Morning Analysis — 2026-09-16

Dek

Power was supposed to be the hard part. Overnight it turned out the harder part is a zoning board in an election year.

The Big Shift

A study commissioned by PJM — the grid operator for 65 million people from Chicago to New Jersey — says that if data-center load keeps growing at current pace, the region could be missing power more than 13 days a year by 2030, "over 100 times worse" than its own planning standard (¹). That is the cleanest official admission yet that electricity, not chips, is what rations AI buildout. It points two ways at once: scarcity pricing and premium value for firm power that already exists inside PJM, and a much louder political fight over who gets blamed for the shortfall.

Analysis

Power. A 2030 reliability gap has only two near-term fixes — build wires or plug into generation that's already there. Both are bullish for the same names for different reasons. Transmission and substation work is Quanta's order book, which is why the PJM number confirms the "transmission is the binding step" pillar; the offset is yesterday's PWR 8-K flagging an executive/board change with no detail, an unexplained data point until the proxy fills it in. For Talen and Constellation, scarcity raises the value of uncontracted nuclear sitting inside the constrained zone and strengthens the case for co-located "behind-the-meter" deals — where a data center plugs straight into a plant rather than the public grid — that FERC is still weighing. NERC's own CEO framing the problem as mindset rather than physics (²) tells you the regulator expects to be pushed.

Politics is now a hard input, not a mood. Republican candidates in tight midterm races are breaking with Trump over data centers (³), and Trump's response is to call the backlash a hoax (⁴) — which escalates rather than settles it. The fight has reached Philadelphia neighborhoods already scarred by heavy industry (⁵). Implication: this splits the thesis book. Sites that are *already energized and interconnected* (WULF's Lake Mariner) gain scarcity value because the alternative got worse. Sites that are still a *pipeline* awaiting permits (KEEL's 2.2 GW, much of it in Pennsylvania) get slower and more expensive. And it exposes a bottleneck the Own-the-Bottleneck thesis doesn't price: social license can't be bought, cornered, or contracted. That's a missing pillar, not a downgrade.

Gas and geopolitics are now one trade. With Hormuz disrupted, global LNG buyers are turning to US cargoes (⁶). Exporting more gas tightens it at home and lifts the domestic price. That hits Bloom Energy's named thesis-breaker directly — on-site fuel cells only pencil out at cheap gas — while simultaneously lifting PJM power prices, which helps the nuclear names. Same event, opposite signs. The security side is thickening in parallel: Seoul is being pressed on a Hormuz mission (⁷) and Japan won't rule out nuclear submarines (⁸). Read that as a slow bid under naval nuclear fuel and long-cycle defense supply chains.

Materials: Washington keeps buying the mine. The US government closed a $35.6M deal for roughly 10% of Trilogy Metals' Alaska copper (⁹), and Pentagon money doubled Bitterroot's stock on a nickel-copper play (¹⁰). Taking equity, not just writing offtake contracts, is a different posture: the state is absorbing development risk private capital wouldn't. Implication for the compute thesis — copper is the shared input for transmission, transformers, and data-hall busbar, so anything that de-risks new supply relaxes the 2030 constraint at the margin, but on a decade lag, not in time for PJM.

Compute demand hasn't blinked. Microsoft filed for a two-building campus in Prince William County (¹¹), Serverfarm closed another $895M to take its development fund to $3.89B (¹²), and the read across the tape is shortage through 2028 (¹³. The counter-narrative — that slowing AI development would repair hyperscaler balance sheets — is a recurring opinion piece with no capex guide behind it. Treat it as noise until a hyperscaler actually cuts a number.

What Would Prove Us Wrong

• PJM or FERC softens the reliability finding. If PJM's own capacity auction or reserve-margin filing lands materially better than this commissioned study — or FERC rules against co-located behind-the-meter deals — the scarcity premium under TLN and CEG deflates, and the "power is the binding constraint" pillar under Own-the-Bottleneck weakens. Watch the next auction result and any FERC co-location order.
• Gas doesn't actually tighten. If Henry Hub stays flat or falls despite LNG pull — mild winter, new supply, Hormuz de-escalation — the Bloom Energy downgrade was wrong and the fuel-cost thesis-breaker never fires. Measurable: front-month Henry Hub over the next 4–6 weeks.
• The backlash loses at the ballot box. If pro-data-center candidates win the contested midterm races Heatmap is tracking (¹⁴), permitting risk reprices down fast. That would hurt the WULF "scarce energized site" call and rescue KEEL's pipeline thesis — the exact inverse of today's read.

