State of Data-Center Power
The binding physical ceiling on the AI buildout — grid interconnect queues, transformer/turbine lead times, and the liquid-cooling cutover, priced as toll-booths.
Own the deep-physics bottlenecks of the AI power crunch — firm generation (CEG/VST), turbines (GEV), and grid gear (ETN/PWR/HUBB) — where lead-time scarcity is durable pricing power; rent the compute/cloud layer, don't own it.
The AI power crunch has flipped the value-chain: with US data-center load heading to ~76 GW in 2026 (from ~50 GW in 2024, IEA/Bloom Energy 2026) against a grid that takes 100+ weeks to deliver a transformer and slots gas turbines out to 2030, the binding constraint is no longer compute — it is firm, fast electrons and the gear that moves them. Margin and pricing power are migrating DOWN the stack into the scarce physical layers (generation, turbines, switchgear, transformers) and away from the abundant top (GPU-cloud rental, where supply floods in and depreciation eats the spread). House call: overweight the toll-booth physical layers with multi-year backlogs and sold-out capacity; underweight/avoid the commoditizing neocloud layer where the same AI demand funds the competitors that erode your rent.
No single layer dominates the way a foundry dominates the chip stack — dominance is distributed across the scarce physical chokepoints, each an oligopoly. GE Vernova effectively dominates heavy-frame gas turbines (with Siemens Energy and Mitsubishi the only alternatives; all sold out), giving it the most durable single-layer pricing power on the board through ~2030. Constellation + Vistra dominate dispatchable nuclear/firm baseload for hyperscaler PPAs, and Eaton/Hubbell/Quanta dominate grid electricals + buildout. This distributed-oligopoly structure is MORE durable than a single chokepoint because each layer is independently capacity-gated by capital, permitting, and skilled labor — none can be bypassed and all are multi-year sold out. The fragile layer is neocloud, where dominance is illusory (capital, not technology, is the moat) and competition is hyperscaler-funded.
Full-stack value chain
The power & cooling delivery chain from the AI campus load down to grid-scale generation. The marker shows where the AI power crunch is a tailwind, headwind, or mixed for each layer’s margin pool. Tap any layer for the full read.
The supply-locked layers — gas turbines, HV transformers, switchgear, liquid cooling — get stronger as the AI power crunch bites, because multi-year backlog and physics, not capital, set the price. Commodity utilities and pure-play SMRs do not.
Shift-point register
Ranked by margin-at-stake × demand-durability × evidence-strength. The flagged rows clear the bar and are promoted to a deep-dive.
Gas-turbine slot scarcity
Heavy-frame gas-turbine slots become the binding constraint on the entire AI build-out — generation OEMs sell out through 2030 and re-price the slot itself
The fastest dispatchable path to firm 24/7 power for an AI campus is on-site or grid-adjacent gas generation, and the world has only ~3 Western heavy-frame turbine OEMs. GE Vernova's gas-turbine reservations+backlog jumped from 83GW to 100GW in a single quarter (Q1 2026) and management targets >=110GW combined and sold-out-through-2030 by YE26 (fact, GEV 8-K Apr 2026), while annual heavy-frame output across all three OEMs is measured in the low tens of GW — demand exceeds near-term deliverable supply by roughly an order of magnitude. With slots scarce, the OEM prices the reservation itself: GEV Q1'26 Power segment EBITDA margin was 16.3%, +470bps, and FY26 guide was raised to 17-19% (fact). The genuinely durable layer beneath the cyclical unit sale is the 15-25yr installed-base SERVICES annuity that each turbine locks in.
Opportunity board
22 names sorted into three buckets — durable compounders, undervalued / high-potential, and short / avoid. Each tile shows its conviction; open one for the thesis, catalyst, and falsifier.
Durable compounder
8Undervalued / high-potential
10Short / avoid
4Views & the Voices
The strongest bull and the strongest bear case, the US-listed ETF expressions of the theme, and where the tracked QAI Voices roster nets out — each stance stamped trackable vs inferred.
Bull case
the longOwn the chip-agnostic toll-booths, not the merchant-GPU pricing premium. Across the 13-layer AI compute stack, durable gross margin and pricing power concentrate in the PHYSICS-BOUND and IP-BOUND chokepoints that every accelerator must cross regardless of who designs it — EDA/IP, leading-edge foundry, HBM, advanced packaging, EUV/semicap, networking/optics, and power. The custom-silicon wave (TPU/Trainium/Maia/MTIA) erodes Nvidia's bare-GPU sliver but DEFENDS or GROWS these layers: more bespoke designs = more tape-outs, mask sets, HBM stacks, CoWoS slots, EUV layers, switch ports, and megawatts. Demand is confirmed real and supply-gated through 2027+, so margins are scarcity-priced, not cyclical.
