Morning Analysis — 2026-06-24

Dek

Tankers are flowing through Hormuz again and oil is cheap — but the constraint on AI isn't a barrel of crude, it's the wire that can't carry the electron.

The Big Shift

Overnight the energy picture split in two directions that matter for every AI thesis. Cheap, abundant fuel is back: with the US–Iran deal reopening the Strait of Hormuz, Brent fell below $75 and the key oil spread flipped to contango — a bearish structure (near-term barrels cheaper than later ones) that signals ample supply, the first such flip since February (¹). Yet the binding limit on AI compute isn't fuel — it's grid delivery, and overnight that bottleneck drew a bipartisan bill to shield households from data-center power costs (²). The signal: when the molecule gets cheap and the electron stays scarce, value keeps migrating to whoever owns the wire, the substation, and the firm-power contract.

Analysis

Power. The connective thread of the day is the one flagged on the radar: the bottleneck was never the chip or the capital — it's the wire. The US already runs only about half its grid capacity because of how the system is planned and dispatched (³), and last week's FERC order telling grid operators to speed up data-center interconnection (or justify not doing so) is the regulatory admission of that (⁴). The new bipartisan ratepayer bill targeting loads of 100MW+ adds a *political* cost to slow interconnection. Implication: this confirms the PWR/own-the-bottleneck thesis — transmission and substation work gets pulled forward, and the firms that build through the queue benefit. It cuts slightly against BE (Bloom): if the grid speeds up, the "buy on-site fuel cells because the queue can't deliver" argument weakens at the margin — though physical transformer and power scarcity, the deeper reason for behind-the-meter generation, persists for years.

Where developers are voting with capital. Watch where new campuses are going: a 275MW campus on a California oil field paired with an oil-and-gas firm (⁵), Chevron fueling a Microsoft Texas site with gas, and a Starbucks HQ floor reworked into a 20MW site to dodge Seattle's data-center moratorium (⁶). Implication: when builders co-locate with fuel and existing interconnections rather than wait in line, it reinforces that the scarce input is *delivered power*, not real estate or chips — the core own-the-bottleneck read.

Firm power and nuclear. Three threads converge: Walmart's nuclear deal with Constellation adds a large *corporate* (non-cloud) buyer of firm power; AtkinsRealis is seeking US approval to sell reactor tech into the AI-driven demand surge (⁷); and DOE's $17.5B loan push backs Westinghouse AP1000 deployment (⁸). Implication: this broadens the firm-power demand base beyond hyperscalers (confirms the CEG PPA-re-rating thesis) but is mostly long-dated reactor demand — it burns regular LEU, not the high-assay HALEU at the center of the enrichment-monopoly thesis, so it supports the backdrop without adding to that specific bet.

Compute. OpenAI and Broadcom's "Jalapeño" custom inference chip is a real efficiency step, but it's incremental — better performance-per-watt, not a demand-flattening break (⁹). Implication: cheaper inference historically *raises* total compute use (Jevons-style), so it doesn't trip the own-the-bottleneck thesis; if anything it deepens the power demand it's meant to relieve.

Materials. The friction stays structural: China blacklisted two US rare-earth firms (¹⁰), sustaining the scarcity premium behind MP, while lithium pulls the other way — prices slid ~10% on speculation CATL restarts a big mine, and CATL is leaning into sodium-ion to dodge lithium volatility. Implication: the materials picture is bifurcating — rare earths tight (geopolitics-driven), lithium loosening (supply-driven) — so treat them as separate trades, not one "critical minerals" basket.

What Would Prove Us Wrong

• PWR / own-the-bottleneck: If interconnection queue times actually shorten and grid-operator backlogs fall over the next two quarters — or transmission/substation capex guidance stays flat despite the FERC order — the "the wire is the constraint" thesis loses its edge. Watch RTO queue stats and Quanta-type backlog/book-to-bill prints.
• BE (Bloom): If faster grid access slows behind-the-meter fuel-cell demand, look for soft data-center bookings or order deferrals in Bloom's next update. Conversely, sustained transformer/power lead-times would confirm the scarcity argument holds despite FERC.
• Materials: A confirmed CATL mine restart driving lithium down further (past the ~10% already lost), or China *lifting* rare-earth controls, would undercut the scarcity-premium leg — measurable directly in Chinese lithium carbonate spot prices and any change to the MP-firm export designations.

Thesis Impact

• PWR | Conviction: UP | Surprise: MED | FERC's 18 June directive ordering grid operators to speed up (or justify) data-center interconnection is concrete movement on the exact queue bottleneck Quanta builds through — more transmission/substation work pulled forward. CONFIRMS pillar 1; an actual order is firmer than the recurring "reform is coming" chatter. | ⁴
• BE | Conviction: DOWN (slight) | Surprise: MED | Same FERC order cuts the other way: faster grid interconnection chips at Bloom's core rationale — that DCs buy on-site fuel cells *because the queue can't deliver*. CONTRADICTS pillar 1 at the margin. Modest only — physical power/transformer scarcity (the deeper reason for behind-the-meter generation) persists, and FERC orders take years to bite. | ⁴
• CEG | Conviction: HOLD | Surprise: MED | Walmart–Constellation nuclear deal adds a large *corporate* (non-hyperscaler) firm-power buyer — CONFIRMS the PPA-re-rating pillar and broadens the demand base beyond clouds. Held, not raised: prior is already high (0.76) and corporate offtake is smaller-scale than the hyperscaler PPAs already in the thesis. | ¹¹
• Own-the-Bottleneck | Conviction: HOLD | Surprise: LOW | Two would-be disconfirmers checked and dismissed: FERC interconnection reform is *process* easing, not physical supply (no transformer/HALEU/copper added), and the OpenAI–Broadcom "Jalapeño" inference chip is incremental efficiency, not a durable demand-flattening step-change. Neither trips a thesis-breaking trigger. | ⁹
• LEU | Conviction: HOLD | Surprise: LOW | The $17.5B DOE nuclear-supply-chain loan / Westinghouse AP1000 push is recurring and aimed at *reactor* deployment (which burns LEU), not HALEU enrichment — supports the long-run demand backdrop but adds nothing to the enrichment-monopoly thesis. | ⁸
• MP | Conviction: HOLD | Surprise: LOW | China's blacklisting of US rare-earth firms is recurring (already in the prior) — it sustains the scarcity premium but carries no new information. No move on a 0.26 prior. | ¹⁰