Thesis Impact

• TLN | Conviction: UP | Surprise: MED | PJM-commissioned study says loss-of-load by 2030 could run "over 100 times worse" than the planning criterion — 13+ outage days/yr — if data-center load keeps growing. Susquehanna sits inside PJM; scarcity pricing and the value of co-located firm nuclear go up, and it strengthens the regulatory case for behind-the-meter structures FERC is still weighing. CONFIRMS pillars 1 and 2. | ¹
• PWR | Conviction: UP | Surprise: MED | Same PJM study — the only near-term fix for a 2030 reliability gap is wires and substations, which is exactly Quanta's order book. CONFIRMS pillar 1 (transmission is the binding step). Offsetting, unexplained: yesterday's 8-K flags an executive/board change with no detail — watch for the 10-K/proxy follow-up. | ¹
• BE | Conviction: DOWN | Surprise: MED | Hormuz disruption is pulling global LNG buyers toward US cargoes, which tightens domestic gas and pressures the fuel cost that on-site SOFC economics depend on. Directly touches the named thesis-breaker ("sustained natural-gas price spike kills the unit economics"). CONTRADICTS pillar 3. T1 wire. | ⁶
• WULF | Conviction: UP | Surprise: MED | The data-center backlash went electoral overnight — GOP candidates in tight midterm races are breaking with Trump on data centers. Greenfield siting gets slower and politically expensive, which raises the scarcity value of an already-energized, already-interconnected site like Lake Mariner. CONFIRMS pillar 1 by making the alternative worse. | ³
• KEEL | Conviction: DOWN | Surprise: MED | Same backlash, opposite side: Keel's whole story is converting a 2.2 GW *pipeline* into leased capacity, and PA is one of the states where DC siting is now a campaign issue. CONTRADICTS pillars 1 and 4 (siting/permitting de-risking). Trump publicly calling the backlash "a hoax" signals the fight escalates rather than settles. | ⁴
• CEG | Conviction: HOLD | Surprise: MED | PJM reliability study plus the Hormuz gas-price pull both push PJM power and capacity prices higher — good for uncontracted nuclear. CONFIRMS pillars 1–2, but this is the expected news, and at P=0.95 there's no room to move. The residual to watch is still Crane timing, not price. | ¹
• Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | PJM's numbers are the cleanest official statement yet that power is the binding constraint (CONFIRMS pillar 1). But the backlash cluster adds a bottleneck the thesis doesn't price: social license and local permitting, which no one "owns" and which can't be bought. Not a downgrade — a missing pillar. | ¹
• GEV | Conviction: UP | Surprise: LOW | A 2030 PJM reliability gap is a gas-turbine order signal, and the political backlash is aimed at data centers rather than generation. CONFIRMS pillar 1. Small move only — nothing new on backlog conversion or margins, which is where this thesis actually lives. | ¹
• ETN | Conviction: HOLD | Surprise: LOW | Microsoft filing a two-building campus in Prince William County and Serverfarm closing another $895M (fund now $3.89B) show capex still flowing despite the noise. CONFIRMS pillar 1. Already in the prior; the offsetting Semafor "slowing AI development" piece is recurring and stays noise until an actual capex guide moves. | ¹¹
• LEU | Conviction: HOLD | Surprise: LOW | SMR costs "still haven't improved" is the right shape of disconfirming item for pillar 4 (the HALEU demand curve) — but it's a T3 Reddit post, recurring, with no primary cost data behind it. LEAD ONLY, no move. Saudi's 110 Mt uranium find is ore, not enrichment, and doesn't touch the chokepoint. | ¹⁵
• ISRG | Conviction: HOLD | Surprise: LOW | Unitree down 40%+ since listing, ~$30B wiped, Beijing now cautious on humanoid IPOs. CONFIRMS pillar 1 — the benchmark gap between proven razor-and-blades robotics and humanoid hype widens. Recurring, already reflected. | ¹⁶

Inflection Radar

[dismissive] Rivian ALSO e-bikes failure | The spin-off is diverting units for public tours rather than fulfilling customer orders, signaling a potential immediate bottleneck or strategic pivot in consumer fulfillment models. | Touches: RIVN | ¹⁷