- Memory is the cleanest toll-booth and it just printed: Micron FQ3-26 (reported 24-Jun-2026) hit a record $41.5B revenue at 84.9% gross margin (Cloud Memory BU 83% GM), with 16 take-or-pay agreements locking ~$100B minimum contracted revenue + $22B upfront cash; HBM sold out through 2026, SK Hynix sold out ~3 years (into CY2027), HBM4 volume shipping Q1-26 for Vera Rubin — a 3-maker oligopoly (97% of HBM wafers) with pricing power (fact, Q2-26).
- Physics chokepoints are sold out and un-routable: TSMC CoWoS sold out through 2026 (lead times 52-78wk) ramping to ~120-130k wpm by end-2026 yet still short; N2 booked into 2028 (78-104wk); ASML €38.8B backlog, FY26 guide raised to €36-40B, ~100% EUV share — no custom chip designs around needing CoWoS + EUV + leading-edge wafer (fact, Q1-26).
- EDA/IP is the single most-leveraged beneficiary of fragmentation: Cadence Q1-26 88.0% non-GAAP GM with record $8.0B backlog; Synopsys Q2 GM hit 83% first time; Design Automation +96% YoY — every one of the proliferating custom designs crosses the same ~3-vendor EDA layer (~70% share), so more in-house silicon = more tool seats + royalties, not fewer (fact/estimate, Q1-Q2 26).
Bear case
the short / avoidThe buildout is financed against a revenue base that doesn't exist yet, on assets that depreciate faster than the books admit, with two structural deflators already in motion. US AI capex is ~$500B/yr (2026-27) against AI revenue still in the tens of billions — roughly $8-10 of capex per $1 of revenue. The fragile middle (GPU-only neoclouds) is levered to a depreciating, collateral-shrinking asset; hyperscalers are quietly shifting from self-funded to debt-financed datacenters; and BOTH a supply deflator (China/CXMT) and a demand deflator (inference-cost collapse) can hit the toll-booths' volumes at once. A profitable-but-late short here is still a falsified thesis — but the avoid/short list is specific.
- The capex-revenue gap is structural, not a timing lag: ~$500B/yr US AI capex vs tens-of-billions end-user revenue (Cahn's $600B question, now meaningfully higher); Cembalest flags Meta capex+R&D ~70% of revenue (vs ~10% S&P median) and a Q4-25 'explosion' of hyperscaler bond/loan/lease financing replacing internal cash; Kedrosky estimates AI drove ~64-80% of recent US GDP growth — a concentration that unwinds violently if ROI disappoints (estimate, 2026).
- The GPU-depreciation time bomb (Chanos/Burry): if GPUs are 3-4yr assets not 6yr, hyperscalers understate depreciation by ~$176B over 2026-28; CoreWeave is the live tell — ~$21B+ debt (from <$8B in 2024), $1.17B 2025 net loss, GPU-collateralized borrowing against an asset whose spot rental rates are falling 50-75% and whose collateral value drops as workloads move to in-house silicon. ~20% of neoclouds modeled not to survive the cycle (estimate, 2026).
- Custom silicon erodes the fattest, most-watched margin pool: Nvidia DC GM ~74-78% and ~$220B/yr gross-profit pool is the crux — Morgan Stanley est. custom ASICs growing ~2-3x faster than merchant GPUs, heading to ~10-25% of accelerator spend; an inference-share migration that compresses the single largest profit pool in the index and the most crowded long (estimate, 2026).
Bullish-but-bifurcating — net long the chip-agnostic chokepoints (memory/HBM, foundry, packaging, EUV/semicap, networking/optics, power), short/avoid the levered GPU-rental middle (neoclouds) and the merchant-GPU pricing premium at the margin.
Dispersion: Moderate-to-high. The independent semi desks and the 13F smart money (Patel, Aasholm, Rasgon, Hosseini, Genovese, Coatue, Light Street) cluster bullish on the toll-booths and confirm the rotation into picks-and-shovels + power. The dispersion is on the demand-sustainability and financing legs: the forensic/macro bears (Chanos, Kedrosky, Cembalest, EnerTuition) attack GPU depreciation, hyperscaler leverage, and the capex-revenue gap, while the balancers (Ray Wang on CXMT, Handy on commodity reversion, Goldberg on ASIC economics, Irrational Analysis on HBM) flag the specific deflators. Aschenbrenner uniquely encodes the split as a position — long the physical buildout, put-hedged on the chips.