Inflection Radar

[AI Compute] Causal Inference & Efficiency Frontier | Academic focus shifts to robust causal discovery, moving beyond correlation in LLMs; includes novel methods for streaming KV-Cache management and prompt steering, signaling a move toward highly constrained, reliable agentic systems. | Touches: NEW | ¹² (Using the most representative paper for the cluster)

[Robotics & Autonomous] Sim-to-Real Validation Rigor | New theoretical frameworks are integrating 'betting' and anytime-valid inference into sim-to-real pipelines, addressing the critical gap in guaranteeing physical safety and performance outside of controlled simulation environments. | Touches: NEW | ¹³

[Energy & Power] Nuclear Buildout Policy Confirmation | Multiple signals (DOE loan programs, Brookfield commitment) confirm sustained federal capital backing for nuclear infrastructure development, suggesting a policy-driven shift in baseload power planning. | Touches: CCJ, CCO | ¹⁴

[Gov & Policy] AI Utility in Government Procurement | DIA's interest in AI platforms to streamline procurement and back-office functions signals the maturation of AI from a research tool into core, efficiency-driving government operational infrastructure. | Touches: NEW | ¹⁵

[Gov & Policy] Energy Demand Regulation (FERC) | FERC actively developing plans for data center energy consumption, signaling that compute density and AI scaling are now primary regulatory concerns at the utility level. | Touches: NEW | ¹⁶

[Gov & Policy] Critical Resource Risk (Uranium) | Expert warnings highlight a potential nuclear blind spot in proposed geopolitical deals concerning the stability and control of national uranium stockpiles. | Touches: NEW | ¹⁷

[Startup/SaaS] AI Agents in High-Friction Processes | Early commercial application of AI agents to automate and structure traditionally manual, high-touch professional processes (e.g., candidate interviewing), signaling a shift from content generation to workflow orchestration. | Touches: NEW | ¹⁸

***

Summary:

The periphery signals confirm a convergence of three major themes: State-backed capital flowing into resilient infrastructure (Nuclear/Energy), the operationalization of advanced AI theory in physical systems (Robotics/Simulation), and regulatory bodies treating compute power as a critical resource (FERC/DIA). The academic focus on causality suggests that future high-value applications will prioritize verifiable reasoning over sheer scale.

Key Points:

1. Infrastructure De-risking: Government policy is actively de-risking and funding the nuclear sector, while simultaneously regulating the energy demands of AI data centers. This creates a clear capital flow signal between Energy/Utilities and Compute.

2. AI Maturity Curve: The research focus has moved past simply *generating* content (LLMs) to *structuring* knowledge (Causality, Sim-to-Real validation), indicating the next wave of enterprise value will come from reliable decision support in physical or complex systems.

3. GovTech Adoption: Government agencies are adopting AI not for public-facing services, but for internal, high-friction processes like procurement and HR, signaling immediate, measurable cost-saving use cases.

Recommendation:

Prioritize tracking the intersection of Energy Policy (Nuclear/Grid) and AI Compute. The regulatory environment is creating a massive capital opportunity where reliable, low-carbon power sources are needed to support AI buildout. Monitor any cross-domain patents or proposals linking advanced compute cooling/power management to grid stability.

QA & Caveats

No issues found.

Sources

1. Key Oil Spread Flips to Contango as Supply From Hormuz Climbs bloomberg.com
2. US lawmakers introduce bipartisan bill to shield ratepayers from data center energy costs datacenterdynamics.com
3. Why the U.S. Uses Only Half of Its Grid Capacity spectrum.ieee.org
4. U.S. Pushes Grid Operators to Connect Data Centers Faster spectrum.ieee.org
5. Canada's Beacon DC targets 275MW data center campus on California oil field datacenterdynamics.com
6. 20MW data center proposed in Starbucks' Seattle HQ building in Washington datacenterdynamics.com
7. Canada’s AtkinsRealis Seeks US Approval for Nuclear Tech to Power AI Boom bloomberg.com
8. US announces $17.5 billion in loans for nuclear power supply chain - WKZO news.google.com
9. OpenAI and Broadcom unveil LLM-optimized inference chip openai.com
10. China blacklists US firms in rare earths war northernminer.com
11. Walmart Goes Nuclear heatmap.news
12. An Introduction to Causal Reinforcement Learning arxiv.org
13. Sim-to-Real Betting on the E-Process: Bringing "simulators" to anytime-valid confidence sequences arxiv.org
14. US government launches $17.5B loan program to accelerate nuclear buildout northernminer.com
15. DIA considering new AI-powered platform to streamline procurement system defensescoop.com
16. FERC Has a New Plan for Data Centers heatmap.news
17. Trump’s new Iran deal faces nuclear blind spot over uranium stockpile, experts warn - AOL.com news.google.com
18. Fika Jobs raises $4M to build a video-first hiring platform where AI agents interview candidates techcrunch.com