Power Cycle Thesis v1
What this is: parallel-v1 thesis to memory-cycle. Same operating shape (tranche-accumulate, no per-trade stops, phase-marker exits, R-unit sizing), different fundamentals. Per the Diamandis innermost-loop framework, power is the named bottleneck layer — $156B of US data-center projects are blocked on local power-infrastructure opposition, making power the highest-asymmetry public-market play right now [1].
The thesis
Hyperscaler AI capex ($750B annualized in 2026 [1]) is being committed faster than US electrical generation + transmission can come online. Interconnection-queue wait times in PJM and ERCOT have stretched to 36–48 months in data-center load-growth zones [2], and PJM's 2025 Long-Term Load Forecast attributes 94% of 2024–2030 peak-load growth to data centers [2]. Three structural facts compound:
- Power-supply capacity is approximately fixed on a 3–7 year horizon. New gas + nuclear + transmission cannot be built into the immediate gap; existing baseload generation is the only available answer until ~2028–2030.
- Hyperscalers are now contracting for baseload nuclear at decade-plus durations at premium-to-spot pricing (Talen-Amazon $18B / 17-year / 1.92GW [3]; Constellation-Microsoft 20-year / 835MW Crane restart [4]). This is structurally new: hyperscalers as anchor customers for the nuclear industry.
- Upstream uranium supply is similarly capacity-constrained. Spot U3O8 ~$86.55/lb May 2026, up 24% YoY [5], and Cameco's Westinghouse stake (49%) gives reactor-tech + fuel-services exposure to the SMR buildout that hyperscalers + Talen have already announced intent on [3].
The power layer monetizes asymmetrically: IPPs lock in premium PPAs against a fleet whose carrying cost is largely sunk, and upstream fuel + reactor-tech companies capture margin on the supply that the new contracts pull through.
Bull case (structural argument)
The binding constraint
Power generation + transmission. NOT chips, NOT capex willingness, NOT data-center construction. The Diamandis filing names it explicitly [1]: $156B of US data-center projects are blocked on local power-infrastructure opposition (per CNBC / Data Center Watch). This thesis exploits the named bottleneck.
Demand-side anchors (cited)
- Hyperscaler AI capex direction. Combined GOOG/AMZN/META/MSFT 2026 capex = $750B (CreditSights revised UP from $650B) [1]. Bullish state: flat or revised up. Bearish flip: any quarterly cycle with combined revision DOWN >10%.
- Hyperscaler PPA cadence at decade+ durations. Talen-Amazon $18B PPA expanded June 2025 [3]; Constellation-Microsoft 20-year Three Mile Island restart with $1B DOE loan landing Q1 2026 [4]; further SMR exploration announced between Talen + Amazon [3]. Bullish state: ≥2 new hyperscaler PPAs at >500MW / >10yr signed per calendar year. Bearish flip: zero new such PPAs across any rolling 4-quarter window.
- Interconnection queue trend. PJM/ERCOT data-center-zone wait times currently 36–48mo [2]. ERCOT large-load queue +300% YoY through end-2025, reaching 233GW of requests [2]. Bullish state: backlog stable or growing (proves demand exceeds buildout). Bearish flip: backlog clearing materially faster than 2026 baseline AND simultaneous slowdown in new project filings.
- Uranium spot price. $86.55/lb May 2026 [5], up 24% YoY. Bullish state: $80+ floor. Bearish flip: sustained move below $65 on supply-side capacity announcements.