[emergent] AI Energy Management Alliance | Nvidia, Google, and Emerald AI forming a coalition to manage compute demand, signaling that the next major constraint for AI scaling is grid infrastructure/demand flexibility. | Touches: NEW | ¹⁸

[emergent] Govt Energy Permitting Reform | FERC soliciting public comment on multiple information collection activities (FERC-555, -561, -912, etc.), indicating continued federal regulatory focus and friction points in energy infrastructure development. | Touches: NEW | ¹⁹

[emergent] Geospatial AI Mapping | Startup developing tools to create georeferenced 2D maps from minimal input images, lowering the barrier to entry for high-resolution geospatial data collection. | Touches: NEW | ²⁰

[emergent] AI Model Efficiency & Cost | Multiple arXiv papers (e.g., 70, 67) focusing on optimal model activation policies and distilling foundation models for low-latency, cost-effective agentic reasoning in hot paths. | Touches: NEW | ²¹

[emergent] Advanced Robotics Control | Research advancing whole-body control (WBC) and compositional manipulation (ManiSkillFormer, 71), moving beyond simple task execution toward generalizable, physically constrained human-level dexterity. | Touches: NEW | ²²

[emergent] Defense Procurement Cycle | Anduril officials noting the dependence on FY27 procurement dollars for CCA production, highlighting the cyclical nature and budgetary constraints in advanced defense manufacturing. | Touches: NEW | ²³

[emergent] AI Safety Architecture | AIUC raising capital led by Anthropic hires, indicating venture focus on specific architectural methods to mitigate or "rein in" risks associated with autonomous/rogue AI agents. | Touches: NEW | ²⁴

***

Summary & Day Ahead:

The signals confirm a major cross-domain convergence: AI compute is now inextricably linked to physical infrastructure (Energy Grid, Robotics). The primary risk surface is the *cost and efficiency* of running advanced AI models in production environments. Watch for follow-up on the FERC docket filings as they represent systemic regulatory friction points that could slow down energy deployment regardless of technological breakthroughs.

QA & Caveats

No issues found.

Sources

1. PJM faces 2030 reliability crisis if data center load growth continues: report utilitydive.com
2. Change 4 mindsets to build a more resilient US grid: NERC CEO Jim Robb utilitydive.com
3. The data-center backlash is so fierce that Republican candidates are willing to cross Trump marketwatch.com
4. Trump’s New Data Center Defense: ‘It’s a Hoax’ heatmap.news
5. The AI data center boom is colliding with cities scarred by big industry techcrunch.com
6. LNG Buyers Turn to US as Hormuz Crisis Forces Supply Rethink bloomberg.com
7. US, South Korean officials to broach thorny Hormuz mission at talks this week defensenews.com
8. navalnews.com navalnews.com
9. U.S. takes 10% Trilogy stake in Alaska push northernminer.com
10. Bitterroot doubles on $5.2M Pentagon funding mining.com
11. Microsoft files to build data center campus in Prince William County, Virginia datacenterdynamics.com
12. Serverfarm expands North American data center development fund to $3.89 billion datacenterdynamics.com
13. AI Infrastructure Shortage May Last Until 2028 as Hyperscaler Capex Keeps Rising - KuCoin news.google.com
14. The Top 5 Midterm Races of the Data Center Backlash heatmap.news
15. SMR Costs Still Haven't Improved. Big Tech Is Paying Anyway. reddit.com
16. Unitree’s US$30 billion stock wipe-off spurs regulatory caution on humanoid robot IPOs scmp.com
17. Rivian’s ALSO e-bikes are finally hitting the road – but not for frustrated customers electrek.co
18. AI Heavyweights Team Up to Promote Demand Flexibility heatmap.news
19. Commission Information Collection Activities (Ferc-555); Comment Request; Extension federalregister.gov
20. From Wildfire Detection to Fast Mapping: AIVE AI Systems Expands Access to Geospatial Data dronelife.com
21. Optimal Model Activation Policies for Inference Networks of Large Language Models arxiv.org
22. ManiSkillFormer: Demonstration-Free Compositional Manipulation via Task-Conditioned Geometric Contracts arxiv.org
23. Anduril officials say company can’t start on CCA production contract without FY27 budget breakingdefense.com
24. Early Anthropic hire, former METR COO have found a way to rein in rogue AI agents techcrunch.com