Morning Analysis — 2026-06-19

Dek

The shooting war fades, and the only fight left is whether all that AI-and-nuclear capex actually pays for itself.

The Big Shift

Washington is quietly building a *second* Western rare-earth champion. The Pentagon handed Energy Fuels a $725M loan to expand domestic rare-earth processing — refining, not just digging — days after inking $1.2B in separate rare-earth mineral loans (¹, ²). Why it matters: MP Materials' whole story is being the *only* integrated Western producer — mine plus magnet-grade processing. A federally funded rival pokes a hole in that moat. What it points to: the U.S. government is now an active venture builder in critical minerals, and "sole supplier" theses everywhere get weaker the more checks Defense writes.

Analysis

Geopolitics moved from the front page to the footnotes. The U.S.-Iran deal is reopening the Strait of Hormuz; an LNG tanker has already transited and Germany is moving warships to Djibouti for a possible escort mission (³, ⁴). The one fresh wrinkle: nuclear talks hit an early "snag" and stocks slipped (⁵). Implication: the war-risk premium that lifted everything from oil to uranium is draining out. From here, energy and materials names have to be carried by demand and execution — not by fear. That raises the bar for every "scarcity" thesis in the book.

On the nuclear-for-AI pillar, the signal is real but already paid for. Centrus (LEU) locked a multi-year deal to feed HALEU — the higher-enriched fuel next-gen reactors need — to Oklo's Aurora SMRs (⁶). That confirms the SMR-demand and sole-U.S.-enricher story. But it's a day-old, recurring item bundled into routine annual-meeting filings, so it changes nothing about the LEU view. Implication: the thesis is intact, not improving — you're holding, not adding.

Meanwhile a reality check on the nuclear hype. Miners — the industrial buyers everyone assumed would snap up small modular reactors — rank SMRs dead last in their clean-energy plans (⁷). Implication: SMR demand is concentrated in AI/data-center and government buyers, not broad industry. The story is narrower than the headlines suggest, which makes it more fragile if even one anchor customer slips.

The power-for-compute crunch keeps showing up in odd places. A new German data center broke ground *outside* Frankfurt purely because there's no grid capacity in the city (⁸), and the DOJ stepped in to defend xAI's gas-turbine-powered data center as "critical to national security" — disclosing that Grok helped target missiles in the Iran war (⁹). Implication: the grid is the binding constraint, AI compute is now treated as defense infrastructure, and that gives the power-buildout thesis a political tailwind that survives even as war risk fades.

Watch the plumbing of the boom. A new note flags circular financing among Nvidia, CoreWeave, and Nebius — chipmaker money cycling back through its own customers (¹⁰. Implication: this is the tell for the whole "capex earns its keep" question — if demand is partly self-funded, reported growth overstates real end-demand. It's the soft spot under every AI-infrastructure long.

What Would Prove Us Wrong

• MP thesis: If Energy Fuels (or another Pentagon-backed name) announces magnet-grade or NdFeB *output* with a named offtake customer — not just processing capacity — MP's "only integrated Western producer" moat is broken, not just dented. Watch for a follow-on 8-K or DoD award naming finished-magnet production.
• Power/compute thesis: A concrete signal that AI data-center demand is softening — a hyperscaler capex cut, a CoreWeave/Nebius contract cancellation, or evidence the Nvidia financing loop is unwinding — would undercut the grid-scarcity and SMR-demand stories at once.
• War-risk-gone read: If the Iran nuclear talks collapse (not just "snag") and a Hormuz transit is actually blocked or a tanker hit, oil and the geopolitical premium snap back — re-rating everything we just marked down on the "geopolitics is now a footnote" call.

Thesis Impact

LEU | Conviction: HOLD | Surprise: LOW | Centrus Energy (=LEU) signed a multi-year HALEU offtake to fuel Oklo's Aurora fast-fission SMRs — directly CONFIRMS the SMR-demand pillar (P4) and the sole-US-enricher thesis. But it's recurring (shown 1d), already in the 0.71 prior, and yesterday's 8-K bundles it with routine annual-meeting items (board change, shareholder vote). No change to prior. | ⁶

MP | Conviction: HOLD | Surprise: LOW | Energy Fuels landed a $725M Pentagon loan to expand domestic rare-earth *processing* — a federally-backed second Western integrated REE player, which CONTRADICTS MP's "only integrated Western producer" moat and pokes the thesis-breaking trigger. Disconfirming and worth watching as a building multi-day thread (Pentagon now funding several REE names), but recurring and already reflected in the beaten-down 0.30 prior. | ¹

Everything else overnight is recurring Iran/Hormuz/data-center buildout already priced into the book. The one genuinely NEW T1 item — Iran nuclear talks hitting a "snag" (stocks slid) — nicks the "war risk is gone" cross-domain read but is too minor to move any tracked name. The dominant question today is execution/capex, not geopolitics: no NEW signal changes that.

Inflection Radar

[emergent] AI Inference Cost Bottlenecks | Academic work detailing methods like Causal Attribution Pruning and cache-aware RAG protocols shows the industry is rapidly shifting focus from model size to optimizing inference cost and efficiency for deployment. | Touches: NEW | ¹¹

[emergent] LLM Scaling Economics | Research analyzing optimal LLM routing under limited user feedback confirms that the primary commercial constraint is not model quality, but cost-optimal resource allocation during inference. | Touches: NEW | ¹²

[emergent] Data Center Grid Mandates | FERC demanding grid operators justify or reform data center connections signals that the energy supply and interconnection process is becoming a major regulatory bottleneck for AI build-out. | Touches: NEW | ¹³

[emergent] Generative Model Safety & Control | New academic principles for "Grounded Inference" highlight the inherent risks and need for deterministic encapsulation when deploying generative models into critical systems. | Touches: NEW | ¹⁴

[emergent] Drone Utility Adoption | AirData surpassing 120 utility customers globally, coupled with the looming FAA Part 108 rule, confirms that critical infrastructure sectors are moving past testing into mandatory operational deployment of drone services. | Touches: NEW | ¹⁵

QA & Caveats

No issues found.

Sources

1. Energy Fuels lands $725M Pentagon loan for rare earths boost mining.com
2. Pentagon inks pair of rare earth mineral loans for $1.2 billion breakingdefense.com
3. The Strait of Hormuz Is Reopening heatmap.news
4. Germany moves two ships to Djibouti, eyeing multinational Hormuz mission this summer defensenews.com
5. Stocks Slide as Iran Nuclear Talks Hit Early Snag: Markets Wrap bloomberg.com
6. Centrus Energy, Oklo sign multi-year nuclear fuel deal mining.com
7. Mining’s nuclear reality check: SMRs are still not on miners’ shopping lists mining.com
8. Firstcolo breaks ground on data center in Rosbach vor der Höhe, Germany datacenterdynamics.com
9. DOJ intervenes on behalf of xAI in data center gas turbine lawsuit utilitydive.com
10. Nvidia, CoreWeave & Nebius: Circular Financing in the GPU Boom - Beth Kindig – Medium news.google.com
11. Pruning via Causal Attribution Preserves Reasoning Performance in Large Language Models arxiv.org
12. Cost-Optimal LLM Routing with Limited User Feedback under User Satisfaction Guarantees arxiv.org
13. FERC orders US grid operators to justify or reform how data centers connect to the grid datacenterdynamics.com
14. Grounded Inference: Principles for Deterministically Encapsulated Generative Models arxiv.org
15. AirData Marks Energy and Utility Growth Milestone as BVLOS Rule Nears dronelife.com