Disqualifying conditions
- Hyperscaler combined capex revised down >10% in any quarterly cycle for 2+ consecutive quarters
- New nuclear or transmission buildouts come online materially faster than projected (closing the bottleneck inside 18 months)
- Regulatory action that forces hyperscalers off behind-the-meter / collocated nuclear deals (the FERC ruling on the Talen-Amazon arrangement is the canary [3])
- Major hyperscaler announces meaningful efficiency breakthrough that flattens forward capex (LLM compute-per-token drops 10x in a way that propagates to training too)
Locked tickers
| Ticker | Layer | Rationale | Max R |
|---|---|---|---|
| TLN (Talen Energy) | IPP / nuclear | Primary. Pure-play nuclear IPP with the highest hyperscaler-deal beta — $18B Amazon PPA expanding [3], Susquehanna 2.5GW dual-unit nuclear plant, behind-the-meter-to-front-of-meter pivot 2026. | 2 R |
| CEG (Constellation Energy) | IPP / nuclear + gas | Large-cap proxy. Most diversified US nuclear fleet (>22GW), Microsoft Crane restart 20-yr PPA [4], Trump-admin $1B DOE loan [4]. Lower beta than TLN but stronger balance sheet. | 1.5 R |
| CCJ (Cameco) | Upstream fuel + reactor tech | Uranium spot leverage [5] + 49% Westinghouse stake = SMR / fuel-services exposure. The hyperscaler nuclear theme pulls through to fuel + reactor-tech regardless of which IPP wins which deal. | 0.5 R |
Total power bucket cap = 4 R (matches memory bucket cap from v2 sizing frame).
Bear case
(500-word devil's-advocate counter. Forbidden from hedging.)
The bull case has three load-bearing assumptions, and at least two of them are softer than the parent filing admits.
First: hyperscaler capex is not destiny. The $750B 2026 number assumes the current Q1 2026 ramp continues linearly. CreditSights revised UP, true — but capex revisions go both directions, and we are now ~36 months into an unprecedented AI capex super-cycle with zero precedent for the magnitude. The dot-com analog is instructive: telecom capex peaked Q1 2001 and was cut 70% over the following 18 months, not because demand for bandwidth disappeared but because supply briefly overshot perceived near-term need. If Anthropic + OpenAI demonstrate that a new training paradigm needs materially less compute per token (the o3 → o4 efficiency pattern, the DeepSeek-R1 distillation pattern, mixture-of-experts becoming default), hyperscaler capex can flatten without any drop in end-user demand. The capex line is what funds the power bucket; flat capex breaks the thesis even if AI usage continues to grow.
Second: the power-supply "shortage" is mostly an interconnection-bureaucracy problem, not a physical generation problem. PJM trimmed its near-term load forecast in late 2025 specifically because the operator started vetting data-center load-additions more strictly — many announced data centers are speculative or double-counted across multiple utility territories [2]. Carbon Direct's queue analysis shows 111.6GW of generation + 64.3GW of storage actively queued in PJM + ERCOT [2]. That is more capacity in queue than the entire current peak load gap. If FERC + state PUCs streamline interconnection (which Trump-admin "energy abundance" rhetoric is explicitly pushing toward), the queue clears faster than the thesis assumes, and the IPP scarcity premium compresses. Talen and Constellation today trade on the assumption that scarcity persists for 5+ years; cut that to 24 months and the multiples re-rate down.
Third: behind-the-meter nuclear deals face regulatory tail risk. The Talen-Amazon arrangement was challenged at FERC precisely because it bypasses the grid-cost-recovery model that funds transmission for everyone else [3]. The 2026 transition to front-of-meter is the regulatory accommodation, but a future ruling could force these deals into grid-tariff structures that strip the premium pricing the IPPs are capturing. Constellation's Microsoft deal is structurally cleaner but politically vulnerable too — restoring a closed nuclear plant for one tech company while ratepayers fund the grid is an obvious 2026 political target.
Specific kill conditions the bull case isn't pricing: (a) a single Anthropic/OpenAI paper showing 10x training-efficiency gain, (b) FERC ruling requiring colocated nuclear loads to pay full transmission tariff, (c) PJM 2026 capacity auction clearing at materially LOWER prices than 2025 (signaling supply additions winning), (d) one hyperscaler publicly trimming capex guidance. Any one of those four prints in the next 4 quarters and the IPP basket loses 30–50% as the scarcity narrative compresses.
The bull case is real. It is not as locked-in as 12-month price action suggests.
Cross-domain translation
This is structurally a cloud-region capacity buildout cycle, run in reverse with a hard physical constraint.
