Morning Analysis — 2026-06-10

Dek

The Persian Gulf is reopening and the fear premium is draining out — which means AI's power hunger now has to pay for these trades all by itself.

The Big Shift

Overnight the Middle East kept de-escalating: Iraq is pushing more oil through the Strait of Hormuz as tankers return to the route, and Trump floated an Iran deal "in two or three days" with the strait reopening "immediately" (¹). Why it matters: a chunk of copper's and energy's recent strength was a *risk-appetite* bid — money pricing in conflict, not demand. As that bid bleeds out, the structural AI-demand story has to carry these prices on its own. The signal: today separates trades backed by real demand from trades that were just riding the geopolitical wave.

Analysis

The demand side, meanwhile, keeps printing big. OpenAI is in talks to lease a 10GW data center from SB Energy on federal land in Ohio (²), and China's roughly $295B data-center rollout will need a matching wave of grid and generation spending to work at all (³). Both reconfirm the core Own-the-Bottleneck thesis: power, not chips, is the thing AI can't get enough of. A 10GW ask is the size of ten large nuclear reactors for one tenant — that is the binding constraint, stated out loud.

But the cadence isn't perfectly smooth. Crusoe paused its 1.8GW Cheyenne, Wyoming site "at the request of our customer" while still touting a 5GW pipeline (⁴). For the VRT equipment thesis ("orders track hyperscaler capex"), this is the first small crack — one site reshuffled, not cancelled, and more than offset by the Ohio talks. What it implies: watch whether "pauses" stay one-offs or start clustering. The tell that the picks-and-shovels floor is real is the steady equipment-block 8-K cadence (VRT again on 06-03) — a sector-wide filing pattern, not one firm's news.

On materials, the de-escalation cuts hardest where the geopolitical beta is highest. FCX, with its Indonesia exposure, loses more of its fear bid than SCCO, widening the SCCO-over-FCX split the radar has tracked for four days (posteriors 0.78 vs 0.52). The structural copper-deficit case is unchanged — Goldman's 350kt mine-supply cut still stands — but near-term price now has to lean on demand alone, not on a Hormuz scare. Implication: if copper holds here *without* the fear premium, that's a genuine demand signal; if it slips, the recent rise was mostly risk appetite.

The one live counter-thesis keeps inching forward: a timing trick cut LLM *training* energy by up to 14% (⁵), feeding the "AI gets more efficient, so the power need is overstated" argument now in its third day. But it touches training — a thin slice — not inference, which is where the durable load lives. It contradicts the firm-baseload CEG/VST thesis only at the margin; not enough to move conviction. The generation-owner thread is holding, not advancing, and is waiting on Oracle's call tonight for fresh demand-side proof.

Two housekeeping items for today. KEEL filed an 8-K that entered a material agreement, took on material debt, *and* sold unregistered equity — verify the size of the raise and the counterparty before reading it either way; the equity sale pressures the "no heavy dilution" pillar, but the unnamed material agreement could be the awaited anchor AI/HPC lease. And Microsoft's Nevada tariff push — making big loads pay for their own infrastructure (⁶) — plus fresh polling showing most Americans want a data-center moratorium (⁷) are the social-license risk to watch under the whole buildout.

What Would Prove Us Wrong

• Crusoe-style pauses cluster. If a second or third hyperscaler site gets paused or cancelled this week — not reshuffled — that breaks "orders track capex" and hits VRT and the equipment-block thesis directly. One pause is noise; three is a cadence.
• Copper keeps sliding as Hormuz reopens. If SCCO falls materially over the next few sessions once the fear premium is fully gone, it confirms the recent rise was risk appetite, not demand — denting the structural-deficit case for SCCO/FCX. Watch the price action against a confirmed strait reopening.
• Oracle guides capex/demand soft tonight. Weak hyperscaler-adjacent demand or capex commentary would undercut the CEG/VST behind-the-meter thesis and feed the efficiency counter-thread. The measurable tell: forward capex and AI-revenue color on the call.

Thesis Impact

Disconfirming first:

• VRT | Conviction: HOLD | Surprise: MED | Crusoe paused its 1.8GW Cheyenne site "at the request of our customer" — the first NEW crack in the buildout cadence and a mild CONTRADICTION of "orders track hyperscaler capex." But it's one site being reshuffled (Crusoe still touts a 5GW pipeline) and OpenAI's separate 10GW Ohio talks more than offset. Backlog thesis intact. | ⁴
• CEG / VST | Conviction: HOLD | Surprise: LOW | A timing trick cutting LLM *training* energy up to 14% (IEEE) feeds the building efficiency/demand-flattening counter-thread (now 3 days). But it hits training — a small slice — not inference, and it's nowhere near a step-change. CONTRADICTS at the margin only; not enough to move firm-baseload conviction. | ⁵
• SCCO | Conviction: HOLD | Surprise: MED | Iraq lifting exports as tankers re-transit Hormuz (T1) confirms Mideast de-escalation, draining the fear premium that had flattered copper. The recent rise was partly risk-appetite, so the structural-deficit thesis must now carry the price alone. Neutral to pillars, near-term price headwind. | ¹
• FCX | Conviction: HOLD | Surprise: MED | Same de-escalation strips more from FCX given its higher geopolitical/Indonesia beta; today removes a bid that flattered the higher-beta name most, widening the SCCO-over-FCX split. CONTRADICTS near-term price, structural deficit unchanged. | ¹
• KEEL | Conviction: HOLD | Surprise: MED | NEW 8-K: entered a material agreement, took on material debt, AND sold unregistered equity. Financing was expected-news, but the equity sale pressures the "no heavy dilution" pillar — while the unspecified material agreement could be the awaited anchor AI/HPC lease. Direction hinges on the detail; verify size of raise and counterparty today. | ⁸

Confirming:

• Own-the-Bottleneck | Conviction: HOLD | Surprise: MED | OpenAI in talks for a 10GW Ohio data center (SB Energy, federal land) — a large NEW demand print reconfirming power as the binding AI constraint. Paired with China's $295B DC rollout needing massive grid spend, the demand side CONFIRMS the thesis even as Crusoe's pause is the offsetting residual. | ²

Inflection Radar

[emergent] AI Inference Bottlenecks | Research highlights fundamental limits in current LLM scaling, specifically addressing the quadratic complexity of attention (60) and the failure modes when quantizing the KV cache (61), suggesting memory/state management is the next major compute hurdle. | Touches: NEW | ⁹

[emergent] Dynamic Knowledge Systems | Current LLM wiki architectures assume static knowledge bases; research is emerging on proactive, materiality-scored pinning mechanisms required for truly time-evolving information landscapes (63). | Touches: NEW | ¹⁰

[emergent] Physical Embodiment & Dexterity | Robotics research is moving beyond simple task execution, focusing on the necessity of contact-rich interaction metrics (77) and developing unified locomotion/navigation policies for versatile terrains (68). | Touches: NEW | ¹¹

[emergent] AI Agent Reliability | The challenge for generalist AI agents is shifting from capability to reliability; work is emerging on robust web navigation (65) and specialized control mechanisms like active inference for constrained decision-making (59). | Touches: NEW | ¹²

[emergent] On-Device Compute Optimization | Hardware-software co-design is critical for edge AI; specific research targets operator fusion (64) and managing state/memory constraints (61) to make large models viable outside data centers. | Touches: NEW | ¹³

[emergent] Regulatory Bottlenecks (Energy) | The interconnection cost allocation debate (76) remains a critical, unresolved regulatory choke point for massive load additions like data centers, signaling policy risk for the sector. | Touches: NEW | ¹⁴

[emergent] Advanced Counter-UAS Integration | The defense sector is rapidly moving from single-domain solutions to integrated, multi-platform counter-UAS suites (83, 84, 87), driven by high-profile geopolitical flashpoints. | Touches: NEW | ¹⁵

[emergent] Human-Robot Interface (HRI) | The frontier in robotics is moving toward intuitive, non-explicit control methods, demonstrated by research into facial recognition control (88) and advanced demonstration curation (66). | Touches: NEW | ¹⁶

QA & Caveats

No issues found.

Sources

1. Iraq Boosts Oil Exports as More Tankers Transit Strait of Hormuz bloomberg.com
2. OpenAI in talks to lease 10GW data center from SB Energy in Ohio datacenterdynamics.com
3. China’s AI Ambitions Depend on Massive Power-Grid Investments bloomberg.com
4. Crusoe pauses work on 1.8GW Cheyenne, Wyoming, data center "at the request of our customer" datacenterdynamics.com
5. Timing Trick Cuts Energy Used in LLM Training by Up to 14 Percent spectrum.ieee.org
6. Microsoft seeks Nevada tariff to shield ratepayers from data center costs utilitydive.com
7. Most Americans Want a National Data Center Moratorium heatmap.news
8. KEEL · 8-K - Current report sec.gov
9. Blurry Window Attention arxiv.org
10. Streaming Knowledge Compilation: Proactive Materiality-Scored Pinning for Time-Evolving LLM Wikis arxiv.org
11. GuideWalk: Learning Unified Autonomous Navigation and Locomotion for Humanoid Robots across Versatile Terrains arxiv.org
12. WebChallenger: A Reliable and Efficient Generalist Web Agent arxiv.org
13. Operator Fusion for LLM Inference on the Tensix Architecture arxiv.org
14. FirstEnergy asks FERC to require data centers to pay for transmission interconnection costs utilitydive.com
15. US approves Kuwait request to buy nearly $2 billion of counter-drone platforms defensenews.com
16. Controlling a robot only using my face reddit.com