilities.

ETF shorthand. No US-listed ETF cleanly tracks this bloc. $XLU is the broad utilities ETF — about 30% of its weight is in the bloc, but the rest is regulated utilities with no AI-load exposure. $NLR (uranium) and $URNM capture nuclear-fuel exposure but not the operators. The bloc is currently better expressed by an equal-weight basket of the seven names than by any single ETF.

What can break the thesis

The honest counter-cases:

1. AI capex pullback. If the major hyperscalers cut 2027+ capex by ~30%+, PPA renegotiation becomes a real risk. Current contracts are mostly locked, but new contracting flow (which drives the forward cash-flow growth that's already in the price) slows materially. Bloc downside in this scenario: ~30-40%.

2. Permitting reform actually works. The bloc is partly priced on scarcity — the 5-7 year queue interconnect backlog protects incumbents. If federal policy compresses that to 2-3 years, new entrants compete down capacity prices. This is a tail risk, not a base case (US permitting reform has a poor track record), but worth flagging.

3. AI compute efficiency gains compress demand. If a generation of GPUs delivers 5x the throughput per watt (architectural shift, not Moore's law-style incremental), the load-growth projection compresses. The current EIA 9-12% datacenter share by 2028 assumes mostly current efficiency curves continuing.

None of these scenarios are immediate. All three are structurally bounded — the bloc has roughly 18-24 months of "the demand signal is overwhelming and there's no near-term offset" before any of them could plausibly materialize.

Why the bloc is underrated by retail

The retail-trader cliché is to play AI through $NVDA or a thematic ETF. Datacenter Power doesn't show up on most "AI stocks 2026" listicles because:

• The constituents have boring industry classifications (utilities, IPPs)
• They don't appear in any thematic AI ETF marketing
• The narrative requires understanding capacity-market mechanics, which most coverage doesn't bother explaining
• The biggest single-name move ($VST) has already happened — the easy multi-bagger is in the rearview

The opportunity now is the broadening — $CEG still has 4-5x the Three Mile Island contracting potential available, $NRG's ERCOT contracted-MW disclosures are still ramping, $SO + $AEP haven't begun the rate-base inflection yet. The first wave was a 2024 phenomenon. The bloc-wide rerating is a 2026-2028 phenomenon.

Bottom line

Datacenter Power is the validated mid-tier bubble in the QuantAbundancia taxonomy that captures the AI buildout one layer downstream of the silicon. The residualized correlation (~0.55) is lower than the marquee blocs (Quantum at 0.76, Semi Equipment at 0.82) but covers a broader, more diversified roster of cash-generative incumbents.

The trade structure is cleaner than the compute trade because it strips out the chip-cycle margin debate, the China export-control risk, and the GPU-architecture competitive dynamics — leaving roughly pure exposure to "how many AI campuses get built × how much electricity each consumes." Both numbers are growing on multi-year contracts already signed.

For a 5-year thematic position in the AI buildout, the bloc is a more durable expression than the compute trade and a more honest one than the failed Hyperscalers basket.

---

Browse all 12 bubbles ranked by empirical realness for context on where Datacenter Power sits, or read the methodology piece on residualized correlation for how the 0.55 number is computed.