The founder's data-engineering muscle has seen this exact shape three times in the AWS era:
2019–2020 us-east-1 capacity crunch. AWS hit physical capacity in northern Virginia, started rationing reserved-instance availability, and forced large customers into us-west-2 or new regions. The companies that owned land + power + fiber in northern Virginia (Digital Realty, Equinix, then-private CoreSite) saw their unit economics flip from commodity to scarce-asset. Re-rate took ~24 months.
2022 GPU capacity crunch. AWS, GCP, Azure all rationed H100/A100 availability. Customers who had committed reserved capacity 12 months earlier paid 1/3 the spot rate; new customers waited 6–12 months for any allocation. The companies that owned the upstream physical input (NVIDIA, then the HBM suppliers a year later) captured the entire surplus.
2024–2025 region-launch lag. Hyperscalers announced new regions but couldn't physically deliver power + cooling on the announced schedule. Customers shifted workloads to regions with available headroom; some abandoned regions entirely. The constraint was always physical infrastructure, not chips.
The power-layer thesis is the SAME shape, one level deeper in the supply chain. AWS/Azure/GCP are now in the seat that the AWS-customer was in during the GPU crunch — they need a physical input (megawatts) that has a multi-year provisioning lead time, and the suppliers who own existing capacity capture the surplus. TLN owns the Susquehanna fleet today; building a comparable 2.5GW nuclear plant takes 8–12 years. CEG owns 22GW of operating nuclear; replacing it takes a decade-plus. CCJ + Westinghouse own the upstream uranium + reactor-tech that any new SMR deployment will pull through, regardless of which IPP wins which contract.
What breaks the analog: in cloud-region buildouts, the constraint cleared in 18–24 months because AWS could pour concrete and order more racks at will. The power constraint doesn't clear that way. Nuclear takes 8+ years. Transmission lines take 5+ years against community opposition. Even gas peakers face 3-year permitting cycles. The supply-side response curve is longer than the demand-side surge curve by a factor of 3–5x. That asymmetry is the structural opportunity.
The founder pattern-recognition test: if you've seen us-east-1 capacity rationing flip Digital Realty's pricing power in 2019, you've seen the shape of this thesis. The difference is the supply-side response time, which is longer here. Same playbook, longer runway.
Locked tickers (detail)
| Ticker | Company | Layer | Rationale | Max position (R) | Notes |
|---|---|---|---|---|---|
| TLN | Talen Energy | Pure-play nuclear IPP | $18B Amazon PPA, 1.92GW supply through 2042 [3]; Susquehanna 2.5GW dual-unit baseload; behind-to-front-of-meter pivot 2026 [3]. Highest hyperscaler-deal beta in the basket. | 2.0 | Primary |
| CEG | Constellation Energy | Diversified IPP (nuclear majority) | Largest US nuclear fleet (~22GW); Microsoft Crane 20-yr PPA + $1B DOE loan [4]; investment-grade balance sheet. Lower beta, stronger downside protection. | 1.5 | Large-cap proxy |
| CCJ | Cameco | Uranium + reactor tech | 49% Westinghouse stake — captures SMR build cycle. Uranium spot $86.55/lb +24% YoY [5]. Pulls through regardless of which IPP wins each PPA. | 0.5 | Upstream basket leg |
Total power-bucket exposure: 4.0 R max (matches memory-bucket cap, per parent project README sizing frame).
Demand-side anchors (the disqualifying conditions)
Same 3-anchor discipline as the memory thesis — all three must remain on trajectory. If any one breaks materially (>30% miss to projection for 2+ consecutive quarters), close the entire bucket.