The consensus is NOT 'buy the GPU'; it is 'own the un-routable layers below and around the chip, plus power.' Where bulls and bears actually agree: Micron's 84.9%-GM/$100B-contracted FQ3-26 print (24-Jun-2026) hard-validated the memory toll-booth, and CoWoS/EUV/N2 are genuinely sold out into 2027-28. They split on whether scarcity pricing is structural (bull) or a distortion that CXMT (supply) and ~10x/yr inference-cost compression (demand) will mean-revert (bear), and on whether ~$500B/yr capex against tens-of-billions of revenue is buildable from the FCF buffer or a credit event waiting on the neocloud refinancing wall. Highest-conviction shared call across the roster: neoclouds (CoreWeave) are the fragile, levered, depreciating-collateral short; EDA/IP + memory + power are the defended longs.
US-listed ETF expressions
7The QAI Voices
stance · trackable / inferredPrediction matrix
Directional calls across Sep'26 / Dec'26 / Jun'27 / Jun'28, confidence decaying high → low over the horizon. Each cell is the call; click a row for the full reasoning, leading indicator, and falsifier.
Across Sep'26 -> Jun'28 the AI data-center power & cooling stack stays a demand-pull, supply-constrained regime: US data-center grid load rises from ~31GW (2025) toward ~66GW by 2027 (Goldman Sachs, Q4 2025), and ~$720B of grid capex through 2030 (Goldman, 2025) keeps order books extending faster than revenue converts. The durable margin pool sits at the capacity-constrained toll-booths — gas-turbine OEMs (GE Vernova ~80GW backlog stretching into 2029, reservations expected sold out through 2030 by YE2026), large-power transformers/switchgear (Eaton DC orders +240% YoY, $22.8B backlog, Q1 2026), and liquid-cooling (Vertiv $15B+ backlog, ~12-18mo coverage, Q1 2026) — where multi-year lead times pin pricing power regardless of which chip wins. The near-term swing factors are (a) FERC's Dec-2025 PJM co-location order converting into actual tariff rules through 2026-27 (re-rates the IPPs Constellation/Vistra/Talen up or down), and (b) whether the gas-turbine/transformer bottleneck loosens by Jun'28 as OEM capacity adds land. Base case: backlogs and pricing stay firm through Jun'27; by Jun'28 the question shifts from 'can they book orders' to 'can they convert backlog to margin without supply-chain cost creep and without a co-location/permitting air-pocket' — the regime most likely to crack the bull thesis.
Valuation scenarios
Every target is scenario-conditional with a probability; the verify produced zero outright buys. Tap a name for its full bull / base / bear ladder.
Regime calls
The four cross-cutting forces and when each bites across the Sep'26 → Jun'28 horizon. Tap any force or modulating risk to read the full call.
Constrained Buildout (Base Case)
Demand pulls hard, supply stays the binding constraint — gas turbines/transformers/cooling sold out years forward, pricing power firm at the toll-booths. Own the bottlenecks (GEV, VRT, ETN); IPPs/SMRs are higher-variance call options. Holds through Jun'27 with high confidence; the question by Jun'28 is whether supply starts loosening.
Supply Catch
OEM capacity adds, transformer capacity, and BTM/onsite alternatives loosen the scarcity by late-2027/2028 — lead times shorten, escalators fade, the pricing-power premium compresses. Equipment names de-rate from scarcity multiples toward normalized cyclicals; advantage rotates to share-gainers and lowest-cost. The natural maturation of the bull regime.
Regulatory Air
FERC/PJM final co-location rules restrict behind-the-meter, or permitting/emissions/interconnection friction stalls projects — IPPs (TLN/CEG/VST) and BTM-gas re-rate down hard, demand-deliverability gap widens, equipment orders pause. Triggered by a restrictive tariff order or a marquee PPA cancellation. Lower probability but the sharpest drawdown.
AI
Hyperscalers cut 2027-28 DC capex on AI-ROI doubt or model-efficiency gains — DC GW forecasts revised down >15%, reserved turbine/transformer slots freed, the whole stack de-rates from backlog-visibility multiples. The macro tail that breaks the demand-pull premise; watch consecutive capex guide-downs.
Power
New generation + efficiency outpace load, PJM/ERCOT forward curves soften, and the merchant/IPP scarcity premium compresses even with PPAs intact — equipment durable, but IPP multiples (20-25x EBITDA) prove too rich. Hits the generation leg specifically, leaves the equipment/cooling toll-booths relatively unscathed.
How they modulate the book
Hyperscaler AI-capex air-pocket (the single correlated factor)
The dominant near-term risk across every layer — generation, grid-gear, cooling, EPC, IPP. A synchronized hyperscaler capex deceleration fires nearly every long's falsifier at once (Vertiv book-to-bill <1.5x, Eaton orders flip negative, IPP PPA pace stalls), so the opportunity set's real independent risk is far below gross and the diversification is partly illusory. FY2026 capex tracked ~$725B (still rising); the binding window is the FY2027 capex-guide season (late-2026/early-2027). LEADING INDICATORS: the four hyperscalers' FY2027 capex guides, Eaton Electrical Americas rolling-12mo orders, and CoreWeave book-to-bill/utilization — all of which pre-date toll-booth income-statement weakness. Express via a portfolio-level AI-capex factor budget, not name-by-name sizing.
Peak-cycle margin + valuation at a capex-cycle high
The electrical/cooling names (VRT >50x fwd, GEV ~37-71x, AAON/Powell/Modine re-rated hard) are short-cycle capital-goods makers priced for sustained peak margins AND peak volume — an unkind base rate. The most relevant reference class (electrical-equipment margin persistence at a capex-cycle peak) mean-reverts. The risk is multiple compression even WITHOUT an earnings collapse: a 25-40% de-rate on the complex is plausible on any growth-rate wobble. Inversion discipline (own the lower-multiple tolls — ETN, GRID, FIX — before the chase-priced pure-plays; size for the roll, not the wedding) is the defense.
Supply response / power glut (the scarcity premium normalizes)
Every long in the generation and IPP layers depends on a binding supply shortage holding pricing — gas-turbine sold-out-to-2030, PJM capacity at the $329-333/MW-day cap, transformer 4-5yr lead times. New turbine-OEM capacity lands 2028-30, SMRs add (slowly) late-decade, on-site fuel cells/gensets (Bloom, CAT) deploy in ~90 days, and PJM capacity-market reform is actively floated. Any of these relieving the bottleneck normalizes the scarcity rent that the generation and IPP theses are built on. The frontier names (Oklo, NuScale, SMR ETFs) are most exposed — their entire value is the supply not arriving on time.
Cooling-architecture & in-house-design disintermediation
The liquid-cooling and rack-power content theses (Vertiv, nVent, Modine, AAON) assume the OEM captures the premium. Two erosion vectors: (1) hyperscalers designing cooling/power in-house and dual-sourcing the gear, and (2) the architecture shifting — direct-to-silicon microfluidics or immersion displacing the direct-to-chip CDU/rack content these names sell. A liquid-cooling-architecture standardization signal (immersion vs direct-to-chip, expected 2026-27) is the watch item; the wrong standardization strands content for some of the cooling pure-plays.
China supply-chain & geopolitics (analysis-only, no China calls)
Discussed here, never a recommendation. Mainland-China names dominate parts of the upstream power-and-cooling supply chain: grain-oriented electrical steel and transformer cores, rare-earth magnets for generators/turbines, and a large share of global power-equipment and HVAC component manufacturing. Export controls, rare-earth retaliation, or tariff escalation could lengthen the already-stretched 4-5yr transformer lead times and raise gear costs — a tail that affects the whole US-listed basket's input economics. No mainland-China-listed equity is recommended or shorted; the exposure is captured by sizing the US-listed gear and EPC tolls (ETN, GRID, FIX, PWR) that benefit from domestic-content reshoring.
Premise pressure-test
The six named 2026-Q3 catalysts the thesis rests on, probability-weighted. Click any premise for the if-true / if-false split.
AI data-center power demand keeps compounding
holdingSupply stays the binding constraint through Jun'27
holdingThe bottleneck owners hold pricing power, not just volume
holdingFERC/PJM co-location rules settle constructively
at-riskLiquid cooling becomes the default for AI racks
holdingSMRs stay optionality, not near-term cash flow
unprovenThe grid-capex / T&D cycle is durable and recoverable
holdingMacro / cost-of-capital doesn't break long-duration capex
holding