| Anchor | 2026 projection | Current state (May 2026) | Quarterly check |
|---|---|---|---|
| Hyperscaler combined capex (GOOG/AMZN/META/MSFT) | $750B annual [1] | CreditSights revised UP from $650B [1] | Earnings-call capex guidance every quarter |
| Hyperscaler nuclear PPAs at >500MW / >10yr | ≥2 per calendar year | TLN-Amazon expanded June 2025 [3]; CEG-Microsoft Crane restart Q1 2026 [4]; both above threshold | PPA announcement watch (IR releases, DCD, Power Magazine) |
| Interconnection-queue wait time in DC growth zones | Stable or growing | 36–48 months in PJM + ERCOT [2]; ERCOT large-load queue +300% YoY [2] | Quarterly Carbon Direct / Utility Dive reports |
Disqualifying conditions
Thesis archives if:
- Hyperscaler combined capex revised DOWN >10% in 2+ consecutive quarters
- A single training-efficiency paper drops compute-per-token by ≥10x AND is replicated by a second lab inside 6 months (compresses forward power demand)
- FERC ruling forces colocated nuclear loads onto grid-tariff structures (strips IPP premium)
- PJM 2026 capacity auction clears at materially LOWER prices than 2025 (signal that supply additions are winning the race)
- Interconnection-queue wait times in DC growth zones drop below 18 months for 2+ consecutive quarterly reports
Phase markers
Power-specific (not memory-recycled). All three are observable via public quarterly cadence — Ray monitors and surfaces flips.
| # | Marker | Bullish state | Bearish flip = exit signal |
|---|---|---|---|
| 1 | PJM + ERCOT annual peak-load growth vs forecast | Growth at or above 3.3% PJM / 9.6% ERCOT 2026 baseline [2] | Materially below baseline for 2 consecutive annual cycles — demand softening |
| 2 | Hyperscaler PPA announcement cadence (>500MW / >10yr nuclear or baseload) | ≥2 such deals per calendar year | Zero deals across any rolling 4-quarter window — anchor-customer model broken |
| 3 | Interconnection-queue backlog trend (DC growth zones) | Wait times stable or growing from 36-48mo baseline [2] | Wait times drop below 18 months across PJM + ERCOT — supply catching demand |
Quarterly review (Ray automates): scan vault + web for current state of all 3 markers, flag any flips, surface to founder.
Position sizing
Same v2 framework as memory thesis. 1R = $5,000 (paper). Sizing in R-units, scale-invariant.
| Ticker | Initial (T1) | Scale-in (T2) | Full (T3) | Max |
|---|---|---|---|---|
| TLN | 0.5 R | +0.5 R (1.0 R) | +1.0 R (2.0 R) | 2.0 R |
| CEG | 0.5 R | +0.5 R (1.0 R) | +0.5 R (1.5 R) | 1.5 R |
| CCJ | 0.25 R | +0.25 R (0.5 R) | — | 0.5 R |
Tranche triggers (same shape as memory):
- T1: thesis-confirmed entry (founder approves /decisions/ page)
- T2: -5% from T1 average price OR thesis-confirming catalyst (new hyperscaler PPA, FERC favorable ruling, capex revision up)
- T3: -10% from T1 OR earnings catalyst with thesis-positive data (capex guidance up, queue backlog growing, uranium spot strength)
No per-trade stops. Exits are phase-marker driven, not price-driven.
Position-deployment plan (initial tranche)
| Ticker | T1 size | T1 dollar (at $5k/R) | Action |
|---|---|---|---|
| TLN | 0.5 R | $2,500 | Market buy on go-decision |
| CEG | 0.5 R | $2,500 | Market buy on go-decision |
| CCJ | 0.25 R | $1,250 | Market buy on go-decision |
Total initial deployment: 1.25 R = $6,250 of paper capital. Leaves 2.75 R of headroom for tranche scale-in across the three positions.
Exit rules
| Trigger | Action |
|---|---|
| Any 1 of 3 phase markers flips bearish for 2+ consecutive quarters | Close ENTIRE power bucket (thesis broken) |
| Single phase marker flips bearish but other 2 hold | Trim each position to 50% of full size (de-risk, await confirmation) |
| Price hits +50% from T1 average | Trim 0.5 R on that name (book profit), let rest run |
| Price hits +100% from T1 average | Trim another 0.5 R (still let core ride) |
| Thesis fundamentals improve materially (e.g. third major hyperscaler signs >2GW deal) | Add 0.5 R to TLN (override 2 R cap, max 2.5 R) |
| Founder kill-switch | Close immediately |
Quarterly review template
Each calendar quarter, Ray produces a theses/<YYYY-Q#>-power-cycle-review.md with:
- Anchor metrics — current value of each of the 3 anchors vs the projection in this doc, vs last quarter's value. Flag any anchor with a Δ that crosses the disqualifying-condition threshold.
- Phase-marker state — current state of each of the 3 phase markers. Flag any flips.
- Position state — current size of each tranche per ticker, P/L in R-units, distance to next tranche-trigger or exit-trigger.
- New evidence — any new hyperscaler PPAs, FERC rulings, capex revisions, uranium price moves, interconnection-queue updates since last quarter. Source-cite each.
- Recommendation — hold / add / trim / archive. Founder approves via /decisions/ page.
First review due: 2026-08-17 (90 days post-deployment).
Sources
- [[06-reference/2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — Diamandis Innermost Loop framework; $156B blocked DC projects; $750B 2026 hyperscaler capex; power-as-named-bottleneck thesis.
- Carbon Direct: PJM and ERCOT interconnection queue analysis + Utility Dive: data center demand spike could drive 79% ERCOT price hike in 2027 + PJM 2025 Year in Review: planning for burgeoning electricity demand — 36-48mo queue wait times in DC growth zones; 233GW ERCOT large-load queue +300% YoY; 94% of PJM peak-load growth attributable to data centers 2024-2030.
- Talen Energy IR: expanded nuclear relationship with Amazon (June 2025) + Power Magazine: Talen-Amazon $18B nuclear PPA + DCD: AWS-Talen 1.92GW PPA — $18B / 17-year PPA; up to 1.92GW through 2042; spring 2026 front-of-meter transition; SMR exploration.
- CNBC: Trump administration backs Three Mile Island restart with $1B loan to Constellation + Constellation news release: Crane Clean Energy Center launch — 20-year Microsoft PPA; 835MW dedicated; $1B DOE loan Q1 2026; 2027 restart timeline.
- Cameco: uranium price + William Blair initiates Cameco at Outperform $165 fair value (24/7 Wall St) — U3O8 spot $86.55/lb May 2026 (+24% YoY); Westinghouse Q4 2025 adjusted EBITDA $780M CAD (Cameco 49% share), +61% YoY.
- [[06-reference/2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] — Packy McCormick framework: scarce asset migrates when adjacent input becomes abundant. Power-as-new-scarce-asset is the canonical instance.
Related
- [[2026-05-12-innermost-loop-ai-infrastructure]] — parent thesis (power = named bottleneck layer)
- [[2026-05-17-memory-cycle-thesis-v1-framing]] — parallel-v1 sibling thesis (same operating shape, different fundamentals)
- [[06-reference/2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — full Diamandis source synthesis
- [[06-reference/2026-04-30-not-boring-scarce-assets-abundance-driven-scarcity]] — Packy abundance-driven scarcity framework
- [[01-projects/investing/README]] — project home, v2 quality process, R-unit sizing frame
- [[feedback_calibrate_overconfidence]] — applies to investing too
Notes on what's intentionally NOT in this thesis
- No BE / OKLO / FLNC / GEV / NRG. Per Diamandis filing [1], BE is at +1,647% 12mo and extended. OKLO + FLNC have strong narratives but no operating revenue (OKLO has never sold a reactor; FLNC margins are still negative). GEV is large-cap and the trade is already 12 months in. NRG is a gas IPP without the nuclear-PPA beta. Could reconsider any of these at v2 review.
- No options sleeve. v1 = stock-only, matching the memory thesis. Options consideration deferred to v2 after 1 quarter of paper track record.
- No private-market frontier-lab exposure. Layer 1 of the innermost loop is out-of-scope for the public-market paper book.
- No short positions. Bear-case kill conditions trigger ARCHIVE (close longs), not short